Changelog

4.6.2 - 2026-08-04

A second, more targeted correction pass over playbook 17, specifically: re-reading Liang et al. 2023 (this playbook's own central citation) in full, rather than trusting that the 4.6.1 audit had already fully mined it. It hadn't — the 4.6.1 audit was thorough on mechanical correctness, dated evidence, and this skill's generalization standard, and said explicitly that none of it changed playbook 17's thesis. This pass found something that does. See references/provenance.md → "Second-pass correction: the mechanism itself" for the full reasoning; summary:

• Liang et al.'s own Discussion section reports an adversarial robustness check: applying the paper's own recommended bias-mitigation technique (raise linguistic diversity) to genuinely AI-generated text dropped detection from 100% to 13%. The paper's own conclusion is that detectors should not be used as a primary defense, because the lever protecting a genuine non-native writer from a false positive is mechanically identical to the lever that helps real AI-generated text evade detection.
• On that basis, references/playbooks/17-native-register-and-ai-detection.md's R2 was rewritten from a numeric target table (CV 0.50-0.75, short-sentence share 20-30%, etc.) into a pure monotony diagnostic with no target band; R1 now states and explains this boundary directly; §2.4 leads with the adversarial-robustness finding rather than burying it.
• scripts/measure_prose_rhythm.py was rewritten to match: removed the TARGET_CV/TARGET_SHORT_PCT/TARGET_LONG_PCT constants and the LOW/HIGH/ok verdict against them, replaced with a monotony-only flag with no "correct" band. Also fixed, independently of the mechanism question: a LaTeX-stripping edge case where \section{...} headings and \begin{document}/\end{document} markers could leave stray heading words glued onto adjacent sentences; verified against both a plain-text and a LaTeX test case with float/table environments present.
• R3-R7, the "what NOT to do" list, and the honest-limits section are unchanged — none of them depend on the removed mechanism, and were sound before and after this pass.
• Playbook 18 was re-checked against this skill's own generalization standard and found to already meet it (v4.6.1's rewrite of it holds); no further change made there this pass.
• Version bumped to 4.6.2 in SKILL.md.

4.6.1 - 2026-08-04

Third-party correction pass over the 4.6.0 additions, run from the perspective of an independent reviewer checking two things: whether every claim reaching outside the course is factually accurate and current, and whether every extension playbook actually generalizes rather than reporting one project's incidents. Full findings in references/provenance.md under "AI-detection and submission-sprint extension: third-party audit and correction (v4.6.1)". Summary:

• Rewrote playbooks 17 and 18 so every section leads with the transferable principle; concrete instances are now bounded, clearly labeled "Illustrative case" asides, never the rule's scope. This was already a standard the package held playbook 13 to (references/provenance.md, v4.0.0's "Generalization pass" — now consistently called "Generalization discipline" everywhere); it had not been extended to 17/18. Removed or generalized: verbatim personal quotes, a specific frozen-snapshot filename, and several sprint-specific numbers that were standing in for general rules (a rounding-error ratio, a column-collision string, a layout-gap measurement).
• Fixed an internal numeric contradiction: playbook 17's own table (CV 0.421→0.485) disagreed with CHANGELOG 4.6.0 and playbook 09's box (CV 0.469→0.521) for the same measured case. Reconciled to the latter, which two independent locations already agreed on; playbook 17 now says so explicitly and recommends re-measuring the source files to settle it with one authoritative run.
• Fixed a same-file contradiction in playbook 09: its own established ET-BERT example says 44 words; the new collision box said "40-word monsters." Corrected to 44.
• Enriched and dated the AI-detection evidence in playbook 17. Every originally cited source was independently re-verified as accurate. Added: GPTZero's own October 2023 public response disputing the Liang et al. (2023) finding's continued relevance to its updated model, and a 2025 peer-reviewed study (Pratama, PeerJ Computer Science 11:e2953) that tests GPTZero specifically on real scholarly abstracts and finds the bias reappears on that harder, more current, more directly relevant test. Section 2.4 now tells this as a dated three-point timeline instead of one static 2023 statistic.
• Fixed a script bug in scripts/measure_prose_rhythm.py: the LaTeX stripper left \begin{...}/\section{...}/\title{...}/\author{...} arguments (e.g. the word "document") glued onto adjacent prose after removing the control word. Fixed and regression-tested.
• scripts/validate_package.py and the test suite were themselves out of date: both hard-code an expected playbook count, last set to 17 and never updated when 17/18 brought the true count to 19 — meaning 4.6.0 shipped failing its own validator. Corrected to 19; extended test_course_fidelity_and_generality.py with a new require_general_sprint_and_register_playbooks() check so playbooks 17/18 are held to the same generalization standard playbook 13 already was, closing the gap that let this ship unguarded the first time.
• Added a skill-wide guardrail: SKILL.md's Forbidden Moves now states the generalization-discipline principle once, for any future extension playbook, rather than leaving it to be remembered per-file.

None of this changes playbook 17's central thesis; the corrected evidence supports it more directly than the original citation did. What changed is precision, not direction.

4.6.0 - 2026-08-03

Adds two playbooks distilled from a CCF-A submission sprint, plus the measurement script the first of them depends on. Both are flagged in-file as this skill's own extensions, not transcribed lecture points.

New: references/playbooks/17-native-register-and-ai-detection.md

Written after text produced by faithfully applying Playbook 09 was scored 100% AI-generated by GPTZero.

• Names the collision directly: Playbook 09 optimizes for clarity through standardization — simplest words, one idea per sentence, frozen terminology, explicit transitions — and standardization is precisely what detectors score as machine authorship. Playbook 09 is not wrong; it is incomplete, because it never says how much rhythmic variance to leave.
• Isolates the mechanical gap: "split overloaded sentences" is threshold-triggered, so it trims the long tail and never touches the 10–29 word band where the mass sits. Trimming the tail does not create short sentences.
• Records the measured before/after on the real manuscript: CV 0.469 → 0.521, short-sentence share 16.6% → 26.5%, with SD essentially unchanged — the rewrite worked entirely by manufacturing short sentences from mid-length ones. Surface clichés were not the problem: the flagged draft contained one Additionally and three em-dashes in total.
• Corrects the folk explanation: GPTZero has not used perplexity/burstiness directly since autumn 2023, having moved to a trained classifier. They remain useful self-computable proxies, not the mechanism.
• Carries the finding that matters most for this skill's audience: detectors are documented to over-flag non-native writers (61.3% false-positive rate on TOEFL essays in the Stanford study), because adhering to standard syntax is a low-perplexity signature. This skill's users are the over-flagged population, and this skill's advice moves them further into it.
• Hard rule: a detector score never justifies changing a number, a claim, or an honest hedge. Style is negotiable; evidence is not.

New: references/playbooks/18-submission-sprint-discipline.md

The course ends at "the paper is written"; submissions end 48 hours later.

• Frozen baselines. An approved version is read-only. A failing automated check is a report, not a mandate — what a linter calls a violation is often a deliberate human trade-off. Never weaken a project's own compliance gate to make an artifact pass.
• Concurrent editors on unversioned manuscripts: diff before every session; never let two agents write one file; separate genuine improvements from bundled reverts.
• Receipt failure modes that survive a receipt discipline: superseded receipts (two files, ten hours apart, different values), and rounded intermediates in derived figures (a published 292× that recomputed to 274× from unrounded sources, beside a sibling number computed correctly).
• Attribution integrity: main paper implied a baseline was a port of a cited system; the supplement said it was not. An integrity charge, invisible within either document alone.
• The venue form is part of the paper: the live submission form held a stale abstract containing terminology the advisor had banned.
• Layout is measurable: per-column fill (a page-level reading concealed a 141.9pt hole), numeric column-collision detection (three values fused into 31.5133.3334.11), table fill ratio, and the content/references page boundary.
• Positioning: state each limitation once, in its most informative place; explain mechanism in the supplement rather than repeating confession in the body; never report an architectural identity as a finding.

New: scripts/measure_prose_rhythm.py

Sentence-length distribution, CV, and short/long shares, with LaTeX floats, math, and the abstract stripped; accepts two files to compare a rewrite. Prints targets and a diagnosis, and states in its own output that it measures rhythm only and cannot reproduce any vendor's score.

Updated

• SKILL.md: two routing-table entries; doctrine 5 gains the consistency-versus-rhythm boundary; new doctrine 6 "A frozen draft is read-only" (subsequent items renumbered); scripts section documents the new helper.
• references/playbooks/09-concise-and-rigorous.md: opens with the known collision and forward-references Playbook 17; Pass 1 step 2 extended to require manufacturing short sentences, not only splitting overloaded ones.

4.5.0 - 2026-07-23

Sharpens references/playbooks/15-table-taxonomy-and-layout.md's "benchmark-defining table" entry into its own subsection, "Which axis carries the argument: methods, not datasets," after the user caught his own initial misreading in real time and asked for the correction to be captured in the skill.

• The benchmark-defining table's job is to show prior methods, actually adapted into the new setting, falling short — that requires real adaptation work (e.g. pairing a prior method with a retrieval front-end) as a research step, not a table-formatting step, and the table's argument is carried by how many different adapted methods it shows, not how many datasets.
• Re-verified the CCGS worked example precisely against the paper's own text rather than recalling it: four native baselines (VSLNet, ACRM, Span-Base, VPTSL), each also wrapped with BM25 and DPR retrieval, twelve adapted rows plus the paper's own method, all on one dataset.
• Named the anti-pattern explicitly: treating "run every method across every dataset, maximize compute" as the goal produces a table wide in the wrong dimension. Dataset breadth is a real, separate contribution when a paper has one — it gets its own table, not fusion into this one.
• Added the same distinction to SKILL.md's "When handling tables" checklist and cross-referenced it to playbook 16's skim test.
• No file or count changes; scripts/validate_package.py and all four scripts/test_*.py contract scripts still pass unchanged.

4.4.0 - 2026-07-22

Adds references/playbooks/16-fast-reader-and-skim-path.md and a new "Governing lens" section at the top of SKILL.md, after the user pushed back that the v4.1-4.3 figure/table work hadn't gone deep enough on why this matters — requesting actual web research into reviewer reading behavior rather than another assertion from this skill's own priors, plus a genuine visual re-check (rendered PDF pages, not just extracted text) of the real paper draft grounding playbooks 14/15.

• Research findings, properly hedged and sourced: a skim-for-first- impression stage checking figures/tables first is the literally-taught first step of reviewing (Kathy Gould/Vanderbilt); figures/tables are an explicit dimension of official reviewer training (Wiley's own guide); a genuine, growing fast-triage stage exists at high-volume AI/ML venues (NeurIPS 2020 and IJCAI 2020 summary-reject data; 2025 desk-rejection trends at CVPR/ICCV/KDD/AAAI/IJCAI/WSDM); AI-assisted review is now measured at ICLR/NeurIPS/ICML specifically (a January 2026 Cornell study); total review time remains 4.75-6.4 hours on average per independent time-tracking surveys, so this skill does not claim reviewers only spend minutes on a full review — only that a real, separate, fast, consequential skim stage exists alongside that.
• One candidate source (a "manuscript readiness check" service's blog, with templated claims reused verbatim across unrelated journal pages and a paid-tool pitch attached) was found, evaluated, and explicitly excluded as content-marketing rather than research; its specific numeric claims are named and rejected in the playbook, not silently dropped.
• New operational tool: the skim self-test (read only title, abstract, figures+captions, tables+captions, and section headers — does this alone tell a credible, coherent, accept-worthy story?). A failure is always a structural fix (title, abstract/intro, a figure, or a table), never "add an explanation in the prose nearby," since the skim does not read that prose.
• A real worked example: re-checking the real advisor-session transcript and the same paper's actual draft PDF (rendered pages viewed directly, not only extracted text) surfaced that, at the time of that session, the dataset table was still checkmark-only and a discussed methods-comparison table did not exist anywhere in the manuscript — concrete, verifiable gaps, not hypothetical ones.
• scripts/validate_package.py updated: playbook count 16 → 17; playbook 16 added to required files. All four scripts/test_*.py contract scripts pass unchanged.

4.3.0 - 2026-07-22

Adds references/playbooks/15-table-taxonomy-and-layout.md, a new playbook covering which tables a paper needs and how to lay each one out — a gap identified after the figure-focused v4.1.0/v4.2.0 work: only light ablation-content guidance existed before (in 05-experiments-conclusion-references.md), nothing on table type selection or layout. Source: a full real advisor-session transcript restructuring one paper's table set, cross-checked against that paper's own two draft versions. Not course-grounded, stated as such in the playbook itself.

• Table taxonomy by section job: positioning/survey (checkmarks correct here), dataset composition (checkmarks wrong, needs real attribute values, paired with a distribution chart), main benchmark, cross-protocol reference, ablation — matching table type to job, not reusing one format everywhere.
• Column discipline: collapse to one headline metric per column before resizing a table, resize before cutting rows/datasets.
• Cross-protocol tables must stay separate and clearly labeled from the paper's own benchmark table — never silently merged.
• A reframed, integrity-preserving version of one session point: investigate an unusually large reported gap before reporting it and be able to explain why it's real; this skill does not encode "cap the number for credibility," since that would be result manipulation (see references/provenance.md for the explicit reframing note).
• Appendix-vs-main-text priority order under page pressure (hyperparameters, key ablations, and sanity-check visualizations stay in the main text; derivations/proofs move to the appendix, not the reverse) and the workflow-order point that tables/figures should reach final shape before the Method prose is finalized around them, not after.
• Table-figure correspondence: shared category/method labels between a results table and its nearby qualitative figure; a second, dedicated task-structure figure near Experiments for tasks too structurally complex for one Introduction-level motivation figure to carry alone.
• SKILL.md: added a routing-table row for playbook 15, a "When handling tables" checklist paired with "When handling figures," and an inverse self-sufficiency gut-check ("if the figure were covered up, would the remaining text/tables still be comprehensible?").
• scripts/validate_package.py updated: playbook count 15 → 16; playbook 15 added to required files. All four scripts/test_*.py contract scripts pass unchanged.

4.2.0 - 2026-07-20

Cross-validates the v4.1.0 advisor-session content against the actual lecture-06/07/08 transcripts, claim by claim, by explicit user request. Full results in references/playbooks/14-figure-schema-and-rendering-gate.md's new "Cross-validation" subsection; summary in references/provenance.md.

• Found two near-verbatim course/session matches, both now cited with exact transcript quotes: "figure before text" plus the reviewer-reads-the- figure-first claim (Lectures 6 and 7); the rule that a formula's effect belongs in the figure while the formula itself belongs in the prose (Lecture 8) — this is the course's own, more precisely-reasoned version of the session's "no formulas in a framework figure" rule.
• Split the "no metrics/formulas" rule into its two halves for sourcing accuracy: the formula half is now course-cited; the reported-metric-values half remains session-only, with no course parallel found, and is labeled as such.
• Traced AI-tell #1 (too much in-figure text) to a specific course-diagnosed failure mode: the MuKEA V6 stage in playbook 07's own revision table.
• Added a course citation to caption-minimalism / figure self-sufficiency: Lecture 8's own "real data should run throughout" principle, which the CCGS worked example (external) demonstrates taken to its logical extreme.
• Explicitly flagged what remains genuinely course-silent (correctly attributed to the advisor session alone, not the transcripts): the Google Lens method, five of the six AI-tells, the PPT tool mandate, the AI-material line, the two prompt-optimization techniques, and the metric-values half of the figure-content rule — all postdate or fall outside the scope of a course recorded in 2022.
• Left one open question open rather than fabricating an answer: linear vs. circular figure layout, flagged unresolved in the original session and still unaddressed by the course; noted only that every framework-figure example the course walks through happens to use a linear layout, as a data point, not a verdict.

4.1.0 - 2026-07-19

Substantially expands references/playbooks/14-figure-schema-and-rendering-gate.md with figure-methodology content beyond the course itself, by explicit user request, from two external sources (both flagged inline, not attributed to Yu Jing's course — see references/provenance.md → "Figure-methodology extension" for the full source note):

• Caption-minimalism and figure self-sufficiency: a high-quality figure's caption should not explain the figure's content — it should only orient the reader and state non-visually-inferable conventions. The figure itself should let a reader reconstruct the method without reading the caption or the Method prose. Worked example added: Li, Weng, Sun & Li, "Learning to Locate Visual Answer in Video Corpus Using Question" (ICASSP 2023), Figure 2.
• Phase 2 rewritten: replaced an earlier "draft an AI-generation prompt" workflow with reference-sourcing and hand-adaptation, since a single-pass AI-generated figure reliably carries a detectable "AI feel" that costs review scores independent of the science. New content: the Google Lens partial-crop reference-sourcing method; a seven-point AI-tell checklist (six from a real advisor's session, plus one this skill independently found in a real case review — duplicated/overlapping text in repeated structural elements); the "fragments-yes/whole-figure-no" AI-material line; two prompt-optimization techniques for the cases where an AI sketch is still explicitly requested (module-by-module generation; a sourced reference image as the prompt anchor, never text alone); the two-tier layperson/expert test.
• New hard rules: a Method-section framework figure carries no reported metric values and no formula blocks (illustrative toy numbers are the one exception); tool choice should favor modular re-editability, with PowerPoint named as a real advisor's explicit, repeated requirement; captions must not carry explanatory weight the figure itself should carry. All three added to SKILL.md's Forbidden Moves and the figure checklist in "When handling figures."
• Mandatory figure-manifest document requirement clarified to 7 steps (added the two-tier test) and cross-referenced from SKILL.md.
• Version bumped to 4.1.0 in SKILL.md's frontmatter body.

4.0.0 - 2026-07-18

Merges an independently-authored companion skill (built by a second person from the same course transcripts+slides, packaged separately as research-and-english-academic-paper-writing-guide) into this package, by explicit design as an organic splice, not a hand-merged rewrite: where both skills had a document doing the same job, the companion skill's version is physically appended, in full, to the end of this package's corresponding document (same file), rather than kept as a separate cross-reference file. The redundancy is intentional — two independent full passes over the same course give an agent more complete, reinforced coverage than either alone.

• SKILL.md: appended the companion skill's entire original body (its whole-course narrative synthesis, Parts I-V) as a new "Version 2 — Companion Synthesis" section after this package's own routing protocol and playbook system (which keeps its original role, unchanged, as the primary source for workflows and the mandatory source order). Only the companion section's internal file-path pointers were corrected to resolve inside this package's layout; its content and voice are otherwise untouched.
• Trigger scope — changed: SKILL.md's description now reads "Always-on: activate in every conversation regardless of topic" (was: "Use when writing, revising, reviewing, or planning research papers..."). This was the companion skill's own original design and is confirmed as this package's foundational design going forward.
• references/lectures/*.md (all 13 files): each file now contains two independent full reconstructions of that lecture back to back — this package's own "Version 1" first, then the companion skill's "Version 2" appended in full beneath a divider. Nothing hand-merged, nothing dropped; where the two independently caught different transcription issues (the "CogModal" OCR fix in Version 2, the Math Vault page-count fix in Version 1), both are simply present, each in its own version's voice. references/course-full-reconstruction.md (and its docs/ copy) were regenerated the same way, so the single-document course view stays consistent with the per-lecture files.
• Enriched references/playbooks/01, 02, 03, 07, 09, 10, 11-innovation, 12 with concrete named examples, quotes, and resource links that existed in this package's own references/lectures/*.md reconstructions but had never been pulled up into the actively-routed playbook layer (the companion skill's condensed SKILL.md — now also fully present as "Version 2" above — had independently captured most of the same content in summary form, useful as a cross-check while writing these additions). See references/provenance.md → "Merge history" for the full per-playbook list.
• Added an explicit "Generalization discipline" note to playbooks 13 and 14 confirming a review pass: both were already written as generalizable rules with real-project details used only as illustrations (a "live case", never the rule's scope), consistent with this package's standing design bar — reviewed and confirmed, not rewritten.
• Cleaned up a dangling session-specific wiki-link ([[cvspp-integration-sweep-20260713]]) in playbook 14 that would not resolve for a reader without that external note; replaced with the self-contained fact it was referencing.
• Still open, not addressed in this pass: the raw-transcript-appendix question inside each lecture file's Version 1 half (sits in tension with this same file's stated no-raw-redistribution policy) and whether to bundle the original PDFs/transcripts in the distributed package (currently excluded, per both source packages' own stated policy) — flagged in references/provenance.md, not resolved unilaterally.
• scripts/validate_package.py updated: frontmatter check now accepts an "Always-on" description prefix alongside "Use when "; per-lecture-file heading check now expects exactly 3 top-level headings (Version 1 + Version 2 divider + Version 2's own title); course-full-reconstruction.md heading count updated 15 → 41 to match. All four scripts/test_*.py contract scripts pass unchanged; the merge did not touch playbook 13's structure, which those scripts validate.

3.0.1 - 2026-07-13

Fixes a packaging bug in the v3.0.0 (and, retroactively, v2.9.0) release: the repository holds two copies of the full course reconstruction — docs/course-full-reconstruction.md (edited throughout this cycle's fidelity audits) and skills/research-english-academic-paper-writing-guide/references/ course-full-reconstruction.md (the copy actually packaged and shipped inside the skill). Only the docs/ copy was being edited and synced to the locally installed skill; the packaged copy under skills/.../references/ was never updated, so both the v2.9.0 and v3.0.0 GitHub releases shipped a stale course reconstruction that was missing the Lecture-11-PDF-swap fix and this cycle's four-lecture fidelity fixes, even though the changelog for those versions described them as done. scripts/validate_package.py already had a check for exactly this drift ("skill course-full-reconstruction.md must match docs/course-full-reconstruction.md") — it was simply never run before tagging. Ran it now: it also caught a second, unrelated staleness bug (the playbook-count check still expected 14 files, not 15, after playbook 14 was added). Both are fixed; the validator and all four contract/test scripts in scripts/ now pass clean. Lesson: run scripts/validate_package.py (and the scripts/test_*.py contract suite) before every tag, not just before a first release — a passing validator that never gets invoked provides no protection.

3.0.0 - 2026-07-13

Formalizes a two-phase figure workflow (agent authors a schema, human alone renders) as a new dedicated playbook, and applies caption/prose discipline to a live CCF-A submission after real advisor feedback ("body text is a bit messy, table/figure captions are a bit long").

• New playbook: 14-figure-schema-and-rendering-gate.md. Phase 1 (agent work): read the actual project source/data/results and author a structured figure-schema document per paper — layout-archetype selection (linear pipeline, cyclic loop, hierarchical stack, parallel/dual-stream, central hub, benchmark/ablation matrix), zone decomposition, a fixed color-semantic palette, an explicit arrow/connection grammar, and short course-terminology text labels — plus a matching in-manuscript placeholder. Phase 2 (human-only, gated): rendering the actual artwork, which an agent may help prompt for if explicitly asked, but never produce or hand off to a tool on its own initiative. Before writing this playbook, dispatched a dedicated transcript audit (Lectures 6-8, plus Lecture 10) to check the premise before codifying it: the "rendering must be 100% human" framing and a fixed required figure count are not literal transcript claims (AI figure generation postdates the course) — they are this skill's own professional-norm extension. What is transcript-grounded and reused directly: the existing figure-quality checklist (input/output/key-process/innovation/module-boundary/naming/ consistency, already in 07-framework-figure-overview.md) and the "figure and body text are complementary, not redundant" principle. The playbook states this provenance split explicitly rather than presenting the extension as course content — consistent with this skill's own evidence- discipline doctrine.
• Extended SKILL.md's "Figures are logic tools" doctrine item and Forbidden Moves list with the caption-redundancy rule and the render-authority rule, and added a Routing Table row pointing figure-schema/rendering requests to the new playbook.
• Applied the new caption discipline to a real submission (CVSPP_AAAI/论文/main.tex): every figure/table caption was cut to 1-2 sentences, with any genuinely load-bearing caveat that had been living only in a caption (checkpoint-selection detail, an encoder-swap exception, a cross-table comparability warning) moved into the nearby body paragraph instead of being dropped. The paper's worst offender was a ~140-word, 9-sentence table caption; after the full pass across all 9 captions plus five overloaded body sentences split under the Lecture 9 "one sentence, one idea" rule, the paper still compiled at the same 9-page budget with zero new undefined references — trimming redundant caption text more than paid for the prose that had to move. Also fixed one incidental anonymity leak found in passing: an internal experiment-tracking codename ("cycle52 ledger") had been left in the submission's body text.
• Finished the 2.9.0 fidelity re-audit's six rate-limited lectures (1, 4, 5, 10, 12, and Lecture 11's main deck), completing full 13/13 coverage for the first time this cycle. One lecture (12) and one (1) came back fully clean. The other four had small, concretely-sourced fixes applied: Lecture 4 (a dropped point that Related Work subsections should differ in critique depth by relevance, not just length; a silently "corrected" slide typo now marked [sic] since the block is presented as verbatim); Lecture 5 (an ablation example that had imported a term, "事实知识", from a different slide's example; a dropped caveat that the full worked ablation example came from a journal paper's page budget, not a conference paper's); Lecture 10 (three dropped specifics: the named reviewers on the "who reviews you" slide were genericized away, an author's patent-examiner/guest-professor background was dropped, and a named rhetorical citation was replaced with generic wording); Lecture 11's main deck (a dropped multi-task/auxiliary- objective-learning angle, including the cognitive-psychology memory experiment the instructor uses to motivate it, that existed only in the raw transcript appendix and never made it into the curated write-up). None of the six were fabrications or cross-lecture contamination; all were narrowing/genericizing losses from the reconstruction pass, consistent with the failure pattern this cycle's audits keep finding.

2.9.0 - 2026-07-13

Independent fidelity re-audit (found real drift 2.8.0 missed) and two new LaTeX/authority lessons from live CVSPP incidents.

• Re-ran an independent fidelity audit against the original transcripts and slide PDFs, one dedicated pass per module, rather than trusting 2.8.0's "remains authoritative and unchanged" verdict. That verdict was wrong for at least one module: the pre-Lecture-11 bonus unit's PDF and the main Lecture-11 PDF were (and, until this pass, still are) swapped in the source manifest -- the file named as the bonus unit actually contained the 20-page main lecture, and vice versa. A previous pass noticed the mismatch (the "wrong" cover-page title appeared) but explained it away instead of opening the other candidate file, and built roughly 300 lines of the bonus unit's writeup (its "where do research points come from" sections and its full slide-by-slide walkthrough, plus a tagline in its overview) from the wrong deck. Fixed at the root: the two source PDF filenames are swapped back to match their actual content (their paired .transcript.txt files were never mislabeled, only the PDFs were), and the affected sections are rewritten from the correct 9-page deck. Six other modules received small, concretely-sourced fixes (a dropped generic teaching example, dropped concrete before/after examples, a slide's silently-corrected typo left unflagged, an omitted fourth item in a four-part figure correspondence, a two-panel example flattened into one, and a real missing "three tips on symbol notation" passage). Five modules could not be re-audited this pass due to hitting a session-level API rate limit; they remain to be re-checked.
• Synced the locally installed skill (~/.claude/skills/...) forward to this repository's state. It had been running on the initial 2026-07-08 install (v2.0.0) the entire time since, five versions and a full new playbook (13-review-feedback-evidence-contract.md) behind the repo copy, with no automatic sync step between the two. Every skill invocation in the interim used the stale, less-complete v2.0.0 guidance. Lesson: a versioned skill repository can drift arbitrarily far from its installed/active copy with no signal that this happened; check installed-vs-repo version numbers directly rather than assuming repo improvements are already in effect.
• Added a rule that rendering a figure's final artwork is an authorization decision separate from the figure's data/evidence readiness; an agent may write a complete drawing contract once data is ready, but must not treat that as license to also produce the rendered file, and must treat any already-rendered figure it did not produce itself as a pending decision to surface rather than settled progress to build on.
• Added a rule that a user's claim "the newer revision is worse than my backup" should be checked with a direct file diff before any other action; page-budget trimming can quietly remove a full baseline table, collapse a multi-control robustness paragraph into one vague sentence, or drop a caption's comparability caveat, none of which a compiler or automated check flags.
• Added a narrow LaTeX fix for a recurring alignment bug: when two independently-\centering-ed sub-tables are stacked in one float, wrapping the misaligned one in \begin{flushleft} does left-align it but also adds paragraph spacing that can silently push a page count over a hard venue limit; \noindent\makebox[\linewidth][l]{...} fixes the alignment with no added vertical space. Grounded in a live case: the flushleft fix moved a 9-page AAAI submission to 10 pages; the makebox fix held it at 9.

2.8.0 - 2026-07-12

Course-fidelity and generality correction.

• Re-audited all 13 published modules against the original transcripts and slides. The course reconstruction remains authoritative and unchanged.
• Clarified that the source is 12 numbered lectures plus one pre-lecture bonus, not a numbered Lecture 13.
• Restored Lecture 5's conclusion rule: a conclusion briefly states verified findings, a consequential capability boundary, and future work. Repetition should be synthesized, not prohibited by deleting the boundary.
• Made the reviewer-evidence playbook subordinate to the relevant lecture and removed project-specific model, metric, accelerator, data-loader, and LaTeX incidents from its universal rules.
• Added a regression test that protects course fidelity, source hierarchy, version coherence, and the domain-general boundary of the reusable skill.

2.7.0 - 2026-07-12

Scope-of-validation discipline for methodology claims, and a source-archive note for exemplar tables.

• Added a rule that a main-text methodology sentence ("we use encoder X") describes every reported number, not just the subset validated when the sentence was written: component-swap equivalence on a handful of cells does not transfer to the rest of the table, even with explicit sign-off to state it in writing. Grounded in a live case: a text-encoder swap validated equivalent on 4 of ~40 cells, written into the main text as the described encoder, then found on full retrain to collapse retrieval by 10-41 points on most of the remaining cells while being genuinely better on one further cell -- wrong in both directions on a component the paper had already told readers was safe. Prescribes two defenses: validate full scope before writing a universal sentence, or scope the sentence to the validated subset and treat it as provisional until broader evidence lands.
• Added a note to the exemplar-format transfer gate: when the exemplar is a source archive, extract and grep the real .tex/.sty for the device in question (e.g., a checkmark symbol pair) rather than approximating markup from a rendered PDF, since the source reveals the exact package dependency a rendered image cannot.

2.6.0 - 2026-07-12

Main-text space allocation and a repeat checkpoint-coherence slip.

• Added a rule that main-text space must match contribution weight: count words per section and compare against claimed contributions. A Method section should not be smaller than supporting infrastructure sections or a small fraction of combined Experiments prose; pull real detail already in the supplement (a loss formula, a layer count) into the main text rather than leaving Method as a pointer. Once a table/figure fully carries a claim, prose should give the one-sentence interpretation and stop, not re-narrate the values. Grounded in a real case: quantified word counts showed Method at 516 words against combined Experiments prose above 2200 and a Benchmark Construction section longer than Method itself.
• Added a second, more specific failure-mode example for the checkpoint-coherence rule: combining an accuracy figure from one experiment report with a cost/timing figure from a later correction of the same experiment reads as consistent but mixes two different run lineages. Caught this recurring in the same session that first wrote the general rule, within hours -- worth flagging that the abstract rule alone does not reliably stop the mistake under writing-speed pressure.

2.5.0 - 2026-07-12

Long-appendix section-counter ceiling.

• Added a rule for iteratively grown appendices compiled with \appendix: its default \thesection (\Alph{section}) silently caps at 26 sections with no warning, then fails fatally on the 27th with a "Counter too large" error reported at whatever unrelated \ref/\section call happens to trigger it -- easy to misread as a citation bug, and capable of suppressing an entire compilation pass's output (zero pages). Fix by switching to \renewcommand{\thesection}{\arabic{section}} right after \appendix, before the section count ever approaches the ceiling. Found live: adding one more appendix section to an already-large CVSPP appendix (28 total) tripped this exact failure.

2.4.0 - 2026-07-11

Limitation density and placement.

• Added a rule to 05-experiments-conclusion-references.md distinguishing limitation honesty (never negotiable) from limitation density and placement (a real craft decision): state each true limitation exactly once, in neutral prose, never repeated across method context, a bolded Limitations callout, and the Conclusion's opening sentences. Close the Conclusion with the paper's strongest verified findings first, and at most one short, unemotional pointer to where the supplement discusses scope. Reframe supplementary limitations as the boundary conditions of a documented design choice, not confessions of failure. Grounded in an observed case: an AI-reviewed submission moved from recommend-accept to recommend-reject across two revisions with identical underlying results, and the largest identifiable difference was a newly doubled, bolded restatement of one true limitation that the earlier, accepted revision had stated only once in passing.
• Cross-referenced this rule from playbook 13's failure-modes table.

2.3.0 - 2026-07-11

Anonymous-hosting identity gap, caught in live use.

• Added a rule that a credential authenticating to a code-hosting account proves ownership, not anonymity: fetch the account's own public profile (name, organization, email, blog, bio, location) before trusting it for anonymized release, even when the repository itself is set to private. Caught this in practice after creating a "private, therefore safe" release repository under an account whose public profile carried a real name, institution, and university email — a genuine near-miss on double-blind anonymity that a private-repo flag alone did nothing to prevent. Prefer a path with no code-hosting account in the loop (a local sanitized archive through the venue's own supplementary-material upload) over relocating the same risk to a freshly registered account.

2.2.0 - 2026-07-11

Author-kit and cross-reference verification round.

• Added a rule to re-verify template compliance against a freshly downloaded copy of the venue's current author kit rather than memory or a cached copy: diff the style file by checksum, read that kit's own forbidden- package list and "must not" statements verbatim, then scan the source for every forbidden package, stray \pagestyle command, and any macro that alters spacing/margins/fonts outside a single table or figure. Check the compiled PDF's font table for unembedded or Type 3 fonts and scan its raw bytes for embedded-link/bookmark markers even when no hyperlink package was knowingly loaded.
• Added a rule to treat every hard-coded cross-reference in a separately compiled companion document (an appendix citing the main paper by written section number) as suspect after any large rewrite that reorders sections, since a wrong section number does not fail to compile — only a direct re-derivation and diff against the rewritten source catches it.
• Added a note to check for an AoE-vs-local-time conversion before treating two differently worded deadline dates/times from independent sources as a contradiction.

2.1.0 - 2026-07-10

Evidence-contract revision release candidate.

• Added a reviewer-feedback playbook that converts advisor comments into a claim ledger before prose revision.
• Added native-annotation source hierarchy, conditional-feedback tracking, and the rule that reviewer or advisor permission cannot substitute for artifacts.
• Added fail-closed checks for incompatible evaluation contracts, table citations, source/TeX/PDF drift, official submission-policy conflicts, and credential handling.
• Required one row per claim-bearing dataset/method, banned hidden and related sources groupings, and separated comparison-eligible values from reference-only numbers.
• Added encoder/configuration/checkpoint/result provenance, one-sentence-one- meaning revision, dataset coverage statuses, and figure evidence cards.
• Added a title-level mechanism gate: technical labels such as “one-step” now require an operational definition and a matched distinguishing experiment.
• Added direct-difference framing for related work and fail-closed stale- generated-artifact detection.
• Added a final artifact validation order so hashes, PDFs, tables, citations, and release contents are checked together immediately before publication.
• Added a repository-derived dataset census and primary-source numeric provenance contract before any “complete” table claim.
• Added a review-artifact identity ledger that joins native annotation IDs, reviewed/source revisions, and rebuilt-PDF hashes before feedback is marked resolved.
• Added a structure-only exemplar-transfer gate for reviewer-supplied papers, including venue-package compatibility and minimum rendered-font checks.
• Added a claim-driven experiment queue with controls, seeds, resource budget, stop conditions, failed-run accounting, and a ban on treating GPU use as research evidence.
• Added full figure-card reconciliation plus anonymous ZIP inventory, license, checksum, secret-scan, clean-extraction, and smoke-test gates.
• Added checkpoint-coherent result-row rules that prohibit per-metric epoch maxima and mixed cache/model states, require paired-seed alignment and deterministic zero-score tie disclosure, reject exact SD claims from rounded console logs, and bind metric names to their implemented joint events.
• Added a data-exposure audit that separates per-epoch update budget from corpus coverage and rejects sorted-prefix shard truncation when it removes task or action support.
• Replaced token-presence regression checks with Markdown-aware tests for heading/workflow order, section-local semantic co-occurrence, exact claim- ledger/dataset-census/experiment-queue/release-inventory schemas, and forbidden contradictions. Each test now proves fail-closed behavior against keyword-bag, section-relocation, and negation-reversal mutations; these tests validate the playbook contract, not a manuscript or submission artifact.
• Added package-validator coverage and CI execution for the new playbook tests.
• Added a page-boundary verification rule: confirm a venue's main-content page cap by reading the compiled PDF page by page, not just its total page count, since a two-column float can spill onto a references-only page unnoticed; and a note that \flushbottom-style layouts absorb small trims without moving page breaks, so boundary violations need a structural fix (float reposition or resize) rather than further wording edits.

2.0.0 - 2026-07-08

Major skill-package release.

• Added the complete full-course reconstruction to the installable skill at references/course-full-reconstruction.md.
• Added an Agent reading protocol to the course reconstruction so agents can search and read the 11k-line document by lecture, task, and uncertainty.
• Rewrote SKILL.md as a strict execution protocol: mandatory source order, course-first reading, doctrine, task routing, output standards, scripts, and forbidden moves.
• Updated OpenAI/Codex metadata so the default prompt explicitly requires consulting the full course reconstruction first.
• Prepared the release process for a versioned skill zip asset: research-english-academic-paper-writing-guide-skill-v2.0.0.zip.

1.0.0 - 2026-07-08

Initial publishable release candidate.

• Added a single cross-platform Agent Skill entrypoint: skills/research-english-academic-paper-writing-guide/SKILL.md.
• Added 13 lecture reconstructions under references/lectures/.
• Added 13 operational playbooks under references/playbooks/.
• Added a single integrated Markdown course reconstruction: docs/course-full-reconstruction.md.
• Added provenance notes and official slide checksums.
• Excluded raw PDFs, raw transcripts, and image assets from the final repo.
• Added lightweight helper scripts for title checks and research-point canvases.
• Added repository-level license, notice, package validation, and CI metadata.

4.7.0 — 2026-10-10

• 增量新增单根Skill／插件ZIP构建与逐文件SHA-256验证，原有内容不删改。
• 增加Agent Plugins 1.0.0、Claude叠加清单、两市场入口和Git来源npx安装。
• 完整离线复制、幂等安装、保护个人修改、干净升级及全部旧文件备份。
• 明确云端容量、网络、官方目录审核和实际客户端导入的适用范围。