Measure the writers this category harms, and let the boundary move

[peopleworks]

What changes

• Corpus: 206 essays by adult learners of English join the calibration corpus (PELIC, Pittsburgh, 2006-2012, CC BY-NC-ND 4.0 - measured, never redistributed). fetch --source pelic reproduces the exact selection: first version, writing classes, 662+ words, one text per student by lowest id. Verified: the tool and an independent Python selection produce the same 206 ids and identical SHA-256s.
• The published number: boundary 25 -> 30/100. At 25 the learners were flagged 9 of 206 (4.4%, interval 2.3-8.1%) against 0 of 90 published texts. At 30 it is 2 of 296 overall (0.2-2.4%). English now supports its own threshold (271 texts, bound 1.4%); Spanish still borrows. Floor 662 -> 649 words.
• Report prose that stopped being true: the paragraph claiming no group sits above the rest is now computed at the recommended boundary and names the group, its count and its interval. The two caveats that said "measured on published articles rather than student work" now say what the corpus holds, in both languages with recomputed sourceHash.
• Defect found by the new group: chat.eager-opener accepted Of course, with a comma (an ordinary concession) alongside Certainly!. Eleven learner essays fired it; zero published ones had, by register rather than merit. Both packs now require !; four tests added and shown to fail against the old pattern.
• Docs carrying the numbers: README, SKILL (items 3 and 5), student sheets EN/ES, corpus README, code comments in VerdictBands, PublishedCalibration, EvidenceReport, RulePack, fixtures and tests. Blog articles and PARAPHRASE.md keep their figures: dated studies.
• Version 0.6.0 in the six places. First change since 0.5.0 that touches Core, so this one carries a number. Merging obliges v0.6.0 and then desktop-v0.6.0, since the download page already points at it.

What deliberately does not change

The human-rate gates (humanRatePer1000) stay fitted on the 90 published texts. Refitting on the pooled corpus was tried: with learners as the majority, twelve English rules lose their gate entirely and the connector gates inflate two- to seven-fold - learner rates spent on everyone, the population trap #59 documents. A gate per group is issue #75, with the numbers.

Verification

• 332 Core + 72 Documents + 7 Onnx + 31 Desktop tests green; boot tests unaffected.
• run regenerates Docs/CALIBRATION.md and the embedded calibration at engine 0.6.0; SkillDocumentTests pins SKILL.md to it.
• Mutation: reverting the regex fails exactly the four new tests.

Closes nothing; opens #75.

🤖 Generated with Claude Code

https://claude.ai/code/session_015PEbbiYSNPw7jE3LrPNhyF

[peopleworks]

Committee round 1 (Fable: method · Copilot: code · DeepSeek: prose, pending)

Both delivered request changes; every claim was verified before acting.

Fixed in cb42a5e:

• Fable, high — the evidence report quoted the language's best bound (1.4%, at ≥35) beside the recommended threshold (30, where the bound is 2.7%). Latent since the report was written; activated the day a language first earned its own threshold with texts flagged at it. PublishedLanguage.RateHighAtThreshold added, report quotes it, mutation-tested.
• Fable, high — CC BY attribution: PELIC's creators were named nowhere. Generated page now has a Sources table with the citation; manifest carries doi:10.5281/zenodo.3991977.
• Fable, medium — Docs/Teaching/committee.{en,es}.md, Docs/Teaching/README.md, SKILL item 6 still described a corpus of ninety. Updated.
• Fable, follow-up — the page now shows the step below the boundary (9 of 206 at 25), computed, not narrated.
• Copilot, medium — run returned 0 with manifest texts missing; now fails unless --allow-missing.
• Copilot, medium — a 200 with HTML would be cached as the CSV for ever; validated before the move, cached junk discarded.
• Copilot, medium — Spread() reported a group with no row at the boundary as flagged nothing; distinguished.
• Copilot, low — duplicate answer_id refused rather than resolved by row order.

Not changed, with reasons:

• Copilot, high — the download button 404s between the Pages deploy and desktop-v0.6.0. The desktop workflow builds the tagged commit and only requires the constant to match, so the release will be cut from this branch head before merging; Pages then deploys against an existing asset. Recorded in the release procedure.
• Copilot, low — CLI cannot write a Spanish report: pre-existing, now The CLI cannot write the evidence report in Spanish #77.
• Fable, follow-ups — in-sample gates inflate the group gap; learner threshold sits at the edge: both on Rules taught in academic English fire two to three times as often on second-language learners #75 as design constraints.

[peopleworks]

Committee round 2 (DeepSeek: prose, honesty, Spanish)

Verdict request changes; applied in 6bf054d except one, with the reason.

Applied: every surface in this PR now says "before 2022" rather than "before generative models existed" — GPT-3 is from 2020, so the old phrase overclaimed on the most-read pages; the rest of the repository is #78. The Spanish student sheet no longer presents a bound with no supported threshold in the same shape as a rate at one ("medida sobre solo veinticinco textos y todavía sin un umbral respaldado"). "Tightens as it grows" promised a direction; README and the generated page now say the numbers follow the data either way. The step-below sentence is computed: it says "the only flags anywhere in the corpus at that boundary" only when that is true, and "why the boundary does not sit at 25" rather than "why it sits where it does". The report caveats name the encyclopedia revisions they had omitted, and the Spanish ones say "de esta versión", "a un umbral", "acotar con precisión", with agreement fixed. Plus "almost none", "studies report", the learner interval in SKILL item 3, and the two Spanish idioms.

Not applied as proposed: dropping the 5% from SKILL item 3. The 5% is not an unsupported number; it is the target the boundary is derived from (the lowest score whose interval stays under it), and the calibration page leads with the same words. What was wrong is that the skill quoted it with no provenance, so an assistant would repeat it bare. It now reads: "The boundary is the lowest score whose 95% interval stays under the project's 5% target … at 30/100 the published build flagged 2 of 296 … interval reaching 2.4%. Quote the interval; never the point estimate or the target alone."

Three reviewers, three lenses, every claim verified. Ready to merge once CI is green; release order in the PR description.