champollion 0.3.3

package/shared/metric-registry.json

@@ -0,0 +1,620 @@
+{
+  "_comment": "SSOT for METRIC IDENTITY, consumed by the Python arena harness (arena/mt_eval_harness) and available to the JS CLI/website. One metric has up to four coordinated names: the canonical_id (this file's key = the run-card scores key), the Python MetricPlugin 'name' attribute that computes it (plugin_name), the language-card evalMetrics key that declares it (card_key), and the denormalized run_cards DB/leaderboard column (db_column). This file maps them so no consumer ever guesses. Definitions come from the scoring spec (cli/website/docs/network/specifications/scoring.md §2) and scoring.py; arena/tests/test_metric_registry_ssot.py FAILS if scoring.py weight tables or publish.py run-card score keys drift from this file. verifier_reproducible means: deterministically re-derivable by the verifier from the sha-pinned corpus + stored entries (+ card-pinned tools like the FST, or the pinned neural model under the fail-closed contract). in_composite means: currently enters at least one composite profile in scoring.PROFILE_REGISTRY when available (declared-but-INACTIVE metrics are false). Confidence-interval columns (chrf_ci_lower/...), boolean flags (has_references, morph_in_composite), model-id fields (comet_model, qe_model, metricx_model) and counts (corpus_size, total, evaluated, errors) are metadata ABOUT metrics, not metrics — they are deliberately not entries here.",
+  "version": 1,
+  "entries": {
+    "exact_match_rate": {
+      "category": "surface",
+      "status": "implemented",
+      "display_name": "Exact Match",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "exact_match_rate",
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": true,
+      "verifier_reproducible": true,
+      "notes": "Harness core (tester.py), no plugin. Binary predicted == reference; corpus rate = matches/total."
+    },
+    "equivalent_match_rate": {
+      "category": "surface",
+      "status": "partial",
+      "display_name": "Equivalent Match",
+      "plugin_name": "crk_linter",
+      "card_key": "lyss-eq",
+      "db_column": "equivalent_match_rate",
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": true,
+      "verifier_reproducible": true,
+      "notes": "CRK proxy today (champollion_lyss.crk.metrics.CrkLinterMetric, declared on the crk card as 'lyss-eq'). publish.py discovers it by the is_equivalence_linter aggregate flag, not by name — any language's linter plugin feeds this same canonical metric."
+    },
+    "chrf_plus_plus": {
+      "category": "surface",
+      "status": "implemented",
+      "display_name": "chrF++",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "chrf_plus_plus",
+      "scale": "0-100",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": true,
+      "verifier_reproducible": true,
+      "notes": "sacrebleu (word_order=2), harness core. Normalized /100 for the composite (scoring.NORMALIZATIONS)."
+    },
+    "bleu": {
+      "category": "surface",
+      "status": "implemented",
+      "display_name": "BLEU",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "corpus_bleu",
+      "scale": "0-100",
+      "direction": "higher",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "sacrebleu, harness core. Reported for MT-literature compatibility, never composited. Run-card key is top-level 'corpus_bleu' (not scores.bleu) — see _proposed_renames."
+    },
+    "ter": {
+      "category": "surface",
+      "status": "implemented",
+      "display_name": "Translation Edit Rate",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "ter",
+      "scale": "0-inf",
+      "direction": "lower",
+      "level": "both",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "sacrebleu corpus_ter. Excluded from composite (correlates with chrF++)."
+    },
+    "length_ratio": {
+      "category": "surface",
+      "status": "implemented",
+      "display_name": "Length Ratio",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "length_ratio",
+      "scale": "0-inf (ideal 1.0)",
+      "direction": "neutral",
+      "level": "both",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "Diagnostic: len(predicted)/len(reference); <0.5 truncation, >2.0 inflation."
+    },
+    "fst_acceptance_rate": {
+      "category": "structural",
+      "status": "implemented",
+      "display_name": "FST Acceptance",
+      "plugin_name": "giellalt_fst_validity",
+      "card_key": null,
+      "db_column": "fst_acceptance_rate",
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": true,
+      "verifier_reproducible": true,
+      "notes": "GiellaLTFSTMetric (plugins/giellalt_fst.py). Declared per-language via the card's resources.fsts install metadata, not evalMetrics. Plugin aggregate key is 'avg_fst_validity' (legacy standalone reports used 'acceptance_rate')."
+    },
+    "morphological_accuracy": {
+      "category": "structural",
+      "status": "implemented",
+      "display_name": "Morphological Accuracy",
+      "plugin_name": "giellalt_fst_validity",
+      "card_key": null,
+      "db_column": "morphological_accuracy",
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": true,
+      "verifier_reproducible": true,
+      "notes": "FST-derived, lemma-matched (same plugin as fst_acceptance_rate). Enters the fst-coverage composite only when morph_coverage >= MORPH_COVERAGE_FLOOR (0.25); verifier re-derives via recompute_corpus_morph."
+    },
+    "morph_coverage": {
+      "category": "structural",
+      "status": "implemented",
+      "display_name": "Morphological Coverage",
+      "plugin_name": "giellalt_fst_validity",
+      "card_key": null,
+      "db_column": "morph_coverage",
+      "scale": "0.0-1.0",
+      "direction": "neutral",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "Companion disclosure for morphological_accuracy (fraction of analyzable predicted words lemma-matched); a coverage figure, not a quality score."
+    },
+    "orthographic_accuracy": {
+      "category": "structural",
+      "status": "planned",
+      "display_name": "Orthographic Accuracy",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": null,
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": "Carries a DECLARED 0.05 weight in the surface-only profile but is in scoring.INACTIVE_METRICS (needs per-language orthographic rule sets) — never scores yet."
+    },
+    "gloss_word_accuracy": {
+      "category": "structural",
+      "status": "implemented",
+      "display_name": "IGT Gloss Word Accuracy",
+      "plugin_name": "igt_gloss",
+      "card_key": null,
+      "db_column": null,
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "SIGMORPHON-2023 glossing shared-task accuracy (plugins/igt_gloss.py, IGTGlossMetric). LANGUAGE-NEUTRAL Layer-1 lane metric for the interlinear-gloss TASK — not MT scoring; shipped 2026-07-04 (COMPUTEL). Headline is word-level positional accuracy; the plugin also emits companion keys gloss_morpheme_accuracy (+ _average_/_overall_), and stem/gram precision/recall/F1 (gloss_classes). ACTIVATION-PENDING: any language activates it by declaring module mt_eval_harness.plugins.igt_gloss on its card's evalMetrics — no card does yet, so it lives in plugin_metrics JSONB with no run_cards column. Deterministic (positional gold comparison), verifier-re-derivable."
+    },
+    "seg_micro_f1": {
+      "category": "structural",
+      "status": "implemented",
+      "display_name": "Morpheme Segmentation F1 (micro)",
+      "plugin_name": "morph_segmentation",
+      "card_key": null,
+      "db_column": null,
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "SIGMORPHON-style morpheme-segmentation F1 (plugins/morph_segmentation.py, MorphSegmentationMetric). LANGUAGE-NEUTRAL Layer-1 lane metric for the segmentation TASK — not MT scoring; shipped 2026-07-04 (COMPUTEL). Headline is corpus micro F1; the plugin also emits seg_micro_precision/recall, seg_macro_f1, and per-entry seg_precision/recall/f1. ACTIVATION-PENDING: declared per-card via evalMetrics module mt_eval_harness.plugins.morph_segmentation — no card does yet, so it lives in plugin_metrics JSONB with no run_cards column. Deterministic multiset match, verifier-re-derivable."
+    },
+    "semantic_score": {
+      "category": "semantic",
+      "status": "partial",
+      "display_name": "Semantic Similarity",
+      "plugin_name": "crk_semantic",
+      "card_key": "lyss-sem",
+      "db_column": "semantic_score",
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": true,
+      "verifier_reproducible": true,
+      "notes": "CRK proxy today (champollion_lyss.crk.metrics.CrkSemanticMetric, declared on the crk card as 'lyss-sem'). publish.py discovers it by the semantic_verdict_counts aggregate field and applies the verdict weights defined there."
+    },
+    "comet_score": {
+      "category": "neural",
+      "status": "implemented",
+      "display_name": "COMET / AfriCOMET",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "comet_score",
+      "scale": "~0.0-1.0",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "Harness core (metrics_comet.py); model auto-selected via the card's metricModelSupport. NEURAL — reported in the separate lane, never composited; verifier re-derives fail-closed (recompute_corpus_comet)."
+    },
+    "qe_score": {
+      "category": "neural",
+      "status": "implemented",
+      "display_name": "Reference-Free QE (AfriCOMET-QE)",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "qe_score",
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "Neural reference-free lane for no-reference corpora; verifier re-derives (recompute_corpus_qe)."
+    },
+    "metricx_score": {
+      "category": "neural",
+      "status": "implemented",
+      "display_name": "MetricX-24",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": null,
+      "scale": "0-25",
+      "direction": "lower",
+      "level": "both",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": "LOWER-IS-BETTER error score (the WMT24++/TranslateGemma reference metric). Lives in run_card JSONB only (dedicated column is a follow-up); not yet re-derived by the verifier."
+    },
+    "code_switching_rate": {
+      "category": "behavioral",
+      "status": "implemented",
+      "display_name": "Code-Switching Rate",
+      "plugin_name": "code_switching",
+      "card_key": null,
+      "db_column": "code_switching_rate",
+      "scale": "0.0-1.0",
+      "direction": "lower",
+      "level": "both",
+      "in_composite": true,
+      "verifier_reproducible": true,
+      "notes": "CodeSwitchingPlugin (language-agnostic, always loaded). Inverted (1-x) for the composite. Plugin aggregate key: avg_code_switching_rate."
+    },
+    "hallucination_rate": {
+      "category": "behavioral",
+      "status": "implemented",
+      "display_name": "Hallucination Rate",
+      "plugin_name": "hallucination",
+      "card_key": null,
+      "db_column": "hallucination_rate",
+      "scale": "0.0-1.0",
+      "direction": "lower",
+      "level": "both",
+      "in_composite": true,
+      "verifier_reproducible": true,
+      "notes": "HallucinationPlugin (language-agnostic, always loaded). Inverted (1-x) for the composite. Plugin aggregate key: avg_hallucination_rate."
+    },
+    "terminology_adherence": {
+      "category": "behavioral",
+      "status": "implemented",
+      "display_name": "Terminology Adherence",
+      "plugin_name": "terminology",
+      "card_key": null,
+      "db_column": "terminology_adherence",
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": true,
+      "verifier_reproducible": true,
+      "notes": "TerminologyPlugin; null (metric inactive) when the run supplies no coaching glossary. Reproducible given the run's glossary provenance. Plugin aggregate key: avg_terminology_adherence."
+    },
+    "consistency_score": {
+      "category": "behavioral",
+      "status": "planned",
+      "display_name": "Cross-Entry Consistency",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": null,
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": "Specified in scoring spec §2.4, not yet implemented."
+    },
+    "compliance_index": {
+      "category": "compliance",
+      "status": "implemented",
+      "display_name": "Double-Pass Compliance",
+      "plugin_name": "double_pass_compliance",
+      "card_key": null,
+      "db_column": null,
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "DoublePassCompliancePlugin. Quality GATE (placeholder/quote/casing integrity), not a quality score; lives in plugin_metrics/report only, no run_cards column."
+    },
+    "repair_effectiveness": {
+      "category": "compliance",
+      "status": "implemented",
+      "display_name": "Repair Effectiveness",
+      "plugin_name": "double_pass_compliance",
+      "card_key": null,
+      "db_column": null,
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "Fraction of compliance violations auto-repaired by post-translation hooks; same plugin as compliance_index."
+    },
+    "spbleu": {
+      "category": "comparator",
+      "status": "implemented",
+      "display_name": "spBLEU (FLORES-200)",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": null,
+      "scale": "0-100",
+      "direction": "higher",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "Comparability sidecar (FLORES/NLLB lingua-franca tokenizer). JSONB only."
+    },
+    "chrf_plain": {
+      "category": "comparator",
+      "status": "implemented",
+      "display_name": "Plain chrF (word_order=0)",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": null,
+      "scale": "0-100",
+      "direction": "higher",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "The chrF figure FLORES/WMT tables report. JSONB only."
+    },
+    "fuse_score": {
+      "category": "comparator",
+      "status": "partial",
+      "display_name": "FUSE-style Comparator (untrained)",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": null,
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": "Opt-in (--fuse); UNTRAINED reimplementation of AmericasNLP-2025 FUSE (Raja & Vats), flagged fuse_untrained=true; needs the 'fuse' extra (LaBSE). Not re-derived by the verifier."
+    },
+    "linted_chrf": {
+      "category": "comparator",
+      "status": "partial",
+      "display_name": "LYSS-normalized chrF++",
+      "plugin_name": "crk_linted_chrf",
+      "card_key": "lyss-chrf",
+      "db_column": null,
+      "scale": "0-100",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "CRK proxy today: the third LYSS plugin (champollion_lyss.crk.metrics.CrkLintedChrF), sibling to crk_linter/crk_semantic. chrF++ where lint-verdict EXACT/EQUIVALENT pairs score 100 and MISS pairs fall back to canonical chrF++ — NOT comparable to published chrF++ numbers (disclosed in the class docstring). Declared on crk.json as 'lyss-chrf' (activated 2026-07-07; runs before that date never computed it). No run_cards column — comparator, plugin_metrics JSONB only, never composited. Deterministic (LYSS lint + sacrebleu chrF), re-derivable given the pinned LYSS release."
+    },
+    "style_consistency_rate": {
+      "category": "behavioral",
+      "status": "implemented",
+      "display_name": "Writing Style Consistency",
+      "plugin_name": "writing_style",
+      "card_key": null,
+      "db_column": "style_consistency_rate",
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "both",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": "WritingStyleMetric — informational only (register/length/formality alignment); profile may be auto-detected from the corpus, so not verifier-contracted."
+    },
+    "composite": {
+      "category": "composite",
+      "status": "implemented",
+      "display_name": "Composite Score (experimental)",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "composite_score",
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "Derived by scoring.compute_composite_score from the profile weight tables (SSOT: scoring spec §4.3). A convenience sort key, NOT a validated quality measurement."
+    },
+    "cost_adjusted": {
+      "category": "composite",
+      "status": "implemented",
+      "display_name": "Cost-Adjusted Score",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": null,
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "scoring.cost_adjusted_score (composite / log2(1 + cost*1000), penalty-only). JSONB only."
+    },
+    "quality_tier": {
+      "category": "composite",
+      "status": "implemented",
+      "display_name": "Quality Tier",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "quality_tier",
+      "scale": "label (baseline/emerging/functional/deployable/fluent/unscored)",
+      "direction": "neutral",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "Heuristic label on the composite (scoring.QUALITY_TIERS); only human review confirms usability."
+    },
+    "tokens_per_second": {
+      "category": "efficiency",
+      "status": "implemented",
+      "display_name": "Tokens / Second",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "tokens_per_second",
+      "scale": "0-inf",
+      "direction": "higher",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": "Speed metric (spec §7) — derived from RunLog timing; environment-dependent, never re-derived."
+    },
+    "entries_per_minute": {
+      "category": "efficiency",
+      "status": "implemented",
+      "display_name": "Entries / Minute",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "entries_per_minute",
+      "scale": "0-inf",
+      "direction": "higher",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": "Speed metric (spec §7)."
+    },
+    "avg_latency_seconds": {
+      "category": "efficiency",
+      "status": "implemented",
+      "display_name": "Avg Latency (s)",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "avg_latency_seconds",
+      "scale": "0-inf",
+      "direction": "lower",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": "Mean per-entry latency across non-error entries."
+    },
+    "median_latency_seconds": {
+      "category": "efficiency",
+      "status": "implemented",
+      "display_name": "Median Latency (s)",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "median_latency_seconds",
+      "scale": "0-inf",
+      "direction": "lower",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": ""
+    },
+    "p95_latency_seconds": {
+      "category": "efficiency",
+      "status": "implemented",
+      "display_name": "P95 Latency (s)",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "p95_latency_seconds",
+      "scale": "0-inf",
+      "direction": "lower",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": ""
+    },
+    "elapsed_seconds": {
+      "category": "efficiency",
+      "status": "implemented",
+      "display_name": "Elapsed (s)",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "elapsed_seconds",
+      "scale": "0-inf",
+      "direction": "lower",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": "Total wall time (top-level run-card field, not in scores)."
+    },
+    "total_cost_usd": {
+      "category": "efficiency",
+      "status": "implemented",
+      "display_name": "Total Cost (USD)",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "total_cost_usd",
+      "scale": "USD",
+      "direction": "lower",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": "Actual spend of THIS run; null = cost UNKNOWN (never coerced to 0). Lives in totals, not scores."
+    },
+    "cost_per_entry_usd": {
+      "category": "efficiency",
+      "status": "implemented",
+      "display_name": "Cost / Entry (USD)",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "cost_per_entry_usd",
+      "scale": "USD",
+      "direction": "lower",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": "Feeds cost_adjusted (spec §6.3). Lives in totals."
+    },
+    "cost_per_source_char": {
+      "category": "efficiency",
+      "status": "implemented",
+      "display_name": "Cost / Source Char (USD)",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "cost_per_source_char",
+      "scale": "USD",
+      "direction": "lower",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": "Tokenization-independent cost normalization. Lives in totals."
+    },
+    "tokens_per_entry": {
+      "category": "efficiency",
+      "status": "implemented",
+      "display_name": "Tokens / Entry",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "tokens_per_entry",
+      "scale": "0-inf",
+      "direction": "neutral",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": "Verbosity diagnostic. Lives in totals."
+    },
+    "cost_per_1k_tokens": {
+      "category": "efficiency",
+      "status": "implemented",
+      "display_name": "Cost / 1K Tokens (USD)",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": "cost_per_1k_tokens",
+      "scale": "USD",
+      "direction": "lower",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": false,
+      "notes": "Provider-pricing comparability. Lives in totals."
+    },
+    "cchrf": {
+      "category": "comparator",
+      "status": "partial",
+      "display_name": "Chance-corrected chrF++ (cchrF++)",
+      "plugin_name": null,
+      "card_key": null,
+      "db_column": null,
+      "scale": "0.0-1.0",
+      "direction": "higher",
+      "level": "corpus",
+      "in_composite": false,
+      "verifier_reproducible": true,
+      "notes": "clamp0((chrF++ - floor)/(100 - floor)) — removes the orthography-specific chance floor so scores are cross-language comparable; the clamp at 0 doubles as the noise rail (at-or-below-floor = indistinguishable from chance). PARTIAL: implemented today only in the connection-quality lane (arena/mt_eval_harness/connection_quality.py cchrf(); JS twins cli/website/src/utils/connectionQuality.mjs + arcStrength.mjs; constants SSOT shared/connection-quality.json cq-v1) — NOT computed by tester.py and never a leaderboard column. Floors: cli/website/src/data/cchrf-floors.json (196 languages, champollion-derived), regenerated from research/cchrf/results/atlas.json (Monte-Carlo N1-unigram floors over FLORES-200 dev monolingual text; the study + paper live in research/cchrf). Known caveat before any RANKING consumer wires this: N1 undershoots fluent-output chance by ~2.3 chrF++ measured on 24/204 languages (research/cchrf REVIEW_2026-07-11 M1); the N1-vs-N_w estimator choice is an open founder decision recorded in docs/METRICS_RESEARCH_PROGRAM_2026-07-10.md. Forbidden on language cards (card-integrity R3)."
+    }
+  },
+  "_proposed_renames": [
+    {
+      "current": "run-card key 'corpus_bleu' (top-level) vs canonical 'bleu' vs db_column 'corpus_bleu'",
+      "proposal": "Move BLEU into scores as scores.bleu (spec §9 already shows it there) and keep db_column corpus_bleu; OR rename nothing and let this registry carry the mapping.",
+      "impact": "Published run cards embed the top-level corpus_bleu key; leaderboard row-expand readers would need a fallback read.",
+      "decision": "KEEP (2026-07-07) — do not rename. corpus_bleu stays the top-level run-card key and the DB column; this registry carries the bleu↔corpus_bleu mapping. Renaming a published key breaks archived-card readers for no functional gain."
+    },
+    {
+      "current": "plugin 'crk_linter' vs card key 'lyss-eq' vs scores/db 'equivalent_match_rate'",
+      "proposal": "Rename the plugin to 'lyss_eq_crk' (LYSS branding + language qualifier) when a second language ships a linter; publish.py already discovers by the is_equivalence_linter flag, so the rename is display-only.",
+      "impact": "Published run cards embed plugin_metrics['crk_linter']; archived-report rescoring tools that key on the name would need the alias.",
+      "decision": "DEFER to LYSS 0.2 / language #2 (2026-07-07) — keep 'crk_linter' for launch. Discovery is by the is_equivalence_linter envelope flag, so the name is display-only; the alias-first rename to lyss_eq_crk lands with the champollion_lyss.roles spine when a second language forces the generalization, not as pre-launch churn."
+    },
+    {
+      "current": "plugin 'crk_semantic' vs card key 'lyss-sem' vs scores/db 'semantic_score'",
+      "proposal": "Rename the plugin to 'lyss_sem_crk' at the same time as crk_linter, same rationale (discovery is by semantic_verdict_counts, display-only).",
+      "impact": "Same as crk_linter: plugin_metrics keys in published cards.",
+      "decision": "DEFER to LYSS 0.2 / language #2 (2026-07-07) — keep 'crk_semantic' for launch, alias-first rename to lyss_sem_crk alongside crk_linter. Discovery is by the semantic_verdict_counts envelope, so the name is display-only."
+    },
+    {
+      "current": "plugin 'giellalt_fst_validity' (aggregate key 'avg_fst_validity') vs scores/db 'fst_acceptance_rate'",
+      "proposal": "None — keep as-is. The plugin name states the tool (GiellaLT FST), the metric name states the measurement; this registry is the bridge.",
+      "impact": "n/a",
+      "decision": "KEEP (2026-07-07) — no rename; the tool-vs-measurement split is intentional and the registry bridges it."
+    }
+  ]
+}

package/shared/model-aliases.json

@@ -0,0 +1,7 @@
+{
+  "_comment": "Convenience aliases. Full model list fetched from OpenRouter at runtime.",
+  "gemini-flash": "google/gemini-3.5-flash",
+  "gemini-pro": "google/gemini-3.1-pro-preview",
+  "claude-sonnet": "anthropic/claude-sonnet-4",
+  "gpt": "openai/gpt-5.5"
+}