paperlint 9.0.0 → 10.0.0

package/skills/find-venue/SKILL.md

@@ -4,7 +4,7 @@ description: "Discover and rank real venues for a given paper, scored by authors
 allowed-tools: [Read, Write, Grep, Glob, WebSearch, WebFetch, Agent]
 ---
 
-<!-- vigiles:sha256:8b332cbc072b46ee compiled from skills/find-venue/SKILL.md.spec.ts -->
+<!-- vigiles:sha256:42d8e283fb7045f1 compiled from skills/find-venue/SKILL.md.spec.ts -->
 
 # find-venue — rank real venues by what earns the credit, then keep or switch
 
@@ -43,7 +43,7 @@ rank = credit-weight × topic-fit × accept-probability × deadline-feasibility
   prestige (a bonus, not a gate). **Non-archival ⇒ credit-weight ≈ 0** — it earns no authorship credit,
   so it sinks to the bottom no matter how nice the CFP reads.
 - **topic-fit** — how squarely the paper matches the CFP scope (a stretch fit raises reject risk).
-- **accept-probability** — realistic odds given the paper's maturity and the venue's bar.
+- **accept-probability** — realistic odds given the paper's maturity and how selective the venue is.
 - **deadline-feasibility** — can the paper (and artifact) be ready by the deadline? A great venue you
   miss is a zero.
 - **remote-friendliness** — remote-OK is a real weight here, not a footnote.
@@ -84,7 +84,9 @@ rank = credit-weight × topic-fit × accept-probability × deadline-feasibility
 - the venue cards, `presets/<venue>.md` in the paperlint package beside each venue preset (e.g.
   `agenticdev.md`, `aisec.md`) — the winner gets a card with its HotCRP quirks beside its preset;
   for a venue with no shipped preset, the project's own `./venues/<name>.md` beside its
-  `./venues/<name>.jsonc`. Plain data, not a new skill.
+  `./venues/<name>.jsonc`. Plain data, not a new skill. That `./venues/` is not the venue's corpus
+  of accepted papers, `<papers-root>/venues/<venue>/` (`study-accepted-papers`): the three places are
+  told apart in `docs/venues.md` in the paperlint package.
 - `extend-paper` — the prestige-upgrade target this skill names but defers.
 
 ## Provenance

package/skills/find-venue/SKILL.md.spec.ts

@@ -49,7 +49,7 @@ rank = credit-weight × topic-fit × accept-probability × deadline-feasibility
   prestige (a bonus, not a gate). **Non-archival ⇒ credit-weight ≈ 0** — it earns no authorship credit,
   so it sinks to the bottom no matter how nice the CFP reads.
 - **topic-fit** — how squarely the paper matches the CFP scope (a stretch fit raises reject risk).
-- **accept-probability** — realistic odds given the paper's maturity and the venue's bar.
+- **accept-probability** — realistic odds given the paper's maturity and how selective the venue is.
 - **deadline-feasibility** — can the paper (and artifact) be ready by the deadline? A great venue you
   miss is a zero.
 - **remote-friendliness** — remote-OK is a real weight here, not a footnote.
@@ -90,7 +90,9 @@ rank = credit-weight × topic-fit × accept-probability × deadline-feasibility
 - the venue cards, \`presets/<venue>.md\` in the paperlint package beside each venue preset (e.g.
   \`agenticdev.md\`, \`aisec.md\`) — the winner gets a card with its HotCRP quirks beside its preset;
   for a venue with no shipped preset, the project's own \`./venues/<name>.md\` beside its
-  \`./venues/<name>.jsonc\`. Plain data, not a new skill.
+  \`./venues/<name>.jsonc\`. Plain data, not a new skill. That \`./venues/\` is not the venue's corpus
+  of accepted papers, \`<papers-root>/venues/<venue>/\` (\`study-accepted-papers\`): the three places are
+  told apart in \`docs/venues.md\` in the paperlint package.
 - \`extend-paper\` — the prestige-upgrade target this skill names but defers.
 
 ## Provenance

package/skills/grade-paper-writing/SKILL.md

@@ -1,17 +1,17 @@
 ---
 name: grade-paper-writing
-description: Use when asking "does this paper read well?" / "is the writing any good?" / "it reads like a wall of text and jargon" / "grade the writing" on a draft. Grades WRITING CRAFT only — ten dimensions 1–5 (Title, Abstract, Intro architecture, Structure, Prose clarity, Jargon discipline, Landing-the-point, Figure economy, Honesty-without-hedge-stacking, Register distance to the target venue) against how the best-written papers read (Peyton Jones, McEnerney, Gopen & Swan; exemplars Trusting Trust, Carlini, Greshake) AND against 2–3 accepted papers of the target venue, naming the offending sentence and the fix for each. Runs the cold-read stall pass with TWO committed personas (a non-academic engineer per section, and a workshop reviewer skimming abstract, introduction and conclusion) whose stall inventory — not the rubric number — is the readability gate other skills consume (pc-panel-review, paper-adversarial-review, harden-paper). Defaults to a blind multi-grader panel; after fixes, mandates the claims-preservation diff (row `claims`). NOT a content/defect review (paper-adversarial-review / pc-panel-review), NOT a structural cut plan (tighten-paper — run that first on a bloated draft), NOT venue-bar content strength (study-accepted-papers). Compose with harden-paper (which calls it), draft-paper (generative counterpart), render-paper.
+description: Grade how a paper reads and fix it sentence by sentence. Use when asked if a paper reads well or to grade the prose, or when the writing is a wall of text, jargon-heavy, or reads like a blog post. Returns 1–5 scores, the sentences cold readers could not follow, and a rewrite for each. Not for cuts (tighten-paper), one edited paragraph (cold-read-diff), scientific flaws (paper-adversarial-review, pc-panel-review), submit readiness (harden-paper) or what the venue rewards (study-accepted-papers).
 allowed-tools: [Read, Write, Edit, Grep, Glob, Agent, Skill]
 ---
 
-<!-- vigiles:sha256:21c8899648c9b223 compiled from skills/grade-paper-writing/SKILL.md.spec.ts -->
+<!-- vigiles:sha256:1897016b69f3f9b8 compiled from skills/grade-paper-writing/SKILL.md.spec.ts -->
 
 # grade-paper-writing — grade how the paper READS, then fix it sentence by sentence
 
-The review skills hunt scientific defects; `study-accepted-papers` checks whether the contribution clears
-the venue's bar. Neither tells you the paper is a slog to read. This skill does exactly that: it grades
-the **writing craft** against how the best-written papers actually read, and turns each low score into a
-concrete, sentence-level fix. Its target failure mode is the common one — *"a wall of text and jargon that
+The review skills hunt scientific defects; `study-accepted-papers` scores the contribution on the
+strong-accept scorecard. Neither tells you the paper is a slog to read. This skill does exactly that: it
+grades the **writing craft** against how the best-written papers actually read, and turns each low score
+into a concrete, sentence-level fix. Its target failure mode is the common one — *"a wall of text and jargon that
 doesn't drive home how crazy the situation is"* — which is a writing problem, not a science problem, and
 which the defect-hunting skills will pass right over.
 
@@ -67,7 +67,7 @@ Things that exist only in the render, every one of which produces a real stall:
 - how much of page 1 is numbers, which is what a reviewer's first impression is actually made of;
 - line numbers and two-column measure, which change where the eye stops.
 
-`repro/build-submission.sh` produces the PDF; `render-paper` turns pages into images. Judge from
+`repro/tools/build-submission.sh` produces the PDF; `render-paper` turns pages into images. Judge from
 those. If only the source is available, say so in the report — a stall inventory taken from source
 is a weaker artifact and must not be presented as the same thing.
 

package/skills/grade-paper-writing/SKILL.md.spec.ts

@@ -9,15 +9,15 @@ import { experimental_skill } from "vigiles/spec";
 export default experimental_skill({
   name: "grade-paper-writing",
   description:
-    'Use when asking "does this paper read well?" / "is the writing any good?" / "it reads like a wall of text and jargon" / "grade the writing" on a draft. Grades WRITING CRAFT only — ten dimensions 1–5 (Title, Abstract, Intro architecture, Structure, Prose clarity, Jargon discipline, Landing-the-point, Figure economy, Honesty-without-hedge-stacking, Register distance to the target venue) against how the best-written papers read (Peyton Jones, McEnerney, Gopen & Swan; exemplars Trusting Trust, Carlini, Greshake) AND against 2–3 accepted papers of the target venue, naming the offending sentence and the fix for each. Runs the cold-read stall pass with TWO committed personas (a non-academic engineer per section, and a workshop reviewer skimming abstract, introduction and conclusion) whose stall inventory — not the rubric number — is the readability gate other skills consume (pc-panel-review, paper-adversarial-review, harden-paper). Defaults to a blind multi-grader panel; after fixes, mandates the claims-preservation diff (row `claims`). NOT a content/defect review (paper-adversarial-review / pc-panel-review), NOT a structural cut plan (tighten-paper — run that first on a bloated draft), NOT venue-bar content strength (study-accepted-papers). Compose with harden-paper (which calls it), draft-paper (generative counterpart), render-paper.',
+    "Grade how a paper reads and fix it sentence by sentence. Use when asked if a paper reads well or to grade the prose, or when the writing is a wall of text, jargon-heavy, or reads like a blog post. Returns 1–5 scores, the sentences cold readers could not follow, and a rewrite for each. Not for cuts (tighten-paper), one edited paragraph (cold-read-diff), scientific flaws (paper-adversarial-review, pc-panel-review), submit readiness (harden-paper) or what the venue rewards (study-accepted-papers).",
   tools: ["Read", "Write", "Edit", "Grep", "Glob", "Agent", "Skill"],
   body: `
 # grade-paper-writing — grade how the paper READS, then fix it sentence by sentence
 
-The review skills hunt scientific defects; \`study-accepted-papers\` checks whether the contribution clears
-the venue's bar. Neither tells you the paper is a slog to read. This skill does exactly that: it grades
-the **writing craft** against how the best-written papers actually read, and turns each low score into a
-concrete, sentence-level fix. Its target failure mode is the common one — *"a wall of text and jargon that
+The review skills hunt scientific defects; \`study-accepted-papers\` scores the contribution on the
+strong-accept scorecard. Neither tells you the paper is a slog to read. This skill does exactly that: it
+grades the **writing craft** against how the best-written papers actually read, and turns each low score
+into a concrete, sentence-level fix. Its target failure mode is the common one — *"a wall of text and jargon that
 doesn't drive home how crazy the situation is"* — which is a writing problem, not a science problem, and
 which the defect-hunting skills will pass right over.
 
@@ -73,7 +73,7 @@ Things that exist only in the render, every one of which produces a real stall:
 - how much of page 1 is numbers, which is what a reviewer's first impression is actually made of;
 - line numbers and two-column measure, which change where the eye stops.
 
-\`repro/build-submission.sh\` produces the PDF; \`render-paper\` turns pages into images. Judge from
+\`repro/tools/build-submission.sh\` produces the PDF; \`render-paper\` turns pages into images. Judge from
 those. If only the source is available, say so in the report — a stall inventory taken from source
 is a weaker artifact and must not be presented as the same thing.
 

package/skills/map-prior-work/SKILL.md

@@ -1,11 +1,11 @@
 ---
 name: map-prior-work
-description: Sweep the competitive landscape for a paper idea BEFORE drafting — find everyone already working on it, date each against your submission (prior vs concurrent is a five-month question, not a vibe), triage by whether they threaten your claimed contribution, deep-read the ones that do, and emit a verdict on WHAT YOU CAN STILL CLAIM plus a related-work skeleton. Saves the full map, not a summary. Use right after research-ideate and before draft-paper, and re-run as a cheap delta pass before submit (hot fields move). Distinct from analyze-sibling-paper (one known competitor, deep) — this one FINDS the competitors and decides which deserve that treatment; distinct from verify-citations (checks cites you already wrote, too late to reshape a contribution); distinct from study-accepted-papers (mines a venue's bar, not your rivals). Compose with research-ideate (upstream), analyze-sibling-paper (it hands off), draft-paper and find-venue (downstream).
+description: Sweep the competitive landscape for a paper idea BEFORE drafting — find everyone already working on it, date each against your submission (prior vs concurrent is a five-month question, not a vibe), triage by whether they threaten your claimed contribution, deep-read the ones that do, and emit a verdict on WHAT YOU CAN STILL CLAIM plus a related-work skeleton. Saves the full map, not a summary. Use right after research-ideate and before draft-paper, and re-run as a cheap delta pass before submit (hot fields move). Distinct from analyze-sibling-paper (one known competitor, deep) — this one FINDS the competitors and decides which deserve that treatment; distinct from verify-citations (checks cites you already wrote, too late to reshape a contribution); distinct from study-accepted-papers (mines what a venue rewards, not your rivals). Compose with research-ideate (upstream), analyze-sibling-paper (it hands off), draft-paper and find-venue (downstream).
 context: fork
 allowed-tools: [WebSearch, WebFetch, Read, Write, Grep, Glob, Bash, Agent]
 ---
 
-<!-- vigiles:sha256:55971644308c470f compiled from skills/map-prior-work/SKILL.md.spec.ts -->
+<!-- vigiles:sha256:474b3d7b8e0542ae compiled from skills/map-prior-work/SKILL.md.spec.ts -->
 
 # map-prior-work — find out who already did it, while you can still change course
 

package/skills/map-prior-work/SKILL.md.spec.ts

@@ -9,7 +9,7 @@ import { experimental_skill } from "vigiles/spec";
 export default experimental_skill({
   name: "map-prior-work",
   description:
-    "Sweep the competitive landscape for a paper idea BEFORE drafting — find everyone already working on it, date each against your submission (prior vs concurrent is a five-month question, not a vibe), triage by whether they threaten your claimed contribution, deep-read the ones that do, and emit a verdict on WHAT YOU CAN STILL CLAIM plus a related-work skeleton. Saves the full map, not a summary. Use right after research-ideate and before draft-paper, and re-run as a cheap delta pass before submit (hot fields move). Distinct from analyze-sibling-paper (one known competitor, deep) — this one FINDS the competitors and decides which deserve that treatment; distinct from verify-citations (checks cites you already wrote, too late to reshape a contribution); distinct from study-accepted-papers (mines a venue's bar, not your rivals). Compose with research-ideate (upstream), analyze-sibling-paper (it hands off), draft-paper and find-venue (downstream).",
+    "Sweep the competitive landscape for a paper idea BEFORE drafting — find everyone already working on it, date each against your submission (prior vs concurrent is a five-month question, not a vibe), triage by whether they threaten your claimed contribution, deep-read the ones that do, and emit a verdict on WHAT YOU CAN STILL CLAIM plus a related-work skeleton. Saves the full map, not a summary. Use right after research-ideate and before draft-paper, and re-run as a cheap delta pass before submit (hot fields move). Distinct from analyze-sibling-paper (one known competitor, deep) — this one FINDS the competitors and decides which deserve that treatment; distinct from verify-citations (checks cites you already wrote, too late to reshape a contribution); distinct from study-accepted-papers (mines what a venue rewards, not your rivals). Compose with research-ideate (upstream), analyze-sibling-paper (it hands off), draft-paper and find-venue (downstream).",
   context: "fork",
   tools: [
     "WebSearch",

package/skills/paper-pipeline/SKILL.md

@@ -4,7 +4,7 @@ description: The orchestrator for writing a research paper end-to-end, from idea
 allowed-tools: [Read, Write, Edit, Grep, Glob, Bash, WebSearch, WebFetch, Agent, Skill]
 ---
 
-<!-- vigiles:sha256:67a3989629952dc1 compiled from skills/paper-pipeline/SKILL.md.spec.ts -->
+<!-- vigiles:sha256:cb2eecdb81ba766d compiled from skills/paper-pipeline/SKILL.md.spec.ts -->
 
 # paper-pipeline — the conductor for the whole organism
 
@@ -150,7 +150,7 @@ Ordered, because each really does gate the next. Two pairs run in parallel; that
 | **Validate the idea** | **`research-ideate`** | Is it a reusable method (survives review) in your lane, and does it feed the credit criteria? Go/no-go + the MVP finding. |
 | **Map the competition** ‖ | **`map-prior-work`** | 🔴 **Who already did this?** Multi-angle sweep; date each hit against your submission (5 months = prior, not concurrent); triage by which layer of your claim it threatens; deep-read the dangerous ones. Output: **what you can still claim** + a related-work skeleton. *Added 2026-07-27 after a five-month-older paper surfaced at T−10 days on a finished draft.* |
 | **Pick the venue** ‖ | **`find-venue`** | Rank by credit-weight (peer-reviewed + indexed) × fit × accept-odds × deadline × remote. |
-| **Learn the venue's bar** | **`study-accepted-papers`** | 🔴 **What does an ACCEPTED paper look like HERE?** Measure the venue's whole accepted volume, deep-read 8–12 of your contribution type, and extract what made them strong: evidence scale, figures, artifact release, how the abstract is built. File the venue card (`presets/<venue>.md` in the package, or the project's own beside its preset) and confirm family rules (`references/acl-venue-rules.md`). *Promoted to its own step 2026-08-03: on `compile-rules-2026` every judgement of the paper's quality — rubric, persona, simulated panel, accept probability — came from instruments this pipeline itself generated, with no external grounding, and the author had to say so at T−3 days. A closed loop optimises against its own model of a reviewer.* |
+| **Learn what the venue rewards** | **`study-accepted-papers`** (mode A idea, B design) | 🔴 **What does this venue REWARD, before the study is designed?** The venue's award winners against its ordinary accepted papers over 5–7 years (plus award papers on the topic at other venues, labelled apart), coded for substance as well as form, written as **requirements for the study**: what the oracle under the headline is, which belief stated on record the claim targets, which factor is a column. The corpus goes to `<papers-root>/venues/<venue>/`, shared by every paper for that venue; the stage is row `venuebar`. **The same skill comes back once text exists** (mode C draft: the strong-accept scorecard, M1–M10, in `PIPELINE-STATUS.md`; mode D final: the abstract and introduction). The skill always asks which mode. Venue format facts go in its preset (`presets/<venue>.jsonc` in the package, or the project's own) and the venue card beside it (`presets/<venue>.md`); confirm family rules (`references/acl-venue-rules.md`). *Promoted to its own step 2026-08-03: every judgement of a paper's quality had come from instruments this pipeline itself generated, with no external grounding. Moved to modes on 2026-10-12 (#207): run only on a draft, its main findings arrived after they could change the design.* |
 | 🔴 **Can you physically submit?** | **`plan-paper-timeline`** | **The step that owns the question no stage used to own.** Create the portal account NOW; drive the profile to **ACTIVE**, not "pending moderation"; confirm the form's fields and the artifact-hosting requirement. OpenReview moderation runs **up to two weeks**, and an author with no institutional email has no expedite route. Scorecard row `access`. *Added 2026-08-03, when this was the sole critical-path item on a finished paper at T−4 days: the science was done and there was nowhere to put it.* |
 | **Schedule it** | **`plan-paper-timeline`** | Deadlines → calendar with buffers, in the author's own zone (AoE − 12h ≠ your clock). Space papers apart. |
 | 🔴 **State the claim** | **`argument-arc`** (frame mode) | **One paragraph: what will this paper claim, and what would have to be true?** Before any run. The frame decides which results matter, and reframing after the data is collected is how experiments get thrown away — observed on `compile-rules-2026`, and the reason `argument-arc` has a rebuild mode at all. Costs a paragraph; saves a study. Scorecard row `frame`. |
@@ -170,6 +170,8 @@ No numbers. Nothing here happens once, and nothing here is "next".
 
 - **`build-benchmark`** — a design that inverts the pitfall; honest stats; a self-checking artifact
   (`references/artifact-checklist.md`).
+- **`build-label-page`** — when the study needs a person to label items (gold labels, a blind review):
+  a phone page with a shared answer store, the answers exported into the study's `data/`.
 - **`draft-paper`** — section skeleton + the accept-and-get-cited properties
   (`references/writing-craft.md`).
 - **`argument-arc`** — 🔴 **does the draft carry the reader to ONE conclusion?** One sentence per
@@ -221,7 +223,7 @@ Use an atom for a quick spot-check; **run the panel for the decision.**
 skill: 🔴 **argument architecture** — does the whole thing carry the reader to one conclusion
 (`argument-arc`, in the LOOP; the axis that failed on `compile-rules-2026` precisely because it was
 nobody's) · **content strength** — does the contribution clear *this venue's* bar (`study-accepted-papers`,
-mines the accepted corpus for Accept→Strong levers; in SETUP, before the prose is set) ·
+from the venue's own award winners: in SETUP as requirements for the study, and again on the draft as the strong-accept scorecard, M1–M10) ·
 **scientific defects** — what a hostile reviewer attacks (`paper-adversarial-review` / `pc-panel-review`)
 · **writing craft** — does it read like a human wrote it (`grade-paper-writing`, the 10-dim rubric in
 `references/writing-craft.md`; called by `harden-paper`). A paper can ace one and fail another — grade all four.
@@ -294,7 +296,7 @@ Over {{corpus.n}} real rules files, with the extractor's proposed mapping
 taken as if an author had written it, {{annotated.failedContradicted}} fail admission.
 ```
 
-`<paper>/repro/numbers.tsv` gives each name its value, the JSON path or **formula** that recomputes
+`<paper>/repro/data/numbers.tsv` gives each name its value, the JSON path or **formula** that recomputes
 it, and two guards. The build refuses a PDF unless: the name resolves · the value still matches the
 data · the **paragraph** names the arm (`requires`) · the **sentence** does not claim a different one
 (`forbids`). Requires and forbids are separate on purpose — the pairing is auditing's, where a control
@@ -310,7 +312,7 @@ binds every row from now on, and the existing rows are *enumerated rather than e
 says out loud how many figures still have no machine-checked path to data. On the paper that has it:
 7 guarded, 248 grandfathered. That ratio is a finding, not a score.
 
-**Adopting it in another paper:** copy `repro/paper_numbers.py`, `repro/paper_numbers.selftest.py`
+**Adopting it in another paper:** copy `repro/tools/paper_numbers.py`, `repro/tools/paper_numbers.selftest.py`
 and the build step that calls `expand()` on `paper.tex` before `pdflatex`; generate the
 grandfather list from the current draft; register the numbers a reviewer will actually check first.
 
@@ -350,7 +352,7 @@ printing two unrelated 22s broke it the same hour. Key on the occurrence.
 🔴 **Hand the gate to a red team the same day you build it.** This one was, and seven holes came
 back, the worst being that the original defect could simply be **retyped as digits** — its value was
 grandfathered, so the guard applied to the macro and not to the claim. Every hole is now a case in
-`repro/paper_numbers.selftest.py` (18/18, including three that must stay **quiet** — a gate that
+`repro/tools/paper_numbers.selftest.py` (18/18, including three that must stay **quiet** — a gate that
 fires on ordinary writing gets muted, and then its silence reads as coverage).
 
 ## 🔴 The page limit is a SUBMISSION gate, not an EDIT gate

package/skills/paper-pipeline/SKILL.md.spec.ts

@@ -167,7 +167,7 @@ Ordered, because each really does gate the next. Two pairs run in parallel; that
 | **Validate the idea** | **\`research-ideate\`** | Is it a reusable method (survives review) in your lane, and does it feed the credit criteria? Go/no-go + the MVP finding. |
 | **Map the competition** ‖ | **\`map-prior-work\`** | 🔴 **Who already did this?** Multi-angle sweep; date each hit against your submission (5 months = prior, not concurrent); triage by which layer of your claim it threatens; deep-read the dangerous ones. Output: **what you can still claim** + a related-work skeleton. *Added 2026-07-27 after a five-month-older paper surfaced at T−10 days on a finished draft.* |
 | **Pick the venue** ‖ | **\`find-venue\`** | Rank by credit-weight (peer-reviewed + indexed) × fit × accept-odds × deadline × remote. |
-| **Learn the venue's bar** | **\`study-accepted-papers\`** | 🔴 **What does an ACCEPTED paper look like HERE?** Measure the venue's whole accepted volume, deep-read 8–12 of your contribution type, and extract what made them strong: evidence scale, figures, artifact release, how the abstract is built. File the venue card (\`presets/<venue>.md\` in the package, or the project's own beside its preset) and confirm family rules (\`references/acl-venue-rules.md\`). *Promoted to its own step 2026-08-03: on \`compile-rules-2026\` every judgement of the paper's quality — rubric, persona, simulated panel, accept probability — came from instruments this pipeline itself generated, with no external grounding, and the author had to say so at T−3 days. A closed loop optimises against its own model of a reviewer.* |
+| **Learn what the venue rewards** | **\`study-accepted-papers\`** (mode A idea, B design) | 🔴 **What does this venue REWARD, before the study is designed?** The venue's award winners against its ordinary accepted papers over 5–7 years (plus award papers on the topic at other venues, labelled apart), coded for substance as well as form, written as **requirements for the study**: what the oracle under the headline is, which belief stated on record the claim targets, which factor is a column. The corpus goes to \`<papers-root>/venues/<venue>/\`, shared by every paper for that venue; the stage is row \`venuebar\`. **The same skill comes back once text exists** (mode C draft: the strong-accept scorecard, M1–M10, in \`PIPELINE-STATUS.md\`; mode D final: the abstract and introduction). The skill always asks which mode. Venue format facts go in its preset (\`presets/<venue>.jsonc\` in the package, or the project's own) and the venue card beside it (\`presets/<venue>.md\`); confirm family rules (\`references/acl-venue-rules.md\`). *Promoted to its own step 2026-08-03: every judgement of a paper's quality had come from instruments this pipeline itself generated, with no external grounding. Moved to modes on 2026-10-12 (#207): run only on a draft, its main findings arrived after they could change the design.* |
 | 🔴 **Can you physically submit?** | **\`plan-paper-timeline\`** | **The step that owns the question no stage used to own.** Create the portal account NOW; drive the profile to **ACTIVE**, not "pending moderation"; confirm the form's fields and the artifact-hosting requirement. OpenReview moderation runs **up to two weeks**, and an author with no institutional email has no expedite route. Scorecard row \`access\`. *Added 2026-08-03, when this was the sole critical-path item on a finished paper at T−4 days: the science was done and there was nowhere to put it.* |
 | **Schedule it** | **\`plan-paper-timeline\`** | Deadlines → calendar with buffers, in the author's own zone (AoE − 12h ≠ your clock). Space papers apart. |
 | 🔴 **State the claim** | **\`argument-arc\`** (frame mode) | **One paragraph: what will this paper claim, and what would have to be true?** Before any run. The frame decides which results matter, and reframing after the data is collected is how experiments get thrown away — observed on \`compile-rules-2026\`, and the reason \`argument-arc\` has a rebuild mode at all. Costs a paragraph; saves a study. Scorecard row \`frame\`. |
@@ -187,6 +187,8 @@ No numbers. Nothing here happens once, and nothing here is "next".
 
 - **\`build-benchmark\`** — a design that inverts the pitfall; honest stats; a self-checking artifact
   (\`references/artifact-checklist.md\`).
+- **\`build-label-page\`** — when the study needs a person to label items (gold labels, a blind review):
+  a phone page with a shared answer store, the answers exported into the study's \`data/\`.
 - **\`draft-paper\`** — section skeleton + the accept-and-get-cited properties
   (\`references/writing-craft.md\`).
 - **\`argument-arc\`** — 🔴 **does the draft carry the reader to ONE conclusion?** One sentence per
@@ -238,7 +240,7 @@ Use an atom for a quick spot-check; **run the panel for the decision.**
 skill: 🔴 **argument architecture** — does the whole thing carry the reader to one conclusion
 (\`argument-arc\`, in the LOOP; the axis that failed on \`compile-rules-2026\` precisely because it was
 nobody's) · **content strength** — does the contribution clear *this venue's* bar (\`study-accepted-papers\`,
-mines the accepted corpus for Accept→Strong levers; in SETUP, before the prose is set) ·
+from the venue's own award winners: in SETUP as requirements for the study, and again on the draft as the strong-accept scorecard, M1–M10) ·
 **scientific defects** — what a hostile reviewer attacks (\`paper-adversarial-review\` / \`pc-panel-review\`)
 · **writing craft** — does it read like a human wrote it (\`grade-paper-writing\`, the 10-dim rubric in
 \`references/writing-craft.md\`; called by \`harden-paper\`). A paper can ace one and fail another — grade all four.
@@ -311,7 +313,7 @@ Over {{corpus.n}} real rules files, with the extractor's proposed mapping
 taken as if an author had written it, {{annotated.failedContradicted}} fail admission.
 \`\`\`
 
-\`<paper>/repro/numbers.tsv\` gives each name its value, the JSON path or **formula** that recomputes
+\`<paper>/repro/data/numbers.tsv\` gives each name its value, the JSON path or **formula** that recomputes
 it, and two guards. The build refuses a PDF unless: the name resolves · the value still matches the
 data · the **paragraph** names the arm (\`requires\`) · the **sentence** does not claim a different one
 (\`forbids\`). Requires and forbids are separate on purpose — the pairing is auditing's, where a control
@@ -327,7 +329,7 @@ binds every row from now on, and the existing rows are *enumerated rather than e
 says out loud how many figures still have no machine-checked path to data. On the paper that has it:
 7 guarded, 248 grandfathered. That ratio is a finding, not a score.
 
-**Adopting it in another paper:** copy \`repro/paper_numbers.py\`, \`repro/paper_numbers.selftest.py\`
+**Adopting it in another paper:** copy \`repro/tools/paper_numbers.py\`, \`repro/tools/paper_numbers.selftest.py\`
 and the build step that calls \`expand()\` on \`paper.tex\` before \`pdflatex\`; generate the
 grandfather list from the current draft; register the numbers a reviewer will actually check first.
 
@@ -367,7 +369,7 @@ printing two unrelated 22s broke it the same hour. Key on the occurrence.
 🔴 **Hand the gate to a red team the same day you build it.** This one was, and seven holes came
 back, the worst being that the original defect could simply be **retyped as digits** — its value was
 grandfathered, so the guard applied to the macro and not to the claim. Every hole is now a case in
-\`repro/paper_numbers.selftest.py\` (18/18, including three that must stay **quiet** — a gate that
+\`repro/tools/paper_numbers.selftest.py\` (18/18, including three that must stay **quiet** — a gate that
 fires on ordinary writing gets muted, and then its silence reads as coverage).
 
 ## 🔴 The page limit is a SUBMISSION gate, not an EDIT gate

package/skills/paper-pipeline/grade-paper-writing-ablation.eval.mjs

@@ -233,8 +233,9 @@
  *      renamed from two-letter codes to words (`Hn` → `claims`) because the two-letter space had
  *      collided. The target's description carried one such
  *      reference, so "…mandates the Hn claim-preservation diff." became "…mandates the
- *      claims-preservation diff (row `claims`)." A0 is still read from disk and every guard still
- *      passes; the numbers below were measured on the pre-rename bytes and are left as recorded.
+ *      claims-preservation diff (row `claims`)." Every guard still passes on A0, which is pinned in
+ *      this file since the live description was rewritten (2026-10-13); the numbers below were
+ *      measured on the pre-rename bytes and are left as recorded.
  *
  *      measurement                                            AH        SH       total
  *      parent, this morning                                3/12 25%  2/12 17%   5/24  21%
@@ -380,6 +381,7 @@ import {
   realpathSync,
 } from "node:fs";
 import { join } from "node:path";
+import { tmpdir } from "node:os";
 import { frontmatterBlock } from "#lib/markdown";
 import { parseFm } from "../../lib/skill-corpus.mjs";
 import { installedSkills } from "./scripts/consumer.mjs";
@@ -395,9 +397,7 @@ const PARENT = join(
   "paper-pipeline",
   "framing-vs-vocabulary.eval.mjs",
 );
-const SCRATCH =
-  process.env.ABLATION_SCRATCH ??
-  "/tmp/claude-0/-home-user-mine/8268acb1-66a2-55ac-858e-2b9e3c81f84e/scratchpad/gpw-ablation";
+const SCRATCH = process.env.ABLATION_SCRATCH ?? join(tmpdir(), "gpw-ablation");
 const NS = "vigiles-loose-skills";
 const TARGET = "grade-paper-writing";
 const STEM = "2026-08-08-grade-paper-writing-ablation";
@@ -493,7 +493,18 @@ const allPrompts = CELLS.flatMap((k) =>
   PROMPTS[k].map((p) => ({ cell: k, prompt: p })),
 );
 
-// ── the baseline description, read from disk (never hardcoded — it would drift) ──
+// ── the baseline description, PINNED ─────────────────────────────────────────
+// This used to be read from the installed SKILL.md, so the arms followed every edit of it. On
+// 2026-10-13 the description was rewritten (499 chars: what it does, when, when not, what it
+// returns), and every arm below is a set of EXACT SUBSTRING DELETIONS: cut from the new text, they
+// find nothing to delete and the file throws before preflight. The arms are an experiment on THIS
+// text, so the text is pinned and the arms build exactly as before. It is the last description the
+// deletion strings were written against (1371 chars), which already differs from the 1174 chars the
+// 2026-08-08 result measured (see "A0 IS TWO TOKENS OFF" in the header, and the register clauses
+// added since). How the LIVE description fires is measured by
+// `grade-paper-writing/grade-paper-writing.eval.mjs` on the shared trigger cases, not here.
+const A0 =
+  'Use when asking "does this paper read well?" / "is the writing any good?" / "it reads like a wall of text and jargon" / "grade the writing" on a draft. Grades WRITING CRAFT only — ten dimensions 1–5 (Title, Abstract, Intro architecture, Structure, Prose clarity, Jargon discipline, Landing-the-point, Figure economy, Honesty-without-hedge-stacking, Register distance to the target venue) against how the best-written papers read (Peyton Jones, McEnerney, Gopen & Swan; exemplars Trusting Trust, Carlini, Greshake) AND against 2–3 accepted papers of the target venue, naming the offending sentence and the fix for each. Runs the cold-read stall pass with TWO committed personas (a non-academic engineer per section, and a workshop reviewer skimming abstract, introduction and conclusion) whose stall inventory — not the rubric number — is the readability gate other skills consume (pc-panel-review, paper-adversarial-review, harden-paper). Defaults to a blind multi-grader panel; after fixes, mandates the claims-preservation diff (row `claims`). NOT a content/defect review (paper-adversarial-review / pc-panel-review), NOT a structural cut plan (tighten-paper — run that first on a bloated draft), NOT content strength against what the venue rewards (study-accepted-papers). Compose with harden-paper (which calls it), draft-paper (generative counterpart), render-paper.';
 if (!existsSync(join(SKILLS_DIR, TARGET, "SKILL.md")))
   throw new Error(`${TARGET} is not installed under ${SKILLS_DIR}`);
 // READING a field is a parser's job. The hand-rolled version below used a lookahead
@@ -511,7 +522,6 @@ const parseDesc = (md) =>
     .trim();
 const parseName = (md) => fmOf(md).name;
 const BASE_MD = readFileSync(join(SKILLS_DIR, TARGET, "SKILL.md"), "utf-8");
-const A0 = parseDesc(BASE_MD);
 if (parseName(BASE_MD) !== TARGET)
   throw new Error(`${TARGET}/SKILL.md declares name: ${parseName(BASE_MD)}`);
 
@@ -525,7 +535,7 @@ const D_SUBORDINATION = [
   "harden-paper (which calls it), ",
 ];
 const D_NEGATION = [
-  " NOT a content/defect review (paper-adversarial-review / pc-panel-review), NOT a structural cut plan (tighten-paper — run that first on a bloated draft), NOT venue-bar content strength (study-accepted-papers).",
+  " NOT a content/defect review (paper-adversarial-review / pc-panel-review), NOT a structural cut plan (tighten-paper — run that first on a bloated draft), NOT content strength against what the venue rewards (study-accepted-papers).",
 ];
 const D_MACHINERY = [
   " (Title, Abstract, Intro architecture, Structure, Prose clarity, Jargon discipline, Landing-the-point, Figure economy, Honesty-without-hedge-stacking, Register distance to the target venue)",
@@ -559,7 +569,7 @@ const D_PERSONA_ONLY = [
   " Grades WRITING CRAFT only — ten dimensions 1–5 (Title, Abstract, Intro architecture, Structure, Prose clarity, Jargon discipline, Landing-the-point, Figure economy, Honesty-without-hedge-stacking, Register distance to the target venue) against how the best-written papers read (Peyton Jones, McEnerney, Gopen & Swan; exemplars Trusting Trust, Carlini, Greshake) AND against 2–3 accepted papers of the target venue, naming the offending sentence and the fix for each.",
   " whose stall inventory — not the rubric number — is the readability gate other skills consume (pc-panel-review, paper-adversarial-review, harden-paper)",
   " Defaults to a blind multi-grader panel; after fixes, mandates the claims-preservation diff (row `claims`).",
-  " NOT a content/defect review (paper-adversarial-review / pc-panel-review), NOT a structural cut plan (tighten-paper — run that first on a bloated draft), NOT venue-bar content strength (study-accepted-papers).",
+  " NOT a content/defect review (paper-adversarial-review / pc-panel-review), NOT a structural cut plan (tighten-paper — run that first on a bloated draft), NOT content strength against what the venue rewards (study-accepted-papers).",
   " Compose with harden-paper (which calls it), draft-paper (generative counterpart), render-paper.",
 ];
 

package/skills/paper-pipeline/references/acceptance-gate.md

@@ -57,7 +57,7 @@ drive**.
 - ❌ this section earns its place
 - ❌ the reader could follow that sentence
 - ❌ the delta over the nearest prior work is real
-- ❌ the draft sits at this venue's bar
+- ❌ the draft passes the strong-accept scorecard
 - ❌ the idea is worth doing
 - ❌ the arc carries the reader end to end
 - ❌ the accept probability is 88%
@@ -105,7 +105,7 @@ at all** — and it has four rungs:
 Type A lives on rung 3 and sometimes 2. **Rung 1 is not in their taxonomy at all**, and it is the
 rung we are actually best at:
 
-- `repro/paper_numbers.py` — a registered quantity is written `{{annotated.failedContradicted}}`, not
+- `repro/tools/paper_numbers.py` — a registered quantity is written `{{annotated.failedContradicted}}`, not
   `22`. The build refuses to emit a PDF when the name does not resolve, when the value has drifted
   from the data, or when the paragraph does not declare which experimental arm it is reporting. This
   is not a gate that judges the number. **The wrong number has no representation in the source.**
@@ -117,12 +117,12 @@ not to find a second model family we do not have — it is to ask whether the qu
 or rung-1-shaped surrogate.** For measurement papers it very often does, and every time we found one
 it caught a defect the judgement gates had passed:
 
-| the taste question                              | the surrogate we built                                                                              | what it caught                                                                                 |
-| ----------------------------------------------- | --------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------- |
-| does this number support this sentence          | `repro/numbers.tsv` guards — the arm word the paragraph must name, the condition it must not claim  | the 2026-08-05 abstract defect, 29 hours before a deadline; five judgement gates had passed it |
-| is this quantity really bound to its data       | `repro/arm_permutation.py` — permute the arm labels, rebuild, an unchanged quantity was never bound | a whole class the artifact reviewer recomputes and passes, because the number is _correct_     |
-| did the number survive into the delivered paper | `repro/delivered_pdf.py` — read the built PDF back                                                  | a superscript minus lost in typesetting                                                        |
-| is this reference real                          | `extract-ref-facts.mjs` + rules `refs/*` — arXiv / CrossRef                                         | nine fabricated titles carrying correct arXiv ids, and one invented award                      |
+| the taste question                              | the surrogate we built                                                                                    | what it caught                                                                                 |
+| ----------------------------------------------- | --------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------- |
+| does this number support this sentence          | `repro/data/numbers.tsv` guards — the arm word the paragraph must name, the condition it must not claim   | the 2026-08-05 abstract defect, 29 hours before a deadline; five judgement gates had passed it |
+| is this quantity really bound to its data       | `repro/tools/arm_permutation.py` — permute the arm labels, rebuild, an unchanged quantity was never bound | a whole class the artifact reviewer recomputes and passes, because the number is _correct_     |
+| did the number survive into the delivered paper | `repro/tools/delivered_pdf.py` — read the built PDF back                                                  | a superscript minus lost in typesetting                                                        |
+| is this reference real                          | `extract-ref-facts.mjs` + rules `refs/*` — arXiv / CrossRef                                               | nine fabricated titles carrying correct arXiv ids, and one invented award                      |
 
 ---
 
@@ -188,7 +188,7 @@ we want more of.
 | `tighten-paper`            | a cover-to-cover read produced no action                                                         | **B**                                 | ❌ drive only                                                                                                         |
 | `grade-paper-writing`      | the stall inventory came back empty                                                              | **B**                                 | ❌ drive only                                                                                                         |
 | `analyze-sibling-paper`    | the delta is real and no text has to change                                                      | **B**                                 | ❌ drive only                                                                                                         |
-| `study-accepted-papers`    | the draft already sits at the venue's bar                                                        | **B**                                 | ❌ drive only (the corpus is external; the comparison is ours)                                                        |
+| `study-accepted-papers`    | the draft already passes the strong-accept scorecard                                             | **B**                                 | ❌ drive only (the corpus is external; the comparison is ours)                                                        |
 | `paper-adversarial-review` | a hostile read found nothing                                                                     | **B**                                 | ❌ drive only                                                                                                         |
 | `pc-panel-review`          | accept with no must-fix                                                                          | **B**                                 | ❌ drive only — the flagship offender                                                                                 |
 | `harden-paper`             | every gate green and every axis PASS/FIXED — _the only verdict that hands off to `submit-paper`_ | **A over B**                          | ❌ the wrapper is A, the thing wrapped is B; the handoff is an acquittal                                              |
@@ -248,16 +248,16 @@ Everything in `.claude/skills/paper-pipeline/scripts/`, `.claude/lib/` and `<pap
 construction — that is why it is trustworthy, and it is not a coincidence but a selection effect: a
 program is what you can write only when the question is Type A.
 
-| unit                                                 | question                                                        | rung                                                                                                                  |
-| ---------------------------------------------------- | --------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------- |
-| `repro/report-submission.py`                         | body pages, overfull boxes, unresolved refs, dropped characters | A (3)                                                                                                                 |
-| `repro/check-anon.sh`, `check-figure.sh`             | deanonymising strings; does the figure fit the column           | A (3)                                                                                                                 |
-| `prose-lint.mjs`                                     | thresholds with published baselines                             | A (3) — _measuring_ a B question, which is legitimate and is not the same as answering it                             |
-| rules `refs/*` (facts by `extract-ref-facts.mjs`)    | does the registry have this record                              | A (3), **external adjudicator**                                                                                       |
-| `paper-lint.mjs`                                     | appendix ratio, shaved passages                                 | A (3)                                                                                                                 |
-| `repro/arm_permutation.py`, `repro/delivered_pdf.py` | is the quantity bound to its arm; did it survive typesetting    | A (3), metamorphic                                                                                                    |
-| **`repro/paper_numbers.py` + `numbers.tsv`**         | —                                                               | **rung 1: no PDF exists in which the number is wrong**                                                                |
-| `paper-edit-guard.hook.ts`                           | writing to a paper from Bash                                    | rung 2 — **leaky**, an interpreter reading its program from stdin still gets through (`papers/CLAUDE.md`, 2026-08-06) |
+| unit                                                             | question                                                        | rung                                                                                                                  |
+| ---------------------------------------------------------------- | --------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------- |
+| `repro/tools/report-submission.py`                               | body pages, overfull boxes, unresolved refs, dropped characters | A (3)                                                                                                                 |
+| `repro/tools/check-anon.sh`, `check-figure.sh`                   | deanonymising strings; does the figure fit the column           | A (3)                                                                                                                 |
+| `prose-lint.mjs`                                                 | thresholds with published baselines                             | A (3) — _measuring_ a B question, which is legitimate and is not the same as answering it                             |
+| rules `refs/*` (facts by `extract-ref-facts.mjs`)                | does the registry have this record                              | A (3), **external adjudicator**                                                                                       |
+| `paper-lint.mjs`                                                 | appendix ratio, shaved passages                                 | A (3)                                                                                                                 |
+| `repro/tools/arm_permutation.py`, `repro/tools/delivered_pdf.py` | is the quantity bound to its arm; did it survive typesetting    | A (3), metamorphic                                                                                                    |
+| **`repro/tools/paper_numbers.py` + `numbers.tsv`**               | —                                                               | **rung 1: no PDF exists in which the number is wrong**                                                                |
+| `paper-edit-guard.hook.ts`                                       | writing to a paper from Bash                                    | rung 2 — **leaky**, an interpreter reading its program from stdin still gets through (`papers/CLAUDE.md`, 2026-08-06) |
 
 `prose-lint.mjs` deserves its note. It puts a **number with a threshold** on a question that is
 Type B — is this sentence readable. That is not

package/skills/paper-pipeline/references/artifact-checklist.md

@@ -27,6 +27,11 @@ the 140-run `reproduce.py` (AgenticDev) and the 46-guard `evaluate.py`/`mutate.p
 
 ## Contents (mirror this file map)
 
+In the paper folder the artifact lives in `repro/`, laid out as `repro/allowed-files-only` requires: the
+scripts the whole paper uses in `repro/tools/`, shared data in `repro/data/`, each study in
+`repro/<slug>-YYYY-MM-DD/` with its `README.md`, and `repro/README.md` as the index. The shipped
+zip below is built from it; the anonymized copy that was uploaded goes in the paper's `submission/`.
+
 - `reproduce.py` / the harness scripts — self-check, print `RESULT: PASS`, exit non-zero on drift.
 - `data/…` — the raw measurements the numbers derive from.
 - `README.md` — a **"what reproduces which paper number"** table (paper location → the line that prints

package/skills/paper-pipeline/references/pipeline-status-template.md

@@ -25,6 +25,35 @@ stages:
 ---
 ```
 
+## One live-state file: the plan, the scorecard, the stages
+
+Everything about a paper that is _live_ (true now, and rewritten when it changes) is a section of this
+one file, so there is one place to read and nothing to keep in step:
+
+| section                              | holds                                                                                                                                          | written by                                                     |
+| ------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------- |
+| **Plan / now**                       | what is being done now and next, the decisions in force (who, when, a link to the record), the schedule                                        | whoever changes the plan, in the same commit                   |
+| **Strong-accept scorecard (M1–M10)** | a row per strong-accept mechanism: status (`pass` / `partial` / `fail` / `—`), evidence as `file:line`, owner, plan or `accept ceiling: <why>` | `study-accepted-papers` (mode C: every row; mode D: M2 and M4) |
+| **Stages** (SETUP … AFTER)           | a row per unit of pipeline work                                                                                                                | the skill that ran it                                          |
+
+- **No `PLAN.md` beside it.** A plan in a second file drifts from the status it plans against;
+  `paper/allowed-files-only` reports a `PLAN.md` at the paper's root. A decision's _history_ (why it was
+  taken, what was rejected) is a record, not live state: it goes to `research/` as a dated file.
+- **Each study gets a dated folder in `repro/`.** The `study` stage's runs go in
+  `repro/<slug>-YYYY-MM-DD/` (or `<slug>-YYYY-MM` for one that ran over weeks), with a `README.md`
+  saying what it asks, how to run it and what it found, and a line in `repro/README.md`. Scripts the
+  whole paper uses go in `repro/tools/`, shared data in `repro/data/`; `repro/allowed-files-only` reports
+  anything else at the top of `repro/`.
+- **The scorecard's verdicts live only here.** The dated report `study-accepted-papers` writes under
+  `reviews/` keeps the reasoning and links to this section; it never copies the table, so there is one
+  truth to update. The mechanisms themselves, with what each is not, are in
+  `study-accepted-papers/references/mechanisms.md`. Nothing checks the table mechanically yet: keep its
+  columns as the template has them.
+- 🔴 **Only stage tables have a first column headed `id`.** A stage-table parser reads every
+  `id`-headed table as stage rows, and a row under a heading that names no stage section is an error
+  there: a fix-list table headed `id` gave a real paper two such errors (2026-10-12). Head every other
+  table with `#` or a word, as the scorecard does. `src/new-paper.test.ts` holds the template to it.
+
 ## The row ids, and what each one runs
 
 The `id` cell is the **join key** — the same string appears in this template, in every paper's
@@ -36,7 +65,7 @@ the scorecard explains itself where it lives:
 | `idea`        | SETUP      | is the idea worth a paper?                                                | `research-ideate`               |
 | `priorwork`   | SETUP      | full competitor sweep, before drafting                                    | `map-prior-work`                |
 | `venue`       | SETUP      | pick where it goes                                                        | `find-venue`                    |
-| `venuebar`    | SETUP      | what this venue rewards; strengthening levers                             | `study-accepted-papers`         |
+| `venuebar`    | SETUP      | what this venue rewards (modes A/B), and later the draft scored (C/D)     | `study-accepted-papers`         |
 | `access`      | SETUP      | 🔴 _can you physically upload?_ — account, ACTIVE profile, portal, form   | `plan-paper-timeline`           |
 | `schedule`    | SETUP      | the backwards plan and its calendar events                                | `plan-paper-timeline`           |
 | `frame`       | SETUP      | the claim, stated **before** the runs                                     | `argument-arc` (frame mode)     |

package/skills/paper-pipeline/repro/2026-08-07-fork-check.log

@@ -49,4 +49,4 @@ ru-original    67%    53%  -13pp   (10/15 vs 8/15)
 No gate: this file measures an effect, it does not enforce one. Read OVERALL first, then SPLIT BY ORIGIN — a gap that appears only on translated Russian is a finding about translation, not about Russian.
 
 total spend: $9.53 API-equivalent
-no baseline at /home/user/mine/.claude/skills/paper-pipeline/pipeline-language.baseline.json — record one with --update-baseline
+no baseline at <project>/.claude/skills/paper-pipeline/pipeline-language.baseline.json — record one with --update-baseline

package/skills/paper-pipeline/repro/2026-08-07-fork-check2.log

@@ -30,4 +30,4 @@ ru-original   100%   100%  +0pp   (6/6 vs 6/6)
 No gate: this file measures an effect, it does not enforce one. Read OVERALL first, then SPLIT BY ORIGIN — a gap that appears only on translated Russian is a finding about translation, not about Russian.
 
 total spend: $5.61 API-equivalent
-no baseline at /home/user/mine/.claude/skills/paper-pipeline/pipeline-language.baseline.json — record one with --update-baseline
+no baseline at <project>/.claude/skills/paper-pipeline/pipeline-language.baseline.json — record one with --update-baseline

package/skills/paper-pipeline/repro/2026-08-07-language-eval-pilot.log

@@ -30,4 +30,4 @@ ru-original   100%   100%  +0pp   (4/4 vs 4/4)
 No gate: this file measures an effect, it does not enforce one. Read OVERALL first, then SPLIT BY ORIGIN — a gap that appears only on translated Russian is a finding about translation, not about Russian.
 
 total spend: $2.03 API-equivalent
-no baseline at /home/user/mine/.claude/skills/paper-pipeline/pipeline-language.baseline.json — record one with --update-baseline
+no baseline at <project>/.claude/skills/paper-pipeline/pipeline-language.baseline.json — record one with --update-baseline

package/skills/paper-pipeline/repro/2026-08-07-language-eval-raw.log

@@ -163,4 +163,4 @@ ru-original    76%    61%  -15pp   (41/54 vs 33/54)
 No gate: this file measures an effect, it does not enforce one. Read OVERALL first, then SPLIT BY ORIGIN — a gap that appears only on translated Russian is a finding about translation, not about Russian.
 
 total spend: $26.59 API-equivalent
-no baseline at /home/user/mine/.claude/skills/paper-pipeline/pipeline-language.baseline.json — record one with --update-baseline
+no baseline at <project>/.claude/skills/paper-pipeline/pipeline-language.baseline.json — record one with --update-baseline