champollion 0.3.3 → 0.4.0

package/shared/catalogue/card-config.json

@@ -992,7 +992,7 @@
         "import": "champollion_lyss",
         "pip": "champollion-lyss>=0.1,<0.2",
         "description": "Champollion-LYSS measurement plugins — Cree LYSS linter + semantic validator. Fetched on demand from PyPI; the harness core ships no language-specific scorer code.",
-        "ipNotice": "Plains Cree (nêhiyawêwin) data sovereignty — please read.\n\nThe Cree LYSS validation standard and any Plains Cree language data are treated as the property of the nêhiyaw language community and are offered for NON-COMMERCIAL, community-benefit use (OCAP®-aspirant; relational-trust governance in progress). OCAP® is a registered trademark of the First Nations Information Governance Centre (fnigc.ca/ocap-training).\n\nGloss data is fetched live from the public itwêwina API (itwewina.altlab.app) and cached locally for your own use only. The underlying dictionary content (Wolvengrey CW, Maskwacîs MD, AECD) is NOT openly licensed — do not commit, redistribute, publish, or bundle the cache. The Cree FST (GiellaLT/ALTLab, AGPL) is downloaded and invoked as a separate tool, never bundled.\n\nBy installing and running this standard you agree to respect these terms. Set CHAMPOLLION_LYSS_ACCEPT_IP=1 to acknowledge in automation."
+        "ipNotice": "Plains Cree (nêhiyawêwin) data sovereignty — please read.\n\nThe Cree LYSS validation standard and any Plains Cree language data are treated as the property of the nêhiyaw language community and are offered for NON-COMMERCIAL, community-benefit use (sovereignty-aspirant — community ownership and control of language data; relational-trust governance in progress).\n\nGloss data is fetched live from the public itwêwina API (itwewina.altlab.app) and cached locally for your own use only. The underlying dictionary content (Wolvengrey CW, Maskwacîs MD, AECD) is NOT openly licensed — do not commit, redistribute, publish, or bundle the cache. The Cree FST (GiellaLT/ALTLab, AGPL) is downloaded and invoked as a separate tool, never bundled.\n\nBy installing and running this standard you agree to respect these terms. Set CHAMPOLLION_LYSS_ACCEPT_IP=1 to acknowledge in automation."
       },
       "evalMetrics": {
         "lyss-eq": {

package/shared/curated-orthography-conventions.json

@@ -1,25 +1,43 @@
 {
   "_comment": [
-    "Curated orthographic-convention register consumed by scripts/derive-orthographies.mjs.",
-    "One entry per language code, keyed by ISO 15924 script code. Each script entry may set",
+    "Curated orthographic-convention register — the decision-layer input to the",
+    "orthographies[] derivation (cli/scripts/derive-orthographies.mjs). NOT WIRED:",
+    "that deriver is a standalone pass, not imported by cli/scripts/cldf/project.mjs,",
+    "so nothing here reaches a card today. An earlier version of this comment claimed",
+    "the wiring existed; it did not. Do not restore that claim before the import does.",
+    "",
+    "ONE entry per language code, keyed by ISO 15924 script code. Each script entry may set",
     "scheme / longVowelMarking / canonicalForMt and MUST set a source string that cites a",
-    "verifiable basis (on-card prose, a tool the card already lists, or a published convention).",
-    "Only add entries whose convention is actually documented — absence means unknown, and the",
-    "deriver writes nothing it cannot cite (index, not arbiter). Do NOT record measured scores here.",
+    "verifiable basis — a source the card already carries, a published convention, or a tool",
+    "in the repo's register. Only add entries whose convention is actually documented:",
+    "absence means unknown, and the deriver writes nothing it cannot cite (index, not",
+    "arbiter). Optional keys are OMITTED when unknown, never guessed. Do NOT record measured",
+    "scores here — a score of method output is a run result and belongs on the leaderboard.",
+    "",
     "Data-over-code (docs/AGENTS.md §1): language-specific conventions live in this file,",
-    "never hardcoded in the deriver."
+    "never hardcoded in the deriver. Decision-layer inputs to the atlas build live under the",
+    "monorepo-root shared/ (parameters.csv, card-field-disposition.json, catalogue/…); this",
+    "file moved here from cli/shared/ on 2026-09-06 for that reason — cli/shared/ is the",
+    "npm-bundled RUNTIME tree, and nothing reads this register at runtime.",
+    "",
+    "STANDING VERDICT (shared/cldf/curated-file-verdicts.json, 2026-08-06):",
+    "'CITABLE BUT NOT MACHINE-FETCHABLE — keep, with a bibliographic citation added, and",
+    "record that it is asserted from the literature rather than fetched.' Every entry below",
+    "is asserted from the literature and from tools the card already lists. None is fetched,",
+    "so none carries a pin; that is the reason this is a curated register and not a source."
   ],
+  "version": 2,
   "conventions": {
     "crk": {
       "Latn": {
         "scheme": "SRO",
         "longVowelMarking": "circumflex",
         "canonicalForMt": true,
-        "source": "manual-curation (SRO = Standard Roman Orthography; long vowels marked with circumflexes as used by itwêwina/ALTLab and the GiellaLT lang-crk FST the card lists in resources.fsts; reference lexicon: Wolvengrey 2001, nēhiýawēwin: itwēwina / Cree: Words; the card's own formality prose names SRO as the working script)"
+        "source": "manual-curation (asserted from the literature and from a tool the card already carries — not machine-fetched). SRO is Standard Roman Orthography for Plains Cree; it writes the seven long vowels with circumflexes (â ê î ô) against unmarked short vowels. AUTHORITY: the GiellaLT / UAlbertaALTLab lang-crk finite-state transducer, which the crk card lists under resources.fsts from the giellalt-resources source (https://github.com/giellalt/lang-crk, https://github.com/UAlbertaALTLab/lang-crk). Both analysers accept the hyphenated preverb spelling (ê-nipâyân, nikî-nipân, kâ-nipât) and return the SAME analysis plus the tag Err/Orth — the FST's own spelling-error tag — for the fused spelling, so 'correct SRO' here is the grammar machine's verdict rather than a house preference. CANONICAL-FOR-MT: crk is written in two scripts and the pipeline's working script is the Roman one (shared/catalogue/card-config.json scriptConverter.crk converts between them); the scripts[] `primary` display flag is NOT this signal. REFERENCE LEXICON, cited and never redistributed (founder ruling 2026-07-19, index-only permanently): Wolvengrey, Arok, ed. 2001. nehiyawewin: itwewina / Cree: Words. Regina: Canadian Plains Research Center."
       },
       "Cans": {
         "canonicalForMt": false,
-        "source": "manual-curation (Syllabics is the display script, SRO the working script — per the card's own formality prose and the crk scriptConverter)"
+        "source": "manual-curation (asserted, not fetched). Unified Canadian Aboriginal Syllabics is an attested writing system for Plains Cree and stays on the card as an alternative — the language is written both ways and an index says so. It is NOT the pipeline's working form: shared/catalogue/card-config.json registers scriptConverter.crk precisely because syllabics is converted to and from the Roman orthography, and the lang-crk analysers cited on the Latn entry above are SRO-facing. longVowelMarking is OMITTED rather than guessed: syllabics marks vowel length by a diacritic over the syllable, which none of the schema's vocabulary terms (circumflex / macron / double-vowel / none) describes."
       }
     }
   }

package/shared/docent/faq.en.json

@@ -4,7 +4,7 @@
     {
       "id": "what-is-champollion",
       "question": "What is Champollion?",
-      "answer": "Champollion is open-source infrastructure for trustworthy machine translation across the world's languages, with a focus on low-resource and Indigenous ones. It has three faces: a translation CLI that translates your app's locale files with one command, the Network (an open MT evaluation network and leaderboard that maps who can translate what, how well), and a data-sovereignty posture built around the First Nations principles of OCAP®. As the Introduction (/docs/intro) and The Champollion Network (/docs/network/) put it, it is built *with* professionals and communities, never scraped from them — and it is forever a work in progress.",
+      "answer": "Champollion is source-available infrastructure for trustworthy machine-translation evaluation across the world's languages, with a focus on low-resource and Indigenous ones. It has three faces: a translation CLI that translates your app's locale files with one command, the Network (an open MT evaluation network and leaderboard that maps who can translate what, how well), and a data-sovereignty posture built around Indigenous data-sovereignty principles — community ownership and control of language data. As the Introduction (/docs/intro) and The Champollion Network (/docs/network/) put it, it is designed to work *with* professionals and communities, and it never hosts their corpora — and it is forever a work in progress.",
       "sources": [
         "/docs/intro",
         "/docs/network/",
@@ -21,7 +21,7 @@
     {
       "id": "free-and-open-source",
       "question": "Is Champollion free? Is it open source? Does it cost anything?",
-      "answer": "Yes on both counts. Champollion is a non-commercial, open-source research project: the CLI is Apache-2.0 and the evaluation harness is AGPL-3.0, per How the Work Is Funded (/docs/network/sovereignty/economic-model). There is no paid API, no metering, and no revenue share — the only costs you'd incur are the API tokens you spend on your chosen model provider when you translate or benchmark. Champollion itself takes nothing.",
+      "answer": "It is free for noncommercial use, and partly open source. Champollion is a non-commercial, source-available research project: the CLI is PolyForm Noncommercial 1.0.0 (free for noncommercial use; commercial use needs permission) and the evaluation harness is open-source AGPL-3.0, per How the Work Is Funded (/docs/network/sovereignty/economic-model). There is no paid API, no metering, and no revenue share — the only costs you'd incur are the API tokens you spend on your chosen model provider when you translate or benchmark. Champollion itself takes nothing.",
       "sources": [
         "/docs/network/sovereignty/economic-model",
         "/docs/intro"
@@ -39,7 +39,7 @@
     {
       "id": "the-cli",
       "question": "What is the champollion CLI and what does it do?",
-      "answer": "The CLI is the deployment end of Champollion: run `npx champollion sync` and it auto-detects your locale files, format, and target languages, translates what's missing, skips what's done, validates every result through a quality gate, and writes clean output. Per How It Works (/docs/how-it-works), it tracks every source string with SHA-256 hashes so re-runs only re-translate what changed, caches results in Translation Memory to save cost, and lets you pick a different translation method per language pair. It's a full i18n pipeline — sync, watch, lint, verify, audit, XLIFF export, and more (/docs/intro).",
+      "answer": "The CLI is the deployment end of Champollion: run `npx champollion sync` and it auto-detects your locale files, format, and target languages, translates what's missing, skips what's done, checks every result for broken output (empty, echoed, looping or wrong-script — not wrong meaning) through a quality gate, and writes clean output. Per How It Works (/docs/how-it-works), it tracks every source string with SHA-256 hashes so re-runs only re-translate what changed, caches results in Translation Memory to save cost, and lets you pick a different translation method per language pair. It's a full i18n pipeline — sync, watch, lint, verify, audit, XLIFF export, and more (/docs/intro).",
       "sources": [
         "/docs/intro",
         "/docs/how-it-works"
@@ -135,7 +135,7 @@
     {
       "id": "submit-a-method",
       "question": "How do I submit a benchmark run or a method to the leaderboard?",
-      "answer": "Install the harness with `pip install mt-eval`, run it against a registered corpus (for example `mt-eval run --corpus eval-amh-fra-globalvoices-test-v1 --model gemini-pro`), review the run card it produces, then publish with `mt-eval publish` — the full walkthrough is Submit a Method (/docs/network/getting-started/submit-a-method). Your run first appears as self-benchmarked; the server then re-scores your outputs against the sha-pinned corpus and, when it reproduces, promotes it to Champollion Verified. A hosted upload API and web UI are planned but not yet live, so `mt-eval publish` or a pull request to the harness repo are the working paths today.",
+      "answer": "Install the harness with `pip install mt-eval-harness`, run it against a registered corpus (for example `mt-eval run --corpus eval-amh-fra-globalvoices-test-v1 --model gemini-pro`), review the run card it produces, then publish with `mt-eval publish` — the full walkthrough is Submit a Method (/docs/network/getting-started/submit-a-method). Your run first appears as self-benchmarked; the server then re-scores your outputs against the sha-pinned corpus and, when it reproduces, promotes it to Champollion Verified. A hosted upload API and web UI are planned but not yet live, so `mt-eval publish` or a pull request to the harness repo are the working paths today.",
       "sources": [
         "/docs/network/getting-started/submit-a-method",
         "/docs/network/leaderboard/rules"
@@ -228,7 +228,7 @@
     {
       "id": "researcher-involved",
       "question": "I'm an ML researcher — how do I get involved?",
-      "answer": "Build a method and benchmark it. The Network Agent Guide (/docs/network/getting-started/agent-guide) walks through installing the harness (`pip install mt-eval`), running baselines against real corpora, and implementing the simple `TranslationMethod` protocol — coached LLM, FST-gated pipeline, fine-tuned model, or anything that produces translations. You get standardized, reproducible, fingerprinted scoring against a shared corpus catalogue, and every run adds a point to a shared map of what works. Get Involved (/get-involved) points to the developer paths, including running the public benchmark queue.",
+      "answer": "Build a method and benchmark it. The Network Agent Guide (/docs/network/getting-started/agent-guide) walks through installing the harness (`pip install mt-eval-harness`), running baselines against real corpora, and implementing the simple `TranslationMethod` protocol — coached LLM, FST-gated pipeline, fine-tuned model, or anything that produces translations. You get standardized, reproducible, fingerprinted scoring against a shared corpus catalogue, and every run adds a point to a shared map of what works. Get Involved (/get-involved) points to the developer paths, including running the public benchmark queue.",
       "sources": [
         "/docs/network/getting-started/agent-guide",
         "/get-involved",
@@ -264,17 +264,15 @@
       ]
     },
     {
-      "id": "ocap-aspirant-meaning",
-      "question": "What does 'OCAP®-aspirant' mean?",
-      "answer": "OCAP® — Ownership, Control, Access, and Possession — is a set of First Nations data-governance principles, and it is a registered trademark of the First Nations Information Governance Centre (FNIGC). Champollion's posture is OCAP®-aspirant: as Contact, Objections & Takedowns (/docs/network/community/contact-objections-takedown) puts it, the design is built so communities *can* exercise ownership, control, access, and possession of their data and anything derived from it — but whether it *achieves* that is for communities to decide, which is exactly why the objections channel exists. We say OCAP®-aspirant deliberately, rather than claiming to be OCAP®-compliant or certified.",
+      "id": "sovereignty-aspirant-meaning",
+      "question": "What does 'sovereignty-aspirant' mean?",
+      "answer": "Champollion's posture is sovereignty-aspirant — built around Indigenous data-sovereignty principles: community ownership and control of language data. As Contact, Objections & Takedowns (/docs/network/community/contact-objections-takedown) puts it, the design is built so communities *can* exercise ownership and control of their data and anything derived from it — but whether it *achieves* that is for communities to decide, which is exactly why the objections channel exists. We say sovereignty-aspirant deliberately, rather than claiming compliance with or certification under any named framework.",
       "sources": [
         "/docs/network/community/contact-objections-takedown",
         "/docs/network/sovereignty/data-sovereignty"
       ],
       "keywords": [
-        "ocap",
-        "ocap aspirant",
-        "fnigc",
+        "sovereignty aspirant",
         "sovereignty principles",
         "ownership control access possession",
         "indigenous"
@@ -283,7 +281,7 @@
     {
       "id": "data-sovereignty",
       "question": "What is data sovereignty here — do you store or scrape my language data?",
-      "answer": "No. Data Stewardship (/docs/network/sovereignty/data-sovereignty) states the position: language data is biodata, so the people who provide a corpus hold the keys to it and to anything measured against it. Champollion never holds the data — corpora are registered as hash-pinned metadata cards and fetched from the steward's own hosting at evaluation time; take your archive offline and evaluation simply stops. Every license and community restriction is respected by gate, not by promise (CI checks and database triggers), and the design is informed by OCAP®, the CARE Principles, Te Mana Raraunga, and Te Hiku Media's Kaitiakitanga License.",
+      "answer": "No. Data Stewardship (/docs/network/sovereignty/data-sovereignty) states the position: language data is biodata, so the people who provide a corpus hold the keys to it and to anything measured against it. Champollion never holds the data — corpora are registered as hash-pinned metadata cards and fetched from the steward's own hosting at evaluation time; take your archive offline and evaluation simply stops. Every license and community restriction is respected by gate, not by promise (local pre-push checks and database triggers), and the design is informed by Indigenous data-sovereignty principles — community ownership and control of language data — including the CARE Principles, Te Mana Raraunga, and Te Hiku Media's Kaitiakitanga License.",
       "sources": [
         "/docs/network/sovereignty/data-sovereignty",
         "/docs/network/sovereignty/registering-corpora"
@@ -341,7 +339,7 @@
     {
       "id": "speakers-paid",
       "question": "Do speakers get paid, and how much?",
-      "answer": "Yes — paying speakers is treated as non-negotiable, and payment is unconditional (you're paid whether or not your ratings are used). How Speakers Get Paid (/docs/network/perspectives/how-speakers-get-paid) publishes the rates: bilingual speaker work at $50–65 CAD per hour, with corpus curation budgeted at $2,500–6,000 and a full metric-validation round at $1,475–1,920. Payment does not buy your data — you are paid for the work and remain the steward of what you build. These are the published rates for the current Plains Cree work; rates for future languages are set with the partner community and published before the work starts.",
+      "answer": "Yes, once the work is funded — no speaker work is funded or underway today. Paying speakers is treated as non-negotiable, and payment is unconditional (you're paid whether or not your ratings are used). How Speakers Get Paid (/docs/network/perspectives/how-speakers-get-paid) publishes the rates: bilingual speaker work at $50–65 CAD per hour, with corpus curation budgeted at $2,500–6,000 and a full metric-validation round at $1,475–1,920. Payment does not buy your data — you are paid for the work and remain the steward of what you build. These are the rates we intend to pay for Plains Cree work; rates for future languages are set with the partner community and published before the work starts.",
       "sources": [
         "/docs/network/perspectives/how-speakers-get-paid",
         "/docs/network/sovereignty/economic-model"
@@ -398,7 +396,7 @@
     {
       "id": "ai-agent-usage",
       "question": "How should an AI agent use Champollion?",
-      "answer": "Two things are built for agents. First, the machine-readable /llms.txt index maps the whole site, the license lanes, and the public data feeds (queue, mesh, corpus registry). Second, the MCP server `@champollion/mcp-server` (on npm) is the agent-facing door — translate with Translation-Memory caching and a deterministic quality gate, browse the benchmark queue, estimate cost, run benchmarks, read the leaderboard, and query per-language metric-trust evidence, all as tools. The Agent Guides (/docs/guides/agent-guide for the CLI, /docs/network/getting-started/agent-guide for benchmarking) give step-by-step recipes. The best pattern is to set up your own agent with a workspace and your own API key — tokens aren't free.",
+      "answer": "Two things are built for agents. First, the machine-readable /llms.txt index maps the whole site, the license lanes, and the public data feeds (queue, mesh, corpus registry). Second, the MCP server `champollion-mcp-server` (on npm) is the agent-facing door — translate with Translation-Memory caching and a deterministic quality gate, browse the benchmark queue, estimate cost, run benchmarks, read the leaderboard, and query per-language metric-trust evidence, all as tools. The Agent Guides (/docs/guides/agent-guide for the CLI, /docs/network/getting-started/agent-guide for benchmarking) give step-by-step recipes. The best pattern is to set up your own agent with a workspace and your own API key — tokens aren't free.",
       "sources": [
         "/llms.txt",
         "/docs/guides/agent-guide",
@@ -418,7 +416,7 @@
     {
       "id": "contribute-compute",
       "question": "How do I contribute compute by running the benchmark queue?",
-      "answer": "The leaderboard has empty squares — (language pair, model, condition) combinations nobody has measured — kept in a public queue at /queue.json. Install the harness (`pipx install mt-eval`), set one provider API key, and run the highest-value open items with `mt-eval queue --top 5` (add `--dry-run` to see the plan and spend nothing first); the full guide is Contributing Compute (/docs/network/getting-started/contributing-compute) and the Contribute page (/contribute). No account is needed — results publish as 'anonymous' unless you sign in to put your name on the board — and every run is a real, citable contribution to low-resource MT evaluation.",
+      "answer": "The leaderboard has empty squares — (language pair, model, condition) combinations nobody has measured — kept in a public queue at /queue.json. Install the harness (`pipx install mt-eval-harness`), set one provider API key, and run the highest-value open items with `mt-eval queue --top 5` (add `--dry-run` to see the plan and spend nothing first); the full guide is Contributing Compute (/docs/network/getting-started/contributing-compute) and the Contribute page (/contribute). No account is needed — results publish as 'anonymous' unless you sign in to put your name on the board — and every run is a real, citable contribution to low-resource MT evaluation.",
       "sources": [
         "/docs/network/getting-started/contributing-compute",
         "/contribute"
@@ -472,7 +470,7 @@
     {
       "id": "write-code-for-me",
       "question": "Can the site guide write code, debug, or run the harness for me?",
-      "answer": "No — the site guide is an index and a docent, not an engineering service; Contact, Objections & Takedowns (/docs/network/community/contact-objections-takedown) is explicit that this isn't a place to request development help or custom scripts. What it can do is point you at the tooling so your own setup does the work: the /llms.txt cookbook, the MCP server `@champollion/mcp-server`, and the Agent Guides (/docs/guides/agent-guide and /docs/network/getting-started/agent-guide). The recommended path is to set up your own agent with a workspace and your own API key — tokens aren't free, and your agent can then run the harness or the CLI for you.",
+      "answer": "No — the site guide is an index and a docent, not an engineering service; Contact, Objections & Takedowns (/docs/network/community/contact-objections-takedown) is explicit that this isn't a place to request development help or custom scripts. What it can do is point you at the tooling so your own setup does the work: the /llms.txt cookbook, the MCP server `champollion-mcp-server`, and the Agent Guides (/docs/guides/agent-guide and /docs/network/getting-started/agent-guide). The recommended path is to set up your own agent with a workspace and your own API key — tokens aren't free, and your agent can then run the harness or the CLI for you.",
       "sources": [
         "/docs/network/community/contact-objections-takedown",
         "/docs/guides/agent-guide",

package/shared/docent/system-prompt.md

@@ -18,7 +18,8 @@ careful, honest, and humble about what this project is.
 
 Champollion is open translation infrastructure for low-resource languages: a
 translation CLI, a machine-translation evaluation network ("the Network"), and
-a data-sovereignty posture built around the First Nations principles of OCAP®.
+a data-sovereignty posture built around Indigenous data-sovereignty
+principles — community ownership and control of language data.
 It is a permanent work in progress, built to be flexible and to support MT
 developers — and especially low-resource-language speaker communities.
 
@@ -40,22 +41,19 @@ which is provided to you fresh with each question under RETRIEVED CONTEXT below.
   the retrieved context. If two sources disagree, present both — never pick a
   winner or manufacture a consensus.
 
-## Sovereignty: how to talk about OCAP® and communities
-
-- OCAP® — Ownership, Control, Access, Possession — is a First Nations data
-  governance framework and a **registered trademark of the First Nations
-  Information Governance Centre (FNIGC)**. Always write it with the ® mark, and
-  attribute it to FNIGC when you explain it.
-- The project is **OCAP®-aspirant**: designed *with the First Nations principles
-  of OCAP® in mind*. That means the design is built so communities *can*
-  exercise ownership, control, access, and possession of their data and anything
-  derived from it. Whether it *achieves* sovereignty is **for communities to
-  decide, not for us to claim.** Say this nuance plainly and without defensiveness
-  when asked.
-- **Never** say the project is OCAP®-compliant, -certified, -aligned, -forward,
-  or that it "follows OCAP®." We do the work so that others can judge it —
-  self-certification is worth nothing. ("OCAP®-forward" was a retired earlier
-  term; the only sanctioned form is "OCAP®-aspirant".)
+## Sovereignty: how to talk about data sovereignty and communities
+
+- The project's posture is built around Indigenous data-sovereignty
+  principles — community ownership and control of language data and anything
+  derived from it.
+- The project is **sovereignty-aspirant**: the design is built so communities
+  *can* exercise ownership and control of their data and anything derived from
+  it. Whether it *achieves* sovereignty is **for communities to decide, not for
+  us to claim.** Say this nuance plainly and without defensiveness when asked.
+- **Never** say the project is compliant with, certified under, or aligned with
+  any named data-governance framework. We do the work so that others can judge
+  it — self-certification is worth nothing. (The only sanctioned house term is
+  "sovereignty-aspirant".)
 - **Never name a specific nation, community, or organization as a partner, key
   custodian, or endorser.** If asked who the custodians are, say they are
   "community key custodians (in confirmation)" and that no one is named publicly
@@ -75,7 +73,7 @@ You are a guide, not a developer, a translator, or a general assistant.
   pair-programming session — and that tokens aren't free. Point them to the
   agent-native path instead: set up their own agent with its own workspace and
   give it `champollion.dev/llms.txt` and the MCP server
-  (`@champollion/mcp-server`), which expose the queue, language metadata, and
+  (`champollion-mcp-server`), which expose the queue, language metadata, and
   translation/benchmark tools as real tools an agent can drive. Offer to tour
   them to the exact docs.
 - **No translation-on-demand.** You don't translate documents or sentences for
@@ -115,7 +113,7 @@ quickly."
 
 If the visitor is itself an agent (or asks how an agent should use Champollion),
 skip the tour and point it straight at `champollion.dev/llms.txt` and the MCP
-server `@champollion/mcp-server`, and briefly name what those expose (benchmark
+server `champollion-mcp-server`, and briefly name what those expose (benchmark
 queue, language metadata, translate/benchmark tools, the coaching "open problem"
 path). That's the token-efficient path for both of us.
 

package/shared/explainers/tc-features.json

@@ -1,6 +1,6 @@
 {
   "_meta": {
-    "generated": "2026-08-12T23:23:44.431Z",
+    "generated": "2026-09-27T08:15:46.589Z",
     "generator": "cli/scripts/build-explainers.mjs",
     "description": "Typological feature explainers: per-feature plain-English names, value code → label maps, MT-relevance notes and citations. Join from a fact via propertyIndex[\"<source>|<property>\"], or from a card typologicalProfile key via propertyIndex[\"card|<key>\"].",
     "counts": {
@@ -7832,8 +7832,8 @@
       "id": "derived-grambank-hasObliqueCase",
       "source": "grambank-1.0.3",
       "property": "hasObliqueCase",
-      "name": "Oblique case (GB071)",
-      "question": "Whether pronouns take case marking for core grammatical roles.",
+      "name": "Oblique case (GB072)",
+      "question": "Whether full nouns take case marking for oblique (non-core) roles such as location or instrument.",
       "values": {
         "true": {
           "label": "Yes",
@@ -7844,10 +7844,10 @@
           "plain": "The language does not have this feature."
         }
       },
-      "mt_relevance": "Pronoun case errors (I/me-type) are highly salient grammatical mistakes in generated text.",
+      "mt_relevance": "Oblique case endings carry roles English expresses with prepositions (in, with, to), so each preposition must be mapped to the right ending.",
       "citation": {
         "authors": "Skirgård, Hedvig, Hannah J. Haynie, Damián E. Blasi, Hedvig Skirgård et al. (Grambank Consortium)",
-        "title": "Oblique case (GB071) — derived from Grambank v1.0.3",
+        "title": "Oblique case (GB072) — derived from Grambank v1.0.3",
         "source_url": "https://grambank.clld.org"
       }
     },
@@ -7878,8 +7878,8 @@
       "id": "derived-grambank-marksPresentTense",
       "source": "grambank-1.0.3",
       "property": "marksPresentTense",
-      "name": "Marks present tense (GB084)",
-      "question": "Whether the grammar has a dedicated marker for present time reference.",
+      "name": "Marks present tense (GB082)",
+      "question": "Whether verbs carry overt morphological marking of present tense.",
       "values": {
         "true": {
           "label": "Yes",
@@ -7893,7 +7893,7 @@
       "mt_relevance": "Present-tense marking interacts with aspect: rendering it as a plain English present often asserts a habituality the source did not.",
       "citation": {
         "authors": "Skirgård, Hedvig, Hannah J. Haynie, Damián E. Blasi, Hedvig Skirgård et al. (Grambank Consortium)",
-        "title": "Marks present tense (GB084) — derived from Grambank v1.0.3",
+        "title": "Marks present tense (GB082) — derived from Grambank v1.0.3",
         "source_url": "https://grambank.clld.org"
       }
     },
@@ -8127,7 +8127,7 @@
           "plain": "absent"
         }
       },
-      "mt_relevance": "Classifier systems require noun-specific function words that have no counterpart in non-classifier languages.",
+      "mt_relevance": "A polite/familiar \"you\" must be chosen on every second-person reference; English marks neither, so MT has to infer the relationship between speaker and addressee.",
       "citation": {
         "authors": "Skirgård, Hedvig, Hannah J. Haynie, Damián E. Blasi, Hedvig Skirgård et al. (Grambank Consortium)",
         "title": "Is there a politeness distinction in 2nd person forms? (Grambank GB415)",
@@ -19258,11 +19258,11 @@
       "name": "Noun classifiers",
       "question": "Whether nouns combine with classifier words that sort them by type.",
       "values": {},
-      "mt_relevance": "Classifier systems require noun-specific function words that have no counterpart in non-classifier languages.",
+      "mt_relevance": "Numeral classifiers must be chosen per noun; the wrong classifier is immediately visible to native speakers.",
       "citation": {
         "authors": "Champollion enrichment pipeline (derived)",
-        "title": "Noun classifiers — derived from Grambank GB415",
-        "source_url": "https://grambank.clld.org/parameters/GB415"
+        "title": "Noun classifiers — derived from Grambank GB057 (numeral classifiers)",
+        "source_url": "https://grambank.clld.org/parameters/GB057"
       }
     },
     "derived-classifierLanguage": {
@@ -19272,11 +19272,11 @@
       "name": "Classifier language",
       "question": "Whether the language routinely uses classifiers when counting or referring to nouns.",
       "values": {},
-      "mt_relevance": "Classifier systems require noun-specific function words that have no counterpart in non-classifier languages.",
+      "mt_relevance": "Numeral classifiers must be chosen per noun; the wrong classifier is immediately visible to native speakers.",
       "citation": {
         "authors": "Champollion enrichment pipeline (derived)",
-        "title": "Classifier language — derived from Grambank GB415 / WALS 55A",
-        "source_url": "https://grambank.clld.org/parameters/GB415"
+        "title": "Classifier language — derived from Grambank GB057 / WALS 55A",
+        "source_url": "https://grambank.clld.org/parameters/GB057"
       }
     },
     "derived-hasNumeralClassifiers": {

package/shared/gettext-plural-forms.json

@@ -0,0 +1,45 @@
+{
+  "_about": "The Plural-Forms header GNU gettext's msginit writes for a new catalog, per language: its built-in table (gettext-tools plural-table.c). champollion gives a catalog it creates the same header, so the catalog's msgstr[] slots are the ones gettext and Django select at runtime. Looked up as msginit does: the full code first (pt_BR), then the language (pt). A language not listed gets no header from msginit; champollion derives one from CLDR (lib/po.js synthesizePluralForms). An existing catalog's own header is never replaced.",
+  "_source": "Probed 2026-10-03 with `msginit --no-translator --no-wrap --locale=<code>` (GNU gettext-tools 1.0, GETTEXTCLDRDIR unset) for every ISO 639-1 code and pt_BR, pt_PT, en_GB, de_AT, es_AR, fr_CA, zh_CN, zh_TW, sr_Latn; region variants not listed resolved to their language's entry.",
+  "forms": {
+    "be": "nplurals=3; plural=(n%10==1 && n%100!=11 ? 0 : n%10>=2 && n%10<=4 && (n%100<10 || n%100>=20) ? 1 : 2);",
+    "bg": "nplurals=2; plural=(n != 1);",
+    "ca": "nplurals=2; plural=(n != 1);",
+    "cs": "nplurals=3; plural=(n==1) ? 0 : (n>=2 && n<=4) ? 1 : 2;",
+    "da": "nplurals=2; plural=(n != 1);",
+    "de": "nplurals=2; plural=(n != 1);",
+    "el": "nplurals=2; plural=(n != 1);",
+    "en": "nplurals=2; plural=(n != 1);",
+    "eo": "nplurals=2; plural=(n != 1);",
+    "es": "nplurals=2; plural=(n != 1);",
+    "et": "nplurals=2; plural=(n != 1);",
+    "fi": "nplurals=2; plural=(n != 1);",
+    "fo": "nplurals=2; plural=(n != 1);",
+    "fr": "nplurals=2; plural=(n > 1);",
+    "ga": "nplurals=3; plural=n==1 ? 0 : n==2 ? 1 : 2;",
+    "he": "nplurals=2; plural=(n != 1);",
+    "hr": "nplurals=3; plural=(n%10==1 && n%100!=11 ? 0 : n%10>=2 && n%10<=4 && (n%100<10 || n%100>=20) ? 1 : 2);",
+    "hu": "nplurals=2; plural=(n != 1);",
+    "it": "nplurals=2; plural=(n != 1);",
+    "ja": "nplurals=1; plural=0;",
+    "ko": "nplurals=1; plural=0;",
+    "lt": "nplurals=3; plural=(n%10==1 && n%100!=11 ? 0 : n%10>=2 && (n%100<10 || n%100>=20) ? 1 : 2);",
+    "lv": "nplurals=3; plural=(n%10==1 && n%100!=11 ? 0 : n != 0 ? 1 : 2);",
+    "nb": "nplurals=2; plural=(n != 1);",
+    "nl": "nplurals=2; plural=(n != 1);",
+    "nn": "nplurals=2; plural=(n != 1);",
+    "no": "nplurals=2; plural=(n != 1);",
+    "pl": "nplurals=3; plural=(n==1 ? 0 : n%10>=2 && n%10<=4 && (n%100<10 || n%100>=20) ? 1 : 2);",
+    "pt": "nplurals=2; plural=(n != 1);",
+    "pt_BR": "nplurals=2; plural=(n > 1);",
+    "ro": "nplurals=3; plural=n==1 ? 0 : (n==0 || (n%100 > 0 && n%100 < 20)) ? 1 : 2;",
+    "ru": "nplurals=3; plural=(n%10==1 && n%100!=11 ? 0 : n%10>=2 && n%10<=4 && (n%100<10 || n%100>=20) ? 1 : 2);",
+    "sk": "nplurals=3; plural=(n==1) ? 0 : (n>=2 && n<=4) ? 1 : 2;",
+    "sl": "nplurals=4; plural=(n%100==1 ? 0 : n%100==2 ? 1 : n%100==3 || n%100==4 ? 2 : 3);",
+    "sr": "nplurals=3; plural=(n%10==1 && n%100!=11 ? 0 : n%10>=2 && n%10<=4 && (n%100<10 || n%100>=20) ? 1 : 2);",
+    "sv": "nplurals=2; plural=(n != 1);",
+    "tr": "nplurals=2; plural=(n != 1);",
+    "uk": "nplurals=3; plural=(n%10==1 && n%100!=11 ? 0 : n%10>=2 && n%10<=4 && (n%100<10 || n%100>=20) ? 1 : 2);",
+    "vi": "nplurals=1; plural=0;"
+  }
+}

package/shared/human-services.json

@@ -13,7 +13,7 @@
     "sovereignty_flags": {
       "off_community_processing": false,
       "attribution_required": true,
-      "ocap_review": "pending"
+      "sovereignty_review": "pending"
     },
     "nc_terms": "Community-set terms; non-commercial use pending custodian confirmation.",
     "consent_attested": false,

package/shared/method-registry.json

@@ -65,6 +65,7 @@
         "OPENAI_API_KEY"
       ],
       "default_base_url": "http://localhost:11434/v1",
+      "keyless": true,
       "homepage": "https://github.com/ollama/ollama",
       "license": "Varies (self-hosted / OpenAI-compatible gateway)",
       "commercialReady": false,
@@ -147,6 +148,7 @@
         "APERTIUM_API_KEY"
       ],
       "default_base_url": "https://apertium.org/apy",
+      "keyless": true,
       "homepage": "https://www.apertium.org",
       "license": "GPL-3.0+ (Apertium)",
       "commercialReady": false,