champollion 0.3.3

package/shared/catalogue/external-results.json

@@ -0,0 +1,3888 @@
+{
+  "_comment": "CITE-ONLY catalogue of externally-published MT results. Each record is ONE concrete, verified datapoint — a single (model, benchmark, NAMED metric, value) tuple with an explicit lower_is_better direction and a source_url. We NEVER re-host the underlying corpus; we record the citation, the metric, and the published value. TWO integrity axes are tracked per record and must NOT be conflated: (1) verified:true — the (model, benchmark, metric, value) was checked against the cited source (citation accuracy); a record may not be added unverified. (2) badge 'unverified — not reproduced by Champollion' — the number is someone else's measurement, run by them on their data, NOT re-run in the Champollion harness. So verified:true does NOT mean we reproduced the result. Metric VARIANTS (legacy-BLEU vs spBLEU vs chrF vs chrF++ vs MetricX-24 vs a specific COMET version) are NOT directly comparable — metric_variant_flag on each record says so. Aggregate/leaderboard pointers (WMT, Papers with Code, OPUS-MT leaderboard) and model reports we have NOT pinned to a single verified number live in manifest.cite, NOT here. This file is an SSOT in the same family as shared/method-registry.json and shared/human-services.json; it is read at build time by the public Index page (cli/website/src/pages/catalogue.js). DO add new verified datapoints; do NOT paste corpus content, gold references, or per-sentence outputs here. Validate against shared/schemas/external-results.schema.json. v3 ADDS three things WITHOUT changing the integrity rules above: (a) optional `pair` {source,target ISO-639-3} on records that are a single directed language pair, so the datapoint surfaces as a FLAGGED EXTERNAL EDGE on /mesh (cite-only, never a chain bridge, never in routing); (b) optional `signal_strength` — the OBJECTIVE A–F grade of the corpus the number was measured on (size · example length · domain breadth · contamination), SEPARATE from provenance: an external result is flagged not-reproduced but is NOT penalized on grade for being external; a high model score on a tiny or HIGH-contamination test set is still a WEAK edge (interim rubric — the coordinator reconciles with the canonical mesh-routing rubric; each grade carries a plain-English rationale); (c) a top-level `methods` availability INDEX cataloguing every method seen in results (availability / weight openness / access / license), a SUPERSET of method-registry.json — entries with runnable_in_champollion:false are cited-only (point users to the method now, implement later).",
+  "version": 3,
+  "results": [
+    {
+      "id": "translategemma-27b-metricx",
+      "model": "TranslateGemma 27B",
+      "source": "TranslateGemma Technical Report",
+      "org": "Google",
+      "year": 2026,
+      "category": "model-report",
+      "benchmark": "WMT24++",
+      "metric": "MetricX-24",
+      "value": 3.09,
+      "lower_is_better": true,
+      "metric_variant_flag": "MetricX-24 — a learned MQM-style error metric on a ~0–25 scale where LOWER is better; correlates with human MQM error counts. NOT comparable to BLEU/chrF/chrF++ (higher-better surface metrics), to COMET-22, or across MetricX versions.",
+      "headline": "Google's open Gemma-3-based MT suite (55 languages); the 27B reaches the report's best MetricX on WMT24++.",
+      "langs_or_pairs": "WMT24++ — 55 language pairs, automatic eval averaged across directions.",
+      "source_url": "https://arxiv.org/abs/2601.09012",
+      "citation": "Finkelstein, Caswell, Domhan, Peter, Juraska, Riley, Deutsch, et al. (2026), 'TranslateGemma Technical Report', arXiv:2601.09012.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "TranslateGemma is a DEDICATED MT model fine-tuned from Gemma 3 — not to be confused with the general-purpose Gemma 2 LLM (see manifest.cite).",
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "Average over 55 WMT24++ out-of-English directions (general-MT, multi-domain, sentence-level). WMT24++ is a recent held-out-style set (MEDIUM contamination). Solid multi-pair breadth, but an AGGREGATE average rather than a single edge → B."
+      },
+      "method_ref": "translategemma"
+    },
+    {
+      "id": "translategemma-27b-comet22",
+      "model": "TranslateGemma 27B",
+      "source": "TranslateGemma Technical Report",
+      "org": "Google",
+      "year": 2026,
+      "category": "model-report",
+      "benchmark": "WMT24++",
+      "metric": "COMET-22",
+      "value": 84.4,
+      "lower_is_better": false,
+      "metric_variant_flag": "COMET-22 — the Unbabel wmt22-comet-da learned metric, reported on a 0–100 scale where HIGHER is better. NOT comparable to BLEU/chrF/chrF++, to MetricX-24, or across COMET model versions.",
+      "headline": "Google's open Gemma-3-based MT suite (55 languages); the 27B reaches the report's best COMET-22 on WMT24++.",
+      "langs_or_pairs": "WMT24++ — 55 language pairs, automatic eval averaged across directions.",
+      "source_url": "https://arxiv.org/abs/2601.09012",
+      "citation": "Finkelstein, Caswell, Domhan, Peter, Juraska, Riley, Deutsch, et al. (2026), 'TranslateGemma Technical Report', arXiv:2601.09012.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "Average over 55 WMT24++ out-of-English directions (general-MT, multi-domain, sentence-level). WMT24++ is a recent held-out-style set (MEDIUM contamination). Solid multi-pair breadth, but an AGGREGATE average rather than a single edge → B."
+      },
+      "method_ref": "translategemma"
+    },
+    {
+      "id": "translategemma-12b-metricx",
+      "model": "TranslateGemma 12B",
+      "source": "TranslateGemma Technical Report",
+      "org": "Google",
+      "year": 2026,
+      "category": "model-report",
+      "benchmark": "WMT24++",
+      "metric": "MetricX-24",
+      "value": 3.6,
+      "lower_is_better": true,
+      "metric_variant_flag": "MetricX-24 — a learned MQM-style error metric on a ~0–25 scale where LOWER is better; correlates with human MQM error counts. NOT comparable to BLEU/chrF/chrF++ (higher-better surface metrics), to COMET-22, or across MetricX versions.",
+      "headline": "On WMT24++ the 12B TranslateGemma surpasses the 27B Gemma 3 baseline (4.04 MetricX) — fine-tuning beats scale.",
+      "langs_or_pairs": "WMT24++ — 55 language pairs, automatic eval averaged across directions.",
+      "source_url": "https://arxiv.org/abs/2601.09012",
+      "citation": "Finkelstein, Caswell, Domhan, Peter, Juraska, Riley, Deutsch, et al. (2026), 'TranslateGemma Technical Report', arXiv:2601.09012.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "Average over 55 WMT24++ out-of-English directions (general-MT, multi-domain, sentence-level). WMT24++ is a recent held-out-style set (MEDIUM contamination). Solid multi-pair breadth, but an AGGREGATE average rather than a single edge → B."
+      },
+      "method_ref": "translategemma"
+    },
+    {
+      "id": "translategemma-12b-comet22",
+      "model": "TranslateGemma 12B",
+      "source": "TranslateGemma Technical Report",
+      "org": "Google",
+      "year": 2026,
+      "category": "model-report",
+      "benchmark": "WMT24++",
+      "metric": "COMET-22",
+      "value": 83.5,
+      "lower_is_better": false,
+      "metric_variant_flag": "COMET-22 — the Unbabel wmt22-comet-da learned metric, reported on a 0–100 scale where HIGHER is better. NOT comparable to BLEU/chrF/chrF++, to MetricX-24, or across COMET model versions.",
+      "headline": "On WMT24++ the 12B TranslateGemma surpasses the 27B Gemma 3 baseline (83.1 COMET-22).",
+      "langs_or_pairs": "WMT24++ — 55 language pairs, automatic eval averaged across directions.",
+      "source_url": "https://arxiv.org/abs/2601.09012",
+      "citation": "Finkelstein, Caswell, Domhan, Peter, Juraska, Riley, Deutsch, et al. (2026), 'TranslateGemma Technical Report', arXiv:2601.09012.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "Average over 55 WMT24++ out-of-English directions (general-MT, multi-domain, sentence-level). WMT24++ is a recent held-out-style set (MEDIUM contamination). Solid multi-pair breadth, but an AGGREGATE average rather than a single edge → B."
+      },
+      "method_ref": "translategemma"
+    },
+    {
+      "id": "translategemma-4b-metricx",
+      "model": "TranslateGemma 4B",
+      "source": "TranslateGemma Technical Report",
+      "org": "Google",
+      "year": 2026,
+      "category": "model-report",
+      "benchmark": "WMT24++",
+      "metric": "MetricX-24",
+      "value": 5.32,
+      "lower_is_better": true,
+      "metric_variant_flag": "MetricX-24 — a learned MQM-style error metric on a ~0–25 scale where LOWER is better; correlates with human MQM error counts. NOT comparable to BLEU/chrF/chrF++ (higher-better surface metrics), to COMET-22, or across MetricX versions.",
+      "headline": "Smallest TranslateGemma size; still a large MetricX gain over the 4B Gemma 3 baseline (6.97).",
+      "langs_or_pairs": "WMT24++ — 55 language pairs, automatic eval averaged across directions.",
+      "source_url": "https://arxiv.org/abs/2601.09012",
+      "citation": "Finkelstein, Caswell, Domhan, Peter, Juraska, Riley, Deutsch, et al. (2026), 'TranslateGemma Technical Report', arXiv:2601.09012.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "Average over 55 WMT24++ out-of-English directions (general-MT, multi-domain, sentence-level). WMT24++ is a recent held-out-style set (MEDIUM contamination). Solid multi-pair breadth, but an AGGREGATE average rather than a single edge → B."
+      },
+      "method_ref": "translategemma"
+    },
+    {
+      "id": "translategemma-4b-comet22",
+      "model": "TranslateGemma 4B",
+      "source": "TranslateGemma Technical Report",
+      "org": "Google",
+      "year": 2026,
+      "category": "model-report",
+      "benchmark": "WMT24++",
+      "metric": "COMET-22",
+      "value": 80.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "COMET-22 — the Unbabel wmt22-comet-da learned metric, reported on a 0–100 scale where HIGHER is better. NOT comparable to BLEU/chrF/chrF++, to MetricX-24, or across COMET model versions.",
+      "headline": "Smallest TranslateGemma size; COMET-22 gain over the 4B Gemma 3 baseline (77.2).",
+      "langs_or_pairs": "WMT24++ — 55 language pairs, automatic eval averaged across directions.",
+      "source_url": "https://arxiv.org/abs/2601.09012",
+      "citation": "Finkelstein, Caswell, Domhan, Peter, Juraska, Riley, Deutsch, et al. (2026), 'TranslateGemma Technical Report', arXiv:2601.09012.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "Average over 55 WMT24++ out-of-English directions (general-MT, multi-domain, sentence-level). WMT24++ is a recent held-out-style set (MEDIUM contamination). Solid multi-pair breadth, but an AGGREGATE average rather than a single edge → B."
+      },
+      "method_ref": "translategemma"
+    },
+    {
+      "id": "gemma3-27b-metricx",
+      "model": "Gemma 3 27B (baseline)",
+      "source": "TranslateGemma Technical Report",
+      "org": "Google",
+      "year": 2026,
+      "category": "model-report",
+      "benchmark": "WMT24++",
+      "metric": "MetricX-24",
+      "value": 4.04,
+      "lower_is_better": true,
+      "metric_variant_flag": "MetricX-24 — a learned MQM-style error metric on a ~0–25 scale where LOWER is better; correlates with human MQM error counts. NOT comparable to BLEU/chrF/chrF++ (higher-better surface metrics), to COMET-22, or across MetricX versions.",
+      "headline": "The pre-fine-tuning Gemma 3 27B MT baseline reported in the TranslateGemma study (fine-tuning takes it from 4.04 to 3.09).",
+      "langs_or_pairs": "WMT24++ — 55 language pairs, automatic eval averaged across directions.",
+      "source_url": "https://arxiv.org/abs/2601.09012",
+      "citation": "Finkelstein, Caswell, Domhan, Peter, Juraska, Riley, Deutsch, et al. (2026), 'TranslateGemma Technical Report', arXiv:2601.09012.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "Gemma 3 is a general-purpose open LLM; this is its baseline MT score in the TranslateGemma report, included to show the fine-tuning delta.",
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "Average over 55 WMT24++ out-of-English directions (general-MT, multi-domain, sentence-level). WMT24++ is a recent held-out-style set (MEDIUM contamination). Solid multi-pair breadth, but an AGGREGATE average rather than a single edge → B."
+      },
+      "method_ref": "gemma-3"
+    },
+    {
+      "id": "gemma3-27b-comet22",
+      "model": "Gemma 3 27B (baseline)",
+      "source": "TranslateGemma Technical Report",
+      "org": "Google",
+      "year": 2026,
+      "category": "model-report",
+      "benchmark": "WMT24++",
+      "metric": "COMET-22",
+      "value": 83.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "COMET-22 — the Unbabel wmt22-comet-da learned metric, reported on a 0–100 scale where HIGHER is better. NOT comparable to BLEU/chrF/chrF++, to MetricX-24, or across COMET model versions.",
+      "headline": "The pre-fine-tuning Gemma 3 27B MT baseline reported in the TranslateGemma study.",
+      "langs_or_pairs": "WMT24++ — 55 language pairs, automatic eval averaged across directions.",
+      "source_url": "https://arxiv.org/abs/2601.09012",
+      "citation": "Finkelstein, Caswell, Domhan, Peter, Juraska, Riley, Deutsch, et al. (2026), 'TranslateGemma Technical Report', arXiv:2601.09012.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "Gemma 3 is a general-purpose open LLM; this is its baseline MT score in the TranslateGemma report, included to show the fine-tuning delta.",
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "Average over 55 WMT24++ out-of-English directions (general-MT, multi-domain, sentence-level). WMT24++ is a recent held-out-style set (MEDIUM contamination). Solid multi-pair breadth, but an AGGREGATE average rather than a single edge → B."
+      },
+      "method_ref": "gemma-3"
+    },
+    {
+      "id": "gemma3-12b-metricx",
+      "model": "Gemma 3 12B (baseline)",
+      "source": "TranslateGemma Technical Report",
+      "org": "Google",
+      "year": 2026,
+      "category": "model-report",
+      "benchmark": "WMT24++",
+      "metric": "MetricX-24",
+      "value": 4.86,
+      "lower_is_better": true,
+      "metric_variant_flag": "MetricX-24 — a learned MQM-style error metric on a ~0–25 scale where LOWER is better; correlates with human MQM error counts. NOT comparable to BLEU/chrF/chrF++ (higher-better surface metrics), to COMET-22, or across MetricX versions.",
+      "headline": "The pre-fine-tuning Gemma 3 12B MT baseline reported in the TranslateGemma study.",
+      "langs_or_pairs": "WMT24++ — 55 language pairs, automatic eval averaged across directions.",
+      "source_url": "https://arxiv.org/abs/2601.09012",
+      "citation": "Finkelstein, Caswell, Domhan, Peter, Juraska, Riley, Deutsch, et al. (2026), 'TranslateGemma Technical Report', arXiv:2601.09012.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "Gemma 3 is a general-purpose open LLM; this is its baseline MT score in the TranslateGemma report, included to show the fine-tuning delta.",
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "Average over 55 WMT24++ out-of-English directions (general-MT, multi-domain, sentence-level). WMT24++ is a recent held-out-style set (MEDIUM contamination). Solid multi-pair breadth, but an AGGREGATE average rather than a single edge → B."
+      },
+      "method_ref": "gemma-3"
+    },
+    {
+      "id": "gemma3-12b-comet22",
+      "model": "Gemma 3 12B (baseline)",
+      "source": "TranslateGemma Technical Report",
+      "org": "Google",
+      "year": 2026,
+      "category": "model-report",
+      "benchmark": "WMT24++",
+      "metric": "COMET-22",
+      "value": 81.6,
+      "lower_is_better": false,
+      "metric_variant_flag": "COMET-22 — the Unbabel wmt22-comet-da learned metric, reported on a 0–100 scale where HIGHER is better. NOT comparable to BLEU/chrF/chrF++, to MetricX-24, or across COMET model versions.",
+      "headline": "The pre-fine-tuning Gemma 3 12B MT baseline reported in the TranslateGemma study.",
+      "langs_or_pairs": "WMT24++ — 55 language pairs, automatic eval averaged across directions.",
+      "source_url": "https://arxiv.org/abs/2601.09012",
+      "citation": "Finkelstein, Caswell, Domhan, Peter, Juraska, Riley, Deutsch, et al. (2026), 'TranslateGemma Technical Report', arXiv:2601.09012.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "Gemma 3 is a general-purpose open LLM; this is its baseline MT score in the TranslateGemma report, included to show the fine-tuning delta.",
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "Average over 55 WMT24++ out-of-English directions (general-MT, multi-domain, sentence-level). WMT24++ is a recent held-out-style set (MEDIUM contamination). Solid multi-pair breadth, but an AGGREGATE average rather than a single edge → B."
+      },
+      "method_ref": "gemma-3"
+    },
+    {
+      "id": "gemma3-4b-metricx",
+      "model": "Gemma 3 4B (baseline)",
+      "source": "TranslateGemma Technical Report",
+      "org": "Google",
+      "year": 2026,
+      "category": "model-report",
+      "benchmark": "WMT24++",
+      "metric": "MetricX-24",
+      "value": 6.97,
+      "lower_is_better": true,
+      "metric_variant_flag": "MetricX-24 — a learned MQM-style error metric on a ~0–25 scale where LOWER is better; correlates with human MQM error counts. NOT comparable to BLEU/chrF/chrF++ (higher-better surface metrics), to COMET-22, or across MetricX versions.",
+      "headline": "The pre-fine-tuning Gemma 3 4B MT baseline reported in the TranslateGemma study.",
+      "langs_or_pairs": "WMT24++ — 55 language pairs, automatic eval averaged across directions.",
+      "source_url": "https://arxiv.org/abs/2601.09012",
+      "citation": "Finkelstein, Caswell, Domhan, Peter, Juraska, Riley, Deutsch, et al. (2026), 'TranslateGemma Technical Report', arXiv:2601.09012.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "Gemma 3 is a general-purpose open LLM; this is its baseline MT score in the TranslateGemma report, included to show the fine-tuning delta.",
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "Average over 55 WMT24++ out-of-English directions (general-MT, multi-domain, sentence-level). WMT24++ is a recent held-out-style set (MEDIUM contamination). Solid multi-pair breadth, but an AGGREGATE average rather than a single edge → B."
+      },
+      "method_ref": "gemma-3"
+    },
+    {
+      "id": "gemma3-4b-comet22",
+      "model": "Gemma 3 4B (baseline)",
+      "source": "TranslateGemma Technical Report",
+      "org": "Google",
+      "year": 2026,
+      "category": "model-report",
+      "benchmark": "WMT24++",
+      "metric": "COMET-22",
+      "value": 77.2,
+      "lower_is_better": false,
+      "metric_variant_flag": "COMET-22 — the Unbabel wmt22-comet-da learned metric, reported on a 0–100 scale where HIGHER is better. NOT comparable to BLEU/chrF/chrF++, to MetricX-24, or across COMET model versions.",
+      "headline": "The pre-fine-tuning Gemma 3 4B MT baseline reported in the TranslateGemma study.",
+      "langs_or_pairs": "WMT24++ — 55 language pairs, automatic eval averaged across directions.",
+      "source_url": "https://arxiv.org/abs/2601.09012",
+      "citation": "Finkelstein, Caswell, Domhan, Peter, Juraska, Riley, Deutsch, et al. (2026), 'TranslateGemma Technical Report', arXiv:2601.09012.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "Gemma 3 is a general-purpose open LLM; this is its baseline MT score in the TranslateGemma report, included to show the fine-tuning delta.",
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "Average over 55 WMT24++ out-of-English directions (general-MT, multi-domain, sentence-level). WMT24++ is a recent held-out-style set (MEDIUM contamination). Solid multi-pair breadth, but an AGGREGATE average rather than a single edge → B."
+      },
+      "method_ref": "gemma-3"
+    },
+    {
+      "id": "opus-en-fr-tatoeba-bleu",
+      "model": "OPUS-MT en-fr",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (eng-fra)",
+      "metric": "BLEU",
+      "value": 50.5,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT en→fr scores 50.5 BLEU on the Tatoeba test set.",
+      "langs_or_pairs": "eng→fra (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-fr",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "fra"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Large Tatoeba crowd-sentence test set (thousands of short, mixed-register sentences); Tatoeba is broadly trained on → MEDIUM contamination. Strong size, informal short examples → B."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-en-fr-tatoeba-chrf",
+      "model": "OPUS-MT en-fr",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (eng-fra)",
+      "metric": "chrF",
+      "value": 0.672,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF (Helsinki chrF2, 0–1 scale) — character-n-gram F-score on a 0–1 scale where HIGHER is better (e.g. 0.672). NOT the same as chrF++ (word+char, 0–100), and NOT comparable to BLEU/spBLEU/COMET/MetricX.",
+      "headline": "OPUS-MT en→fr: chrF 0.672 on Tatoeba.",
+      "langs_or_pairs": "eng→fra (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-fr",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "fra"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Same Tatoeba eng-fra test; large but short crowd sentences, MEDIUM contamination → B."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-en-fr-newstest2013-bleu",
+      "model": "OPUS-MT en-fr",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "newstest2013 (eng-fra)",
+      "metric": "BLEU",
+      "value": 33.2,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT en→fr: 33.2 BLEU on WMT newstest2013 (news).",
+      "langs_or_pairs": "eng→fra (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-fr",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "fra"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": 3000,
+        "example_length": "sentence",
+        "domain_breadth": "single-domain",
+        "contamination": "MEDIUM",
+        "rationale": "WMT newstest2013: ~3000 professional NEWS sentences — single-domain, and an older WMT set that is broadly trained on (MEDIUM-HIGH contamination) → C."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-en-es-tatoeba-bleu",
+      "model": "OPUS-MT en-es",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (eng-spa)",
+      "metric": "BLEU",
+      "value": 54.9,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT en→es: 54.9 BLEU on Tatoeba.",
+      "langs_or_pairs": "eng→spa (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-es",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "spa"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Large Tatoeba eng-spa test; MEDIUM contamination → B."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-en-es-tatoeba-chrf",
+      "model": "OPUS-MT en-es",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (eng-spa)",
+      "metric": "chrF",
+      "value": 0.721,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF (Helsinki chrF2, 0–1 scale) — character-n-gram F-score on a 0–1 scale where HIGHER is better (e.g. 0.672). NOT the same as chrF++ (word+char, 0–100), and NOT comparable to BLEU/spBLEU/COMET/MetricX.",
+      "headline": "OPUS-MT en→es: chrF 0.721 on Tatoeba.",
+      "langs_or_pairs": "eng→spa (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-es",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "spa"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Same eng-spa Tatoeba test → B."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-en-de-tatoeba-bleu",
+      "model": "OPUS-MT en-de",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (eng-deu)",
+      "metric": "BLEU",
+      "value": 47.3,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT en→de: 47.3 BLEU on Tatoeba.",
+      "langs_or_pairs": "eng→deu (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-de",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "deu"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Large Tatoeba eng-deu test; MEDIUM contamination → B."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-en-de-tatoeba-chrf",
+      "model": "OPUS-MT en-de",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (eng-deu)",
+      "metric": "chrF",
+      "value": 0.664,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF (Helsinki chrF2, 0–1 scale) — character-n-gram F-score on a 0–1 scale where HIGHER is better (e.g. 0.672). NOT the same as chrF++ (word+char, 0–100), and NOT comparable to BLEU/spBLEU/COMET/MetricX.",
+      "headline": "OPUS-MT en→de: chrF 0.664 on Tatoeba.",
+      "langs_or_pairs": "eng→deu (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-de",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "deu"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Same eng-deu Tatoeba test → B."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-en-de-newstest2019-bleu",
+      "model": "OPUS-MT en-de",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "newstest2019 (eng-deu)",
+      "metric": "BLEU",
+      "value": 40.9,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT en→de: 40.9 BLEU on WMT newstest2019.",
+      "langs_or_pairs": "eng→deu (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-de",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "deu"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": 2000,
+        "example_length": "sentence",
+        "domain_breadth": "single-domain",
+        "contamination": "MEDIUM",
+        "rationale": "WMT newstest2019: news single-domain, broadly trained on → C."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-en-ru-tatoeba-bleu",
+      "model": "OPUS-MT en-ru",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (eng-rus)",
+      "metric": "BLEU",
+      "value": 48.4,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT en→ru: 48.4 BLEU on Tatoeba.",
+      "langs_or_pairs": "eng→rus (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-ru",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "rus"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Large Tatoeba eng-rus test; MEDIUM contamination → B."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-en-ru-tatoeba-chrf",
+      "model": "OPUS-MT en-ru",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (eng-rus)",
+      "metric": "chrF",
+      "value": 0.669,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF (Helsinki chrF2, 0–1 scale) — character-n-gram F-score on a 0–1 scale where HIGHER is better (e.g. 0.672). NOT the same as chrF++ (word+char, 0–100), and NOT comparable to BLEU/spBLEU/COMET/MetricX.",
+      "headline": "OPUS-MT en→ru: chrF 0.669 on Tatoeba.",
+      "langs_or_pairs": "eng→rus (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-ru",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "rus"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Same eng-rus Tatoeba test → B."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-tcbig-en-tr-tatoeba-bleu",
+      "model": "OPUS-MT tc-big en-tr",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "tatoeba-test-v2021-08-07 (eng-tur)",
+      "metric": "BLEU",
+      "value": 42.3,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT tc-big en→tr: 42.3 BLEU on 13,907-sentence Tatoeba test.",
+      "langs_or_pairs": "eng→tur (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-tc-big-en-tr",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "tur"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 13907,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Large stated test set (13,907 sentences) of mixed crowd sentences; MEDIUM contamination → B. Size is explicit on the card."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-tcbig-en-tr-tatoeba-chrf",
+      "model": "OPUS-MT tc-big en-tr",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "tatoeba-test-v2021-08-07 (eng-tur)",
+      "metric": "chrF",
+      "value": 0.68726,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF (Helsinki chrF2, 0–1 scale) — character-n-gram F-score on a 0–1 scale where HIGHER is better (e.g. 0.672). NOT the same as chrF++ (word+char, 0–100), and NOT comparable to BLEU/spBLEU/COMET/MetricX.",
+      "headline": "OPUS-MT tc-big en→tr: chrF 0.687 on Tatoeba.",
+      "langs_or_pairs": "eng→tur (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-tc-big-en-tr",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "tur"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 13907,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Same 13,907-sentence eng-tur test → B."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-tcbig-en-tr-flores-bleu",
+      "model": "OPUS-MT tc-big en-tr",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "FLORES-101 devtest (eng-tur)",
+      "metric": "BLEU",
+      "value": 31.4,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT tc-big en→tr: 31.4 BLEU on FLORES-101 (relative-only).",
+      "langs_or_pairs": "eng→tur (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-tc-big-en-tr",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "tur"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "HIGH",
+        "rationale": "FLORES-101 devtest (1012 multi-domain Wikipedia sentences) — strong size and breadth, BUT FLORES is broadly memorized (HIGH contamination) so it is RELATIVE-ONLY / illustration; contamination caps the signal at C."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-tcbig-en-tr-flores-chrf",
+      "model": "OPUS-MT tc-big en-tr",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "FLORES-101 devtest (eng-tur)",
+      "metric": "chrF",
+      "value": 0.62829,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF (Helsinki chrF2, 0–1 scale) — character-n-gram F-score on a 0–1 scale where HIGHER is better (e.g. 0.672). NOT the same as chrF++ (word+char, 0–100), and NOT comparable to BLEU/spBLEU/COMET/MetricX.",
+      "headline": "OPUS-MT tc-big en→tr: chrF 0.628 on FLORES-101 (relative-only).",
+      "langs_or_pairs": "eng→tur (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-tc-big-en-tr",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "tur"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "HIGH",
+        "rationale": "Same eng-tur FLORES-101 devtest; HIGH contamination → relative-only → C."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-en-ha-tatoeba-bleu",
+      "model": "OPUS-MT en-ha",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (eng-hau)",
+      "metric": "BLEU",
+      "value": 17.6,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT en→ha: 17.6 BLEU on a small Tatoeba test.",
+      "langs_or_pairs": "eng→hau (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-ha",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "hau"
+      },
+      "signal_strength": {
+        "grade": "D",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "narrow",
+        "contamination": "LOW",
+        "rationale": "Low-resource Tatoeba eng-hau test (small; size not stated on the card). Little contamination (LOW) but the SMALL low-resource test set is the dominant weakness → D."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-ha-en-tatoeba-bleu",
+      "model": "OPUS-MT ha-en",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (hau-eng)",
+      "metric": "BLEU",
+      "value": 39.0,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT ha→en: 39.0 BLEU — high score, small test → still a weak edge.",
+      "langs_or_pairs": "hau→eng (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-ha-en",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "hau",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "D",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "narrow",
+        "contamination": "LOW",
+        "rationale": "Reverse direction scores higher (39.0) but is measured on the SAME small low-resource Tatoeba set — a high score on a weak test is still a WEAK edge → D."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-en-mul-amh-bleu",
+      "model": "OPUS-MT en-mul",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (eng-amh)",
+      "metric": "BLEU",
+      "value": 8.5,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT (multilingual) en→am: 8.5 BLEU.",
+      "langs_or_pairs": "eng→amh (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-mul",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "amh"
+      },
+      "signal_strength": {
+        "grade": "D",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "narrow",
+        "contamination": "LOW",
+        "rationale": "Low-resource pair via the multilingual en-mul model; small Tatoeba eng-amh test → weak signal D."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-mul-en-amh-bleu",
+      "model": "OPUS-MT mul-en",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (amh-eng)",
+      "metric": "BLEU",
+      "value": 25.6,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT (multilingual) am→en: 25.6 BLEU.",
+      "langs_or_pairs": "amh→eng (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-mul-en",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "amh",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "D",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "narrow",
+        "contamination": "LOW",
+        "rationale": "amh→en via multilingual mul-en; small low-resource test → D."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-en-mul-uig-bleu",
+      "model": "OPUS-MT en-mul",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (eng-uig)",
+      "metric": "BLEU",
+      "value": 2.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT (multilingual) en→ug: 2.1 BLEU — essentially no signal.",
+      "langs_or_pairs": "eng→uig (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-mul",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "uig"
+      },
+      "signal_strength": {
+        "grade": "F",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "narrow",
+        "contamination": "LOW",
+        "rationale": "Near-zero BLEU (2.1) on a tiny low-resource eng-uig test — essentially no usable signal → F."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-mul-en-uig-bleu",
+      "model": "OPUS-MT mul-en",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (uig-eng)",
+      "metric": "BLEU",
+      "value": 11.4,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT (multilingual) ug→en: 11.4 BLEU.",
+      "langs_or_pairs": "uig→eng (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-mul-en",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "uig",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "D",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "narrow",
+        "contamination": "LOW",
+        "rationale": "ug→en via multilingual mul-en; small low-resource test → D."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-en-mul-sah-bleu",
+      "model": "OPUS-MT en-mul",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (eng-sah)",
+      "metric": "BLEU",
+      "value": 0.2,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT (multilingual) en→sah: 0.2 BLEU — the model essentially fails this pair.",
+      "langs_or_pairs": "eng→sah (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-mul",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "sah"
+      },
+      "signal_strength": {
+        "grade": "F",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "narrow",
+        "contamination": "LOW",
+        "rationale": "BLEU 0.2 / chrF 0.007 — the model essentially fails this direction; a near-empty-signal measurement → F. The starkest weak edge."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-mul-en-sah-bleu",
+      "model": "OPUS-MT mul-en",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (sah-eng)",
+      "metric": "BLEU",
+      "value": 5.0,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT (multilingual) sah→en: 5.0 BLEU.",
+      "langs_or_pairs": "sah→eng (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-mul-en",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "sah",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "F",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "narrow",
+        "contamination": "LOW",
+        "rationale": "sah→en 5.0 BLEU on a tiny low-resource test → very weak signal F."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "opus-en-trk-kaz-bleu",
+      "model": "OPUS-MT en-trk (Turkic)",
+      "source": "OPUS-MT model card (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2020,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (eng-kaz)",
+      "metric": "BLEU",
+      "value": 11.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (OPUS-MT model-card, tokenized) — a tokenized n-gram precision score on a 0–100 scale where HIGHER is better. The exact tokenization is the OPUS-MT card's, NOT a pinned sacreBLEU signature, so it is NOT directly comparable to spBLEU, sacreBLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "OPUS-MT (Turkic) en→kk: 11.1 BLEU.",
+      "langs_or_pairs": "eng→kaz (single directed pair).",
+      "source_url": "https://huggingface.co/Helsinki-NLP/opus-mt-en-trk",
+      "citation": "Tiedemann & Thottingal (2020), 'OPUS-MT — Building open translation services for the World', EAMT 2020; benchmark table from the model's Hugging Face card.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "kaz"
+      },
+      "signal_strength": {
+        "grade": "D",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "narrow",
+        "contamination": "LOW",
+        "rationale": "eng→kaz via the Turkic-family en-trk model; small low-resource Tatoeba test → D."
+      },
+      "method_ref": "opus-mt"
+    },
+    {
+      "id": "nllb-eng-hau-flores-spbleu",
+      "model": "NLLB-200",
+      "source": "No Language Left Behind (NLLB-200)",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "paper",
+      "benchmark": "FLORES-101 devtest (eng-hau)",
+      "metric": "spBLEU",
+      "value": 33.6,
+      "lower_is_better": false,
+      "metric_variant_flag": "spBLEU (FLORES SentencePiece-tokenized BLEU) — 0–100, HIGHER better; uses a fixed SPM tokenizer so cross-language sums are comparable WITHIN FLORES but NOT comparable to tokenized BLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "NLLB-200 en→ha: 33.6 spBLEU on FLORES-101 (relative-only).",
+      "langs_or_pairs": "eng→hau (single directed pair).",
+      "source_url": "https://arxiv.org/abs/2207.04672",
+      "citation": "NLLB Team et al. (2022), 'No Language Left Behind: Scaling Human-Centered Machine Translation', arXiv:2207.04672.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "FLORES is HIGH-contamination illustration-only data (the data-boundaries doctrine): cited as a relative reading, never an absolute quality claim.",
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "HIGH",
+        "rationale": "FLORES-101 devtest (1012 multi-domain sentences): strong size/breadth, but FLORES is broadly memorized (HIGH contamination) so it is RELATIVE-ONLY — contamination caps the signal at C."
+      },
+      "method_ref": "nllb-200",
+      "pair": {
+        "source": "eng",
+        "target": "hau"
+      }
+    },
+    {
+      "id": "nllb-eng-hau-flores-chrfpp",
+      "model": "NLLB-200",
+      "source": "No Language Left Behind (NLLB-200)",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "paper",
+      "benchmark": "FLORES-101 devtest (eng-hau)",
+      "metric": "chrF++",
+      "value": 53.5,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ (word+character n-gram F-score, 0–100) — HIGHER better. NOT comparable to chrF (0–1), to BLEU/spBLEU, or to COMET/MetricX.",
+      "headline": "NLLB-200 en→ha: 53.5 chrF++ on FLORES-101 (relative-only).",
+      "langs_or_pairs": "eng→hau (single directed pair).",
+      "source_url": "https://arxiv.org/abs/2207.04672",
+      "citation": "NLLB Team et al. (2022), 'No Language Left Behind: Scaling Human-Centered Machine Translation', arXiv:2207.04672.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "FLORES relative-only; chrF++ (0–100) ≠ OPUS-MT chrF (0–1).",
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "HIGH",
+        "rationale": "Same eng-hau FLORES-101 devtest; HIGH contamination → relative-only → C."
+      },
+      "method_ref": "nllb-200",
+      "pair": {
+        "source": "eng",
+        "target": "hau"
+      }
+    },
+    {
+      "id": "nllb-hau-eng-flores-chrfpp",
+      "model": "NLLB-200",
+      "source": "No Language Left Behind (NLLB-200)",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "paper",
+      "benchmark": "FLORES-101 devtest (hau-eng)",
+      "metric": "chrF++",
+      "value": 57.3,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ (word+character n-gram F-score, 0–100) — HIGHER better. NOT comparable to chrF (0–1), to BLEU/spBLEU, or to COMET/MetricX.",
+      "headline": "NLLB-200 ha→en: 57.3 chrF++ on FLORES-101 (relative-only).",
+      "langs_or_pairs": "hau→eng (single directed pair).",
+      "source_url": "https://arxiv.org/abs/2207.04672",
+      "citation": "NLLB Team et al. (2022), 'No Language Left Behind: Scaling Human-Centered Machine Translation', arXiv:2207.04672.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "FLORES relative-only.",
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "HIGH",
+        "rationale": "ha→en on FLORES-101; HIGH contamination → relative-only → C."
+      },
+      "method_ref": "nllb-200",
+      "pair": {
+        "source": "hau",
+        "target": "eng"
+      }
+    },
+    {
+      "id": "nllb-eng-amh-tico-bleu",
+      "model": "NLLB-200",
+      "source": "No Language Left Behind (NLLB-200)",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "paper",
+      "benchmark": "TICO (eng-amh)",
+      "metric": "BLEU",
+      "value": 13.7,
+      "lower_is_better": false,
+      "metric_variant_flag": "BLEU (detokenized, as reported in the NLLB paper Tico table) — 0–100, HIGHER better; tokenization is the paper's, NOT a pinned sacreBLEU signature, so not comparable to spBLEU/chrF/chrF++/COMET.",
+      "headline": "NLLB-200 en→am: 13.7 BLEU on the TICO (health) set.",
+      "langs_or_pairs": "eng→amh (single directed pair).",
+      "source_url": "https://arxiv.org/abs/2207.04672",
+      "citation": "NLLB Team et al. (2022), 'No Language Left Behind: Scaling Human-Centered Machine Translation', arXiv:2207.04672.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "Reported on the TICO public-health set in the NLLB paper (Table 58).",
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "single-domain",
+        "contamination": "MEDIUM",
+        "rationale": "NLLB paper Tico table (public-health domain): single-domain but a genuine held-out set; MEDIUM contamination → C."
+      },
+      "method_ref": "nllb-200",
+      "pair": {
+        "source": "eng",
+        "target": "amh"
+      }
+    },
+    {
+      "id": "nllb-eng-amh-tico-chrfpp",
+      "model": "NLLB-200",
+      "source": "No Language Left Behind (NLLB-200)",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "paper",
+      "benchmark": "TICO (eng-amh)",
+      "metric": "chrF++",
+      "value": 36.7,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ (word+character n-gram F-score, 0–100) — HIGHER better. NOT comparable to chrF (0–1), to BLEU/spBLEU, or to COMET/MetricX.",
+      "headline": "NLLB-200 en→am: 36.7 chrF++ on TICO.",
+      "langs_or_pairs": "eng→amh (single directed pair).",
+      "source_url": "https://arxiv.org/abs/2207.04672",
+      "citation": "NLLB Team et al. (2022), 'No Language Left Behind: Scaling Human-Centered Machine Translation', arXiv:2207.04672.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "TICO health domain (NLLB Table 58).",
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "single-domain",
+        "contamination": "MEDIUM",
+        "rationale": "Same eng-amh TICO set; single-domain health → C."
+      },
+      "method_ref": "nllb-200",
+      "pair": {
+        "source": "eng",
+        "target": "amh"
+      }
+    },
+    {
+      "id": "nllb-amh-eng-tico-chrfpp",
+      "model": "NLLB-200",
+      "source": "No Language Left Behind (NLLB-200)",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "paper",
+      "benchmark": "TICO (amh-eng)",
+      "metric": "chrF++",
+      "value": 60.2,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ (word+character n-gram F-score, 0–100) — HIGHER better. NOT comparable to chrF (0–1), to BLEU/spBLEU, or to COMET/MetricX.",
+      "headline": "NLLB-200 am→en: 60.2 chrF++ on TICO.",
+      "langs_or_pairs": "amh→eng (single directed pair).",
+      "source_url": "https://arxiv.org/abs/2207.04672",
+      "citation": "NLLB Team et al. (2022), 'No Language Left Behind: Scaling Human-Centered Machine Translation', arXiv:2207.04672.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "TICO health domain (NLLB Table 58).",
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "single-domain",
+        "contamination": "MEDIUM",
+        "rationale": "am→en on TICO; single-domain health, MEDIUM contamination → C."
+      },
+      "method_ref": "nllb-200",
+      "pair": {
+        "source": "amh",
+        "target": "eng"
+      }
+    },
+    {
+      "id": "nllb-flores101-avg-spbleu",
+      "model": "NLLB-200",
+      "source": "No Language Left Behind (NLLB-200)",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "paper",
+      "benchmark": "FLORES-101 devtest (101-language average)",
+      "metric": "spBLEU",
+      "value": 24.0,
+      "lower_is_better": false,
+      "metric_variant_flag": "spBLEU (FLORES SentencePiece-tokenized BLEU) — 0–100, HIGHER better; uses a fixed SPM tokenizer so cross-language sums are comparable WITHIN FLORES but NOT comparable to tokenized BLEU, chrF/chrF++, or COMET/MetricX.",
+      "headline": "NLLB-200 averages 24.0 spBLEU across all FLORES-101 directions (relative-only headline).",
+      "langs_or_pairs": "Average across all FLORES-101 devtest directions.",
+      "source_url": "https://arxiv.org/abs/2207.04672",
+      "citation": "NLLB Team et al. (2022), 'No Language Left Behind: Scaling Human-Centered Machine Translation', arXiv:2207.04672.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "Table 30 average (101 languages). FLORES relative-only; aggregate, not a single edge.",
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "HIGH",
+        "rationale": "Average over all FLORES-101 devtest directions; FLORES HIGH contamination → relative-only → C. Aggregate, not a single pair."
+      },
+      "method_ref": "nllb-200"
+    },
+    {
+      "id": "nllb-flores101-avg-chrfpp",
+      "model": "NLLB-200",
+      "source": "No Language Left Behind (NLLB-200)",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "paper",
+      "benchmark": "FLORES-101 devtest (101-language average)",
+      "metric": "chrF++",
+      "value": 41.7,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ (word+character n-gram F-score, 0–100) — HIGHER better. NOT comparable to chrF (0–1), to BLEU/spBLEU, or to COMET/MetricX.",
+      "headline": "NLLB-200 averages 41.7 chrF++ across all FLORES-101 directions (relative-only headline).",
+      "langs_or_pairs": "Average across all FLORES-101 devtest directions.",
+      "source_url": "https://arxiv.org/abs/2207.04672",
+      "citation": "NLLB Team et al. (2022), 'No Language Left Behind: Scaling Human-Centered Machine Translation', arXiv:2207.04672.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": "Table 30 average (101 languages). FLORES relative-only; aggregate.",
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "HIGH",
+        "rationale": "FLORES-101 devtest 101-language average chrF++; HIGH contamination → relative-only → C."
+      },
+      "method_ref": "nllb-200"
+    },
+    {
+      "id": "opus-afr-deu-tatoeba-chrf2",
+      "model": "OPUS-MT afr-deu",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (afr-deu)",
+      "metric": "chrF2",
+      "value": 67.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT afr→deu: chrF2 67.1 on Tatoeba-test.",
+      "langs_or_pairs": "afr→deu (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "afr",
+        "target": "deu"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-afr-eng-tatoeba-chrf2",
+      "model": "OPUS-MT afr-eng",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (afr-eng)",
+      "metric": "chrF2",
+      "value": 73.8,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT afr→eng: chrF2 73.8 on Tatoeba-test.",
+      "langs_or_pairs": "afr→eng (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "afr",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-afr-epo-tatoeba-chrf2",
+      "model": "OPUS-MT afr-epo",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (afr-epo)",
+      "metric": "chrF2",
+      "value": 41.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT afr→epo: chrF2 41.1 on Tatoeba-test.",
+      "langs_or_pairs": "afr→epo (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "afr",
+        "target": "epo"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-afr-rus-tatoeba-chrf2",
+      "model": "OPUS-MT afr-rus",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (afr-rus)",
+      "metric": "chrF2",
+      "value": 58,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT afr→rus: chrF2 58 on Tatoeba-test.",
+      "langs_or_pairs": "afr→rus (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "afr",
+        "target": "rus"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-afr-spa-tatoeba-chrf2",
+      "model": "OPUS-MT afr-spa",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (afr-spa)",
+      "metric": "chrF2",
+      "value": 68,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT afr→spa: chrF2 68 on Tatoeba-test.",
+      "langs_or_pairs": "afr→spa (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "afr",
+        "target": "spa"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-ara-deu-tatoeba-chrf2",
+      "model": "OPUS-MT ara-deu",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (ara-deu)",
+      "metric": "chrF2",
+      "value": 63,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT ara→deu: chrF2 63 on Tatoeba-test.",
+      "langs_or_pairs": "ara→deu (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ara",
+        "target": "deu"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-ara-ell-tatoeba-chrf2",
+      "model": "OPUS-MT ara-ell",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (ara-ell)",
+      "metric": "chrF2",
+      "value": 64.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT ara→ell: chrF2 64.1 on Tatoeba-test.",
+      "langs_or_pairs": "ara→ell (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ara",
+        "target": "ell"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-ara-eng-tatoeba-chrf2",
+      "model": "OPUS-MT ara-eng",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (ara-eng)",
+      "metric": "chrF2",
+      "value": 61.7,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT ara→eng: chrF2 61.7 on Tatoeba-test.",
+      "langs_or_pairs": "ara→eng (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ara",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-ara-epo-tatoeba-chrf2",
+      "model": "OPUS-MT ara-epo",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (ara-epo)",
+      "metric": "chrF2",
+      "value": 37.6,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT ara→epo: chrF2 37.6 on Tatoeba-test.",
+      "langs_or_pairs": "ara→epo (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ara",
+        "target": "epo"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-ara-fra-tatoeba-chrf2",
+      "model": "OPUS-MT ara-fra",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (ara-fra)",
+      "metric": "chrF2",
+      "value": 56.2,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT ara→fra: chrF2 56.2 on Tatoeba-test.",
+      "langs_or_pairs": "ara→fra (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ara",
+        "target": "fra"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-ara-heb-tatoeba-chrf2",
+      "model": "OPUS-MT ara-heb",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (ara-heb)",
+      "metric": "chrF2",
+      "value": 60.5,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT ara→heb: chrF2 60.5 on Tatoeba-test.",
+      "langs_or_pairs": "ara→heb (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ara",
+        "target": "heb"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-ara-ita-tatoeba-chrf2",
+      "model": "OPUS-MT ara-ita",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (ara-ita)",
+      "metric": "chrF2",
+      "value": 63.9,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT ara→ita: chrF2 63.9 on Tatoeba-test.",
+      "langs_or_pairs": "ara→ita (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ara",
+        "target": "ita"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-ara-jpn-tatoeba-chrf2",
+      "model": "OPUS-MT ara-jpn",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (ara-jpn)",
+      "metric": "chrF2",
+      "value": 19.6,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT ara→jpn: chrF2 19.6 on Tatoeba-test.",
+      "langs_or_pairs": "ara→jpn (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ara",
+        "target": "jpn"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-ara-rus-tatoeba-chrf2",
+      "model": "OPUS-MT ara-rus",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (ara-rus)",
+      "metric": "chrF2",
+      "value": 60.6,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT ara→rus: chrF2 60.6 on Tatoeba-test.",
+      "langs_or_pairs": "ara→rus (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ara",
+        "target": "rus"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-ara-tur-tatoeba-chrf2",
+      "model": "OPUS-MT ara-tur",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (ara-tur)",
+      "metric": "chrF2",
+      "value": 61.4,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT ara→tur: chrF2 61.4 on Tatoeba-test.",
+      "langs_or_pairs": "ara→tur (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ara",
+        "target": "tur"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-ben-eng-tatoeba-chrf2",
+      "model": "OPUS-MT ben-eng",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (ben-eng)",
+      "metric": "chrF2",
+      "value": 63.9,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT ben→eng: chrF2 63.9 on Tatoeba-test.",
+      "langs_or_pairs": "ben→eng (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ben",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-ber-eng-tatoeba-chrf2",
+      "model": "OPUS-MT ber-eng",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (ber-eng)",
+      "metric": "chrF2",
+      "value": 9.5,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT ber→eng: chrF2 9.5 on Tatoeba-test.",
+      "langs_or_pairs": "ber→eng (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ber",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-bre-eng-tatoeba-chrf2",
+      "model": "OPUS-MT bre-eng",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (bre-eng)",
+      "metric": "chrF2",
+      "value": 25.6,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT bre→eng: chrF2 25.6 on Tatoeba-test.",
+      "langs_or_pairs": "bre→eng (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "bre",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-bre-fra-tatoeba-chrf2",
+      "model": "OPUS-MT bre-fra",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (bre-fra)",
+      "metric": "chrF2",
+      "value": 23.3,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT bre→fra: chrF2 23.3 on Tatoeba-test.",
+      "langs_or_pairs": "bre→fra (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "bre",
+        "target": "fra"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-bul-deu-tatoeba-chrf2",
+      "model": "OPUS-MT bul-deu",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (bul-deu)",
+      "metric": "chrF2",
+      "value": 67.7,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT bul→deu: chrF2 67.7 on Tatoeba-test.",
+      "langs_or_pairs": "bul→deu (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "bul",
+        "target": "deu"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-bul-eng-tatoeba-chrf2",
+      "model": "OPUS-MT bul-eng",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (bul-eng)",
+      "metric": "chrF2",
+      "value": 72,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT bul→eng: chrF2 72 on Tatoeba-test.",
+      "langs_or_pairs": "bul→eng (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "bul",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-bul-epo-tatoeba-chrf2",
+      "model": "OPUS-MT bul-epo",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (bul-epo)",
+      "metric": "chrF2",
+      "value": 43.8,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT bul→epo: chrF2 43.8 on Tatoeba-test.",
+      "langs_or_pairs": "bul→epo (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "bul",
+        "target": "epo"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-bul-fra-tatoeba-chrf2",
+      "model": "OPUS-MT bul-fra",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (bul-fra)",
+      "metric": "chrF2",
+      "value": 69.3,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT bul→fra: chrF2 69.3 on Tatoeba-test.",
+      "langs_or_pairs": "bul→fra (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "bul",
+        "target": "fra"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-bul-ita-tatoeba-chrf2",
+      "model": "OPUS-MT bul-ita",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (bul-ita)",
+      "metric": "chrF2",
+      "value": 65.3,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT bul→ita: chrF2 65.3 on Tatoeba-test.",
+      "langs_or_pairs": "bul→ita (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "bul",
+        "target": "ita"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-bul-jpn-tatoeba-chrf2",
+      "model": "OPUS-MT bul-jpn",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (bul-jpn)",
+      "metric": "chrF2",
+      "value": 24.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT bul→jpn: chrF2 24.1 on Tatoeba-test.",
+      "langs_or_pairs": "bul→jpn (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "bul",
+        "target": "jpn"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-bul-spa-tatoeba-chrf2",
+      "model": "OPUS-MT bul-spa",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (bul-spa)",
+      "metric": "chrF2",
+      "value": 66.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT bul→spa: chrF2 66.1 on Tatoeba-test.",
+      "langs_or_pairs": "bul→spa (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "bul",
+        "target": "spa"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-bul-tur-tatoeba-chrf2",
+      "model": "OPUS-MT bul-tur",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (bul-tur)",
+      "metric": "chrF2",
+      "value": 68.6,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT bul→tur: chrF2 68.6 on Tatoeba-test.",
+      "langs_or_pairs": "bul→tur (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "bul",
+        "target": "tur"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-bul-ukr-tatoeba-chrf2",
+      "model": "OPUS-MT bul-ukr",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (bul-ukr)",
+      "metric": "chrF2",
+      "value": 68.3,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT bul→ukr: chrF2 68.3 on Tatoeba-test.",
+      "langs_or_pairs": "bul→ukr (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "bul",
+        "target": "ukr"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-cat-deu-tatoeba-chrf2",
+      "model": "OPUS-MT cat-deu",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (cat-deu)",
+      "metric": "chrF2",
+      "value": 59.3,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT cat→deu: chrF2 59.3 on Tatoeba-test.",
+      "langs_or_pairs": "cat→deu (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "cat",
+        "target": "deu"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "opus-cat-eng-tatoeba-chrf2",
+      "model": "OPUS-MT cat-eng",
+      "source": "Tatoeba-Challenge results (Helsinki-NLP)",
+      "org": "Helsinki-NLP",
+      "year": 2021,
+      "category": "model-report",
+      "benchmark": "Tatoeba-test (cat-eng)",
+      "metric": "chrF2",
+      "value": 66.8,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF2 (chrF, beta=2) on the Tatoeba-test set — 0–100, higher better, read from the Tatoeba-Challenge results tables. NOT comparable to BLEU/spBLEU, COMET/MetricX, or chrF++ from other setups.",
+      "headline": "OPUS-MT cat→eng: chrF2 66.8 on Tatoeba-test.",
+      "langs_or_pairs": "cat→eng (single directed pair).",
+      "source_url": "https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master/results/tatoeba-results-chrF2-sorted-langpair.md",
+      "citation": "Tiedemann (2020), The Tatoeba Translation Challenge; per-pair chrF from the repo results tables.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "cat",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "C",
+        "corpus_size": null,
+        "example_length": "sentence",
+        "domain_breadth": "broad",
+        "contamination": "MEDIUM",
+        "rationale": "Tatoeba-test, read from the Tatoeba-Challenge results tables. Test-set size varies by pair (often small for low-resource); general crowd-sourced sentences widely available online (MEDIUM contamination). Conservative C default for this cite-only tier."
+      }
+    },
+    {
+      "id": "nllb-amh-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (amh-eng)",
+      "metric": "chrF++",
+      "value": 59,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B amh→eng: chrF++ 59 on FLORES-200.",
+      "langs_or_pairs": "amh→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "amh",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-arb-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (arb-eng)",
+      "metric": "chrF++",
+      "value": 65.8,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B arb→eng: chrF++ 65.8 on FLORES-200.",
+      "langs_or_pairs": "arb→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "arb",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-ben-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (ben-eng)",
+      "metric": "chrF++",
+      "value": 61.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B ben→eng: chrF++ 61.1 on FLORES-200.",
+      "langs_or_pairs": "ben→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ben",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-deu-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (deu-eng)",
+      "metric": "chrF++",
+      "value": 67.4,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B deu→eng: chrF++ 67.4 on FLORES-200.",
+      "langs_or_pairs": "deu→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "deu",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-ell-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (ell-eng)",
+      "metric": "chrF++",
+      "value": 62,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B ell→eng: chrF++ 62 on FLORES-200.",
+      "langs_or_pairs": "ell→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ell",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-fra-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (fra-eng)",
+      "metric": "chrF++",
+      "value": 68.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B fra→eng: chrF++ 68.1 on FLORES-200.",
+      "langs_or_pairs": "fra→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "fra",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-hin-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (hin-eng)",
+      "metric": "chrF++",
+      "value": 65.9,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B hin→eng: chrF++ 65.9 on FLORES-200.",
+      "langs_or_pairs": "hin→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "hin",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-isl-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (isl-eng)",
+      "metric": "chrF++",
+      "value": 56.6,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B isl→eng: chrF++ 56.6 on FLORES-200.",
+      "langs_or_pairs": "isl→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "isl",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-ita-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (ita-eng)",
+      "metric": "chrF++",
+      "value": 61.2,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B ita→eng: chrF++ 61.2 on FLORES-200.",
+      "langs_or_pairs": "ita→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ita",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-jpn-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (jpn-eng)",
+      "metric": "chrF++",
+      "value": 55.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B jpn→eng: chrF++ 55.1 on FLORES-200.",
+      "langs_or_pairs": "jpn→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "jpn",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-kor-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (kor-eng)",
+      "metric": "chrF++",
+      "value": 56.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B kor→eng: chrF++ 56.1 on FLORES-200.",
+      "langs_or_pairs": "kor→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "kor",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-nld-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (nld-eng)",
+      "metric": "chrF++",
+      "value": 59,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B nld→eng: chrF++ 59 on FLORES-200.",
+      "langs_or_pairs": "nld→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "nld",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-pol-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (pol-eng)",
+      "metric": "chrF++",
+      "value": 57,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B pol→eng: chrF++ 57 on FLORES-200.",
+      "langs_or_pairs": "pol→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "pol",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-por-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (por-eng)",
+      "metric": "chrF++",
+      "value": 71.3,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B por→eng: chrF++ 71.3 on FLORES-200.",
+      "langs_or_pairs": "por→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "por",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-ron-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (ron-eng)",
+      "metric": "chrF++",
+      "value": 68.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B ron→eng: chrF++ 68.1 on FLORES-200.",
+      "langs_or_pairs": "ron→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ron",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-rus-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (rus-eng)",
+      "metric": "chrF++",
+      "value": 61.3,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B rus→eng: chrF++ 61.3 on FLORES-200.",
+      "langs_or_pairs": "rus→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "rus",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-spa-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (spa-eng)",
+      "metric": "chrF++",
+      "value": 59.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B spa→eng: chrF++ 59.1 on FLORES-200.",
+      "langs_or_pairs": "spa→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "spa",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-swh-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (swh-eng)",
+      "metric": "chrF++",
+      "value": 65,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B swh→eng: chrF++ 65 on FLORES-200.",
+      "langs_or_pairs": "swh→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "swh",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-tur-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (tur-eng)",
+      "metric": "chrF++",
+      "value": 63.9,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B tur→eng: chrF++ 63.9 on FLORES-200.",
+      "langs_or_pairs": "tur→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "tur",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-ukr-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (ukr-eng)",
+      "metric": "chrF++",
+      "value": 64.2,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B ukr→eng: chrF++ 64.2 on FLORES-200.",
+      "langs_or_pairs": "ukr→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "ukr",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-vie-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (vie-eng)",
+      "metric": "chrF++",
+      "value": 61.5,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B vie→eng: chrF++ 61.5 on FLORES-200.",
+      "langs_or_pairs": "vie→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "vie",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-yor-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (yor-eng)",
+      "metric": "chrF++",
+      "value": 44.5,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B yor→eng: chrF++ 44.5 on FLORES-200.",
+      "langs_or_pairs": "yor→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "yor",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-zho-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (zho-eng)",
+      "metric": "chrF++",
+      "value": 56.2,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B zho→eng: chrF++ 56.2 on FLORES-200.",
+      "langs_or_pairs": "zho→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "zho",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-zul-eng-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (zul-eng)",
+      "metric": "chrF++",
+      "value": 60.9,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B zul→eng: chrF++ 60.9 on FLORES-200.",
+      "langs_or_pairs": "zul→eng (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "zul",
+        "target": "eng"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-arb-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-arb)",
+      "metric": "chrF++",
+      "value": 55,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→arb: chrF++ 55 on FLORES-200.",
+      "langs_or_pairs": "eng→arb (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "arb"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-ben-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-ben)",
+      "metric": "chrF++",
+      "value": 48.7,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→ben: chrF++ 48.7 on FLORES-200.",
+      "langs_or_pairs": "eng→ben (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "ben"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-deu-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-deu)",
+      "metric": "chrF++",
+      "value": 62.8,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→deu: chrF++ 62.8 on FLORES-200.",
+      "langs_or_pairs": "eng→deu (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "deu"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-fra-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-fra)",
+      "metric": "chrF++",
+      "value": 69.6,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→fra: chrF++ 69.6 on FLORES-200.",
+      "langs_or_pairs": "eng→fra (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "fra"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-hin-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-hin)",
+      "metric": "chrF++",
+      "value": 57,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→hin: chrF++ 57 on FLORES-200.",
+      "langs_or_pairs": "eng→hin (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "hin"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-ita-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-ita)",
+      "metric": "chrF++",
+      "value": 57.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→ita: chrF++ 57.1 on FLORES-200.",
+      "langs_or_pairs": "eng→ita (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "ita"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-jpn-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-jpn)",
+      "metric": "chrF++",
+      "value": 25.2,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→jpn: chrF++ 25.2 on FLORES-200.",
+      "langs_or_pairs": "eng→jpn (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "jpn"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-kor-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-kor)",
+      "metric": "chrF++",
+      "value": 34.3,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→kor: chrF++ 34.3 on FLORES-200.",
+      "langs_or_pairs": "eng→kor (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "kor"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-nld-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-nld)",
+      "metric": "chrF++",
+      "value": 54.9,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→nld: chrF++ 54.9 on FLORES-200.",
+      "langs_or_pairs": "eng→nld (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "nld"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-por-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-por)",
+      "metric": "chrF++",
+      "value": 69.4,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→por: chrF++ 69.4 on FLORES-200.",
+      "langs_or_pairs": "eng→por (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "por"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-rus-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-rus)",
+      "metric": "chrF++",
+      "value": 56.1,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→rus: chrF++ 56.1 on FLORES-200.",
+      "langs_or_pairs": "eng→rus (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "rus"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-spa-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-spa)",
+      "metric": "chrF++",
+      "value": 54.2,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→spa: chrF++ 54.2 on FLORES-200.",
+      "langs_or_pairs": "eng→spa (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "spa"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-swh-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-swh)",
+      "metric": "chrF++",
+      "value": 60,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→swh: chrF++ 60 on FLORES-200.",
+      "langs_or_pairs": "eng→swh (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "swh"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-tur-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-tur)",
+      "metric": "chrF++",
+      "value": 57.8,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→tur: chrF++ 57.8 on FLORES-200.",
+      "langs_or_pairs": "eng→tur (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "tur"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-vie-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-vie)",
+      "metric": "chrF++",
+      "value": 59.3,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→vie: chrF++ 59.3 on FLORES-200.",
+      "langs_or_pairs": "eng→vie (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "vie"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    },
+    {
+      "id": "nllb-eng-zho-flores-chrfpp",
+      "model": "NLLB-200 3.3B",
+      "source": "NLLB-200 published FLORES-200 metrics",
+      "org": "Meta AI",
+      "year": 2022,
+      "category": "model-report",
+      "benchmark": "FLORES-200 devtest (eng-zho)",
+      "metric": "chrF++",
+      "value": 22.3,
+      "lower_is_better": false,
+      "metric_variant_flag": "chrF++ on FLORES-200 devtest — 0–100, higher better, from the NLLB-200 3.3B public metrics. NOT comparable to BLEU, COMET, or MetricX.",
+      "headline": "NLLB-200 3.3B eng→zho: chrF++ 22.3 on FLORES-200.",
+      "langs_or_pairs": "eng→zho (single directed pair).",
+      "source_url": "https://dl.fbaipublicfiles.com/large_objects/nllb/models/nllb_200_dense_3b/metrics.csv",
+      "citation": "NLLB Team et al. (2022), No Language Left Behind, arXiv:2207.04672; FLORES-200 devtest chrF++ from the released metrics.",
+      "verified": true,
+      "badge": "unverified — not reproduced by Champollion",
+      "notes": null,
+      "pair": {
+        "source": "eng",
+        "target": "zho"
+      },
+      "signal_strength": {
+        "grade": "B",
+        "corpus_size": 1012,
+        "example_length": "sentence",
+        "domain_breadth": "multi-domain",
+        "contamination": "MEDIUM",
+        "rationale": "FLORES-200 devtest = 1012 professionally-translated, multi-parallel sentences (multi-domain). Strong standard benchmark, but a widely-used public set (MEDIUM contamination) → B."
+      }
+    }
+  ],
+  "manifest": {
+    "run": [
+      {
+        "family": "flores",
+        "name": "FLORES-200 / FLORES+",
+        "license": "CC-BY-SA-4.0",
+        "contamination_posture": "HIGH — relative-only (illustrative, never an absolute quality claim)",
+        "note": "Many-to-many devtest used for breadth; the most broadly trained-on suite, so scored on the relative-only lane and used to illustrate coverage, not to rank absolute quality."
+      },
+      {
+        "family": "tico19",
+        "name": "TICO-19 (COVID-19 translation)",
+        "license": "CC0-1.0",
+        "contamination_posture": "varies — eligibility-gated per dataset",
+        "note": "Public-health domain parallel set fetched from source; widely available so treated cautiously by the contamination lane."
+      },
+      {
+        "family": "in22",
+        "name": "IN22 (Indian languages)",
+        "license": "CC-BY-4.0",
+        "contamination_posture": "varies — eligibility-gated per dataset",
+        "note": "Conversational + general Indic evaluation pairs (gated download; fetch-from-source)."
+      },
+      {
+        "family": "tatoeba",
+        "name": "Tatoeba clean bridges",
+        "license": "CC-BY-2.0",
+        "contamination_posture": "varies — eligibility-gated per dataset",
+        "note": "Per-pair clean bridges rebuilt fetch-from-source (Challenge tar + OPUS moses), including the flagship spa→que Quechua bridge."
+      },
+      {
+        "family": "globalvoices",
+        "name": "GlobalVoices news",
+        "license": "CC-BY-3.0",
+        "contamination_posture": "varies — eligibility-gated per dataset",
+        "note": "Citizen-media news parallel corpus, fetch-from-source."
+      },
+      {
+        "family": "smol",
+        "name": "SMOL (sentence + document)",
+        "license": "CC-BY-4.0",
+        "contamination_posture": "varies — eligibility-gated per dataset",
+        "note": "Low-resource sentence- and document-level pairs, per-pair sha-pinned."
+      },
+      {
+        "family": "alt",
+        "name": "ALT (Asian Language Treebank)",
+        "license": "CC-BY-4.0",
+        "contamination_posture": "varies — eligibility-gated per dataset",
+        "note": "Multi-way SEA evaluation (Lao, Khmer, Burmese, Filipino, Malay…); only the bare parallel layer (CC-BY-4.0) is used, not the NC treebank annotations."
+      },
+      {
+        "family": "turkicxwmt",
+        "name": "TurkicX-WMT",
+        "license": "CC-BY-SA-4.0",
+        "contamination_posture": "varies — eligibility-gated per dataset",
+        "note": "Turkic-language evaluation pairs with non-uniform per-pair sizes (data-driven size map)."
+      },
+      {
+        "family": "wmt24pp",
+        "name": "WMT24++ ",
+        "license": "Apache-2.0",
+        "contamination_posture": "varies — eligibility-gated per dataset",
+        "note": "Per-pair built corpus, sha-pinned; out-of-English directions. (Same benchmark family the cited TranslateGemma WMT24++ numbers are reported on — but those numbers are the authors', not a Champollion run.)"
+      },
+      {
+        "family": "gamayun",
+        "name": "Gamayun kit (TWB)",
+        "license": "LicenseRef-TWB-Gamayun",
+        "contamination_posture": "non-commercial research lane",
+        "note": "Translators-without-Borders humanitarian mini-kits; carved into the non-commercial research lane."
+      },
+      {
+        "family": "prize",
+        "name": "Prize / contest sets",
+        "license": "varies",
+        "contamination_posture": "held-out where applicable",
+        "note": "Contest-specific evaluation corpora."
+      }
+    ],
+    "cite": [
+      {
+        "source": "TranslateGemma Technical Report",
+        "result_ids": [
+          "translategemma-27b-metricx",
+          "translategemma-27b-comet22",
+          "translategemma-12b-metricx",
+          "translategemma-12b-comet22",
+          "translategemma-4b-metricx",
+          "translategemma-4b-comet22",
+          "gemma3-27b-metricx",
+          "gemma3-27b-comet22",
+          "gemma3-12b-metricx",
+          "gemma3-12b-comet22",
+          "gemma3-4b-metricx",
+          "gemma3-4b-comet22"
+        ],
+        "note": "Google's open MT suite fine-tuned from Gemma 3 (55 languages). We cite the verified WMT24++ MetricX-24 / COMET-22 numbers for TranslateGemma 4B/12B/27B and their Gemma 3 baselines (see the External results tab); we have not re-run the models."
+      },
+      {
+        "source": "Gemma 2 (general-purpose open LLM)",
+        "note": "A general open-weights LLM (Gemma Team, arXiv:2403.08295 / Gemma 2, arXiv:2408.00118) whose translation ability is incidental and reported inside a broad capability suite — NOT a dedicated MT system, and NOT to be confused with TranslateGemma (a separate Google MT model fine-tuned from Gemma 3). Cited for context; no dedicated-MT single number is pinned because the source publishes none on a stated MT protocol."
+      },
+      {
+        "source": "Frontier and open MT model reports",
+        "note": "Model papers/cards whose translation results we point to but have NOT pinned to a single verified datapoint here (different metric variants and test suites, or no stated MT protocol): MADLAD-400 MT models (arXiv:2309.04662), SeamlessM4T (arXiv:2308.11596), Aya / Aya 23 (arXiv:2402.07827 / 2405.15032), GPT-4 (arXiv:2303.08774), Gemini (arXiv:2312.11805), Claude 3 (model card). Cited as the field's reference systems; numbers are not pooled and not reproduced. (NLLB-200 per-pair numbers ARE now pinned — see its own entry.)"
+      },
+      {
+        "source": "WMT General MT shared task",
+        "note": "The field's reference human-ranked event (Kocmi et al., WMT24, https://www2.statmt.org/wmt24/); cited, not re-run — the test sets are broadly trained on. Human ESA/MQM is the official ranking; automatic COMET/MetricX/chrF are version-specific and not pooled."
+      },
+      {
+        "source": "Papers with Code — MT leaderboards",
+        "note": "Community aggregator of self-reported SOTA per benchmark (https://paperswithcode.com/task/machine-translation); predominantly legacy tokenized BLEU with inconsistent tooling — a literature map, not a like-for-like ranking."
+      },
+      {
+        "source": "OPUS-MT / Tatoeba-Challenge leaderboard",
+        "note": "Open, reproducible neural-MT models + BLEU/chrF benchmark cards across a very large number of pairs incl. low-resource directions (Tiedemann 2020; https://github.com/Helsinki-NLP/OPUS-MT-leaderboard). We pin a sample of VERIFIED per-pair card numbers (see the External results tab) spanning strong high-resource edges and weak low-resource ones; Champollion has not re-run them and runs its own clean Tatoeba bridges fetch-from-source rather than re-hosting.",
+        "result_ids": [
+          "opus-en-fr-tatoeba-bleu",
+          "opus-en-fr-tatoeba-chrf",
+          "opus-en-fr-newstest2013-bleu",
+          "opus-en-es-tatoeba-bleu",
+          "opus-en-es-tatoeba-chrf",
+          "opus-en-de-tatoeba-bleu",
+          "opus-en-de-tatoeba-chrf",
+          "opus-en-de-newstest2019-bleu",
+          "opus-en-ru-tatoeba-bleu",
+          "opus-en-ru-tatoeba-chrf",
+          "opus-tcbig-en-tr-tatoeba-bleu",
+          "opus-tcbig-en-tr-tatoeba-chrf",
+          "opus-tcbig-en-tr-flores-bleu",
+          "opus-tcbig-en-tr-flores-chrf",
+          "opus-en-ha-tatoeba-bleu",
+          "opus-ha-en-tatoeba-bleu",
+          "opus-en-mul-amh-bleu",
+          "opus-mul-en-amh-bleu",
+          "opus-en-mul-uig-bleu",
+          "opus-mul-en-uig-bleu",
+          "opus-en-mul-sah-bleu",
+          "opus-mul-en-sah-bleu",
+          "opus-en-trk-kaz-bleu"
+        ]
+      },
+      {
+        "source": "NLLB-200 (No Language Left Behind)",
+        "result_ids": [
+          "nllb-eng-hau-flores-spbleu",
+          "nllb-eng-hau-flores-chrfpp",
+          "nllb-hau-eng-flores-chrfpp",
+          "nllb-eng-amh-tico-bleu",
+          "nllb-eng-amh-tico-chrfpp",
+          "nllb-amh-eng-tico-chrfpp",
+          "nllb-flores101-avg-spbleu",
+          "nllb-flores101-avg-chrfpp"
+        ],
+        "note": "Meta's 200-language open model (arXiv:2207.04672). We pin VERIFIED per-pair numbers read from the paper: FLORES-101 eng↔hau (relative-only, FLORES is HIGH-contamination) and the TICO health-domain eng↔amh directions, plus the FLORES-101 101-language average. Weights are CC-BY-NC-4.0 (non-commercial) — see the method index. Champollion has not re-run NLLB-200; numbers are the authors'."
+      }
+    ],
+    "exclude": [
+      {
+        "name": "Bible corpora / JW300",
+        "reason_code": "contaminated-colonial",
+        "reason": "Broadly memorized by modern models (high contamination) AND rooted in colonial/missionary translation programs; excluded on both integrity and sovereignty grounds — per the data-boundaries doctrine."
+      },
+      {
+        "name": "OPUS-100 (wholesale)",
+        "reason_code": "mixed-license",
+        "reason": "A grab-bag of sub-corpora under mixed and sometimes unclear licenses; we instead pull specific, license-clean pairs fetch-from-source rather than ingesting the whole bundle."
+      },
+      {
+        "name": "MTEB / BUCC",
+        "reason_code": "embedding-not-mt",
+        "reason": "Sentence-embedding / bitext-mining benchmarks, not machine-translation generation — out of scope for a translation-quality index."
+      },
+      {
+        "name": "MADLAD-400 corpus",
+        "reason_code": "monolingual-training",
+        "reason": "A monolingual training corpus (419 languages), not a parallel evaluation set. We cite the MADLAD-400 MT MODEL results but do not index the corpus as a benchmark."
+      },
+      {
+        "name": "FLEURS",
+        "reason_code": "speech-out-of-scope",
+        "reason": "A speech (ASR/speech-translation) benchmark; Champollion's Network indexes text machine translation, so speech corpora are out of scope."
+      },
+      {
+        "name": "EdTeKLA easy crk subsets (62 / 90-phase1 / 124 / crk-master)",
+        "reason_code": "improper-subset",
+        "reason": "Small, easy slices that must never rank as if they were full held-out evaluations — enforced by the datasets.quarantined flag + DB trigger. The full EdTeKLA protocol stays in the non-commercial research lane and corpus CONTENT is never hosted."
+      },
+      {
+        "name": "Wolvengrey Cree dictionary data",
+        "reason_code": "no-redistribute",
+        "reason": "Redistribution permission is pending; never copied into corpora, cards, or exports."
+      }
+    ]
+  },
+  "methods": [
+    {
+      "id": "opus-mt",
+      "name": "OPUS-MT",
+      "org": "Helsinki-NLP",
+      "paradigm": "neural-mt",
+      "availability": "open-weights",
+      "weights": "open",
+      "access": [
+        "huggingface",
+        "download",
+        "pip",
+        "github"
+      ],
+      "license": "CC-BY-4.0 / Apache-2.0 (per model)",
+      "commercial_use": true,
+      "homepage": "https://github.com/Helsinki-NLP/Opus-MT",
+      "weights_url": "https://huggingface.co/Helsinki-NLP",
+      "source_url": "https://github.com/Helsinki-NLP/Opus-MT",
+      "runnable_in_champollion": false,
+      "notes": "1000+ open Marian-NMT models (single-pair + multilingual). Weights freely downloadable via HF (MarianMTModel) or .zip from object.pouta.csc.fi. Not yet a first-class Champollion method — reachable today by self-hosting behind the local/OpenAI-compatible lane; a native adapter is implement-for-them later."
+    },
+    {
+      "id": "nllb-200",
+      "name": "NLLB-200",
+      "org": "Meta AI",
+      "paradigm": "neural-mt",
+      "availability": "open-weights",
+      "weights": "open",
+      "access": [
+        "huggingface",
+        "download",
+        "github"
+      ],
+      "license": "CC-BY-NC-4.0",
+      "commercial_use": false,
+      "homepage": "https://github.com/facebookresearch/fairseq/tree/nllb",
+      "weights_url": "https://huggingface.co/facebook/nllb-200-3.3B",
+      "source_url": "https://arxiv.org/abs/2207.04672",
+      "runnable_in_champollion": false,
+      "notes": "200-language single model; weights openly downloadable (facebook/nllb-200-3.3B, -distilled-600M) but the CC-BY-NC-4.0 license is NON-COMMERCIAL — carved out of any commercial lane. Research model, not for production deployment."
+    },
+    {
+      "id": "translategemma",
+      "name": "TranslateGemma",
+      "org": "Google",
+      "paradigm": "dedicated-mt",
+      "availability": "open-weights",
+      "weights": "open",
+      "access": [
+        "huggingface",
+        "download"
+      ],
+      "license": "Gemma Terms of Use",
+      "commercial_use": true,
+      "homepage": "https://ai.google.dev/gemma",
+      "weights_url": "https://huggingface.co/google",
+      "source_url": "https://arxiv.org/abs/2601.09012",
+      "runnable_in_champollion": false,
+      "notes": "Dedicated MT suite fine-tuned from Gemma 3 (4B/12B/27B), 55 languages. Open weights under the Gemma license. Cited-only here; a native adapter is implement-for-them later."
+    },
+    {
+      "id": "gemma-3",
+      "name": "Gemma 3",
+      "org": "Google",
+      "paradigm": "llm",
+      "availability": "open-weights",
+      "weights": "open",
+      "access": [
+        "huggingface",
+        "download",
+        "api"
+      ],
+      "license": "Gemma Terms of Use",
+      "commercial_use": true,
+      "homepage": "https://ai.google.dev/gemma",
+      "weights_url": "https://huggingface.co/google",
+      "source_url": "https://arxiv.org/abs/2601.09012",
+      "runnable_in_champollion": false,
+      "notes": "General-purpose open LLM; its MT scores here are the pre-fine-tuning baseline reported in the TranslateGemma study. As a general LLM it is reachable via OpenRouter/local lanes, but not registered as a dedicated MT method."
+    }
+  ]
+}