Academic Research Skills for Claude Code: research → write → review → revise → finalize

文件	最后提交记录	最后更新时间
.claude-plugin	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
.claude	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
.github	feat(review-form): fixed-point reminder that the author decides the review form (#921) (#922) Whether to run a systematic review is the author's decision. ARS now shows one fixed note listing five review forms (systematic, scoping, narrative or integrative, rapid, no formal review), each with one line on what it serves and roughly what it costs, with no default, ranking, or recommendation. - shared/references/review_form_note.md holds the rules and the note text in English and Traditional Chinese; every other language gets the English text. deep-research/SKILL.md and academic-paper/SKILL.md carry the block verbatim. - The note appears at the first of two author actions: selecting lit-review mode, or confirming the research question (full mode's RQ Brief before Phase 2, or the Socratic closing RQ Brief or RQ Summary). It never depends on the question's content, is skipped when the author already named a form, and appears at most once per project or run. The session waits for the reply; skipping counts as a decision. With a passport file the note and the reply go to the run ledger as a review-form-note checkpoint, and an unanswered checkpoint is resumed rather than reopened. - Paths that could move into systematic-review mode on the model's reading now require the author's choice (three-way-scan escalation, the lit-review -> systematic-review transition, lit-review trigger examples), and requests to conduct a systematic review no longer route to academic-paper lit-review. - scripts/check_review_form_note_sync.py keeps the copies byte-identical and checks each note's five labels, order, neutrality sentence, quoting, ranking words, and selection marks; 34 mutation tests; wired into spec-consistency and the pytest manifest. - Routing fixtures 13-16 add the negative case, its content-independence pair, a named-form case, and a broader-synthesis case; 02 and 04 now expect the note. The fixtures have not been run; session behaviour is prompt-level and unmeasured. Cost lines are sourced in the canonical file (Borah et al. 2017; Tricco et al. 2018; Garritty et al. 2021; Whittemore & Knafl 2005). Cross-model review (gpt-6-astra, xhigh) converged in three rounds: 9 -> 2 -> 0 P1/P2. Closes #921 Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	1 个月前
academic-paper-reviewer	feat(reviewer): check the panel's findings against the manuscript (#1002, part two) (#1017) * feat(reviewer): check the panel's findings against the manuscript (#1002) An advisory attachment to the Editorial Decision Letter. After the decision is final, the session sends the five Phase 2 reports' findings and the manuscript to one subagent with the part 1 auditor prompt, seats hidden behind neutral ids. scripts/audit_panel_claims.py builds the prompt and renders the reply; a false verdict shows as contradicted only when its quote passes the part 1 mechanical check. The report never changes the decision, the consensus counts, or the roadmap. ARS_PANEL_CLAIM_CHECK=0 turns it off. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(reviewer): name the tested judge models in the attachment; CHANGELOG (#1002) Refs #1002 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(changelog): match the #1002 entry to the skill text Refs #1002 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(reviewer): claim check review round 1 (duplicates, unread seats, one-pass prompt, CLI stops, blinding wording) Refs #1002 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(reviewer): count only string claim ids (review round 2) Refs #1002 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(reviewer): the #1001 attachment now precedes the #1002 claim check Refs #1002 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(reviewer): claim check security review round 1 (literal letter text, quote-only wording, unreadable W headings) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(reviewer): claim check security review round 2 (Severity lines outside W headings, quote-words-only wording) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(reviewer): claim check security review round 3 (Severity lines read by position) The scoring-report check now walks the Review Body line by line, as w_claims does. A Severity line must sit under a heading that w_claims reads; a heading line is a title and does not count. Matching totals could no longer hide a finding under an unreadable heading. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(reviewer): claim check summary counts answered findings, wrapped quotations, tested models The summary now says how many findings the judge answered, not only how many there were. A quotation in a finding can wrap onto the next line. The note names every model that passed the #1002 contrast test. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	10 天前
academic-paper	docs(guard): state the write-scope fence's actual coverage; pin real hook payloads (#973) (#974) Closes #973 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	22 天前
academic-pipeline	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
agents	docs(guard): state the write-scope fence's actual coverage; pin real hook payloads (#973) (#974) Closes #973 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	22 天前
audits	docs(models): support Claude Opus 5.5 alongside Fable 5.1 (#883) (#884) Two-reader audit of the Claude Opus 5.5 system card (audits/harness-retirement-2026-09-opus-5-5.md); no guardrail retired. The docs name both session models with effort, cost, tiering, and paste guidance; shared/model_tiering.md defines family and tier and corrects the classifier-fallback note; the revision coach inlines the instruction/data boundary (DG-1), with the lint's mutation tests parametrized over every hot-spot agent; docs/RISK_REGISTER.md R1/R3/R5/R11 point to the new evidence. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SK9a1EanckyNXLe1ZdGXEx	1 个月前
commands	feat(review-form): fixed-point reminder that the author decides the review form (#921) (#922) Whether to run a systematic review is the author's decision. ARS now shows one fixed note listing five review forms (systematic, scoping, narrative or integrative, rapid, no formal review), each with one line on what it serves and roughly what it costs, with no default, ranking, or recommendation. - shared/references/review_form_note.md holds the rules and the note text in English and Traditional Chinese; every other language gets the English text. deep-research/SKILL.md and academic-paper/SKILL.md carry the block verbatim. - The note appears at the first of two author actions: selecting lit-review mode, or confirming the research question (full mode's RQ Brief before Phase 2, or the Socratic closing RQ Brief or RQ Summary). It never depends on the question's content, is skipped when the author already named a form, and appears at most once per project or run. The session waits for the reply; skipping counts as a decision. With a passport file the note and the reply go to the run ledger as a review-form-note checkpoint, and an unanswered checkpoint is resumed rather than reopened. - Paths that could move into systematic-review mode on the model's reading now require the author's choice (three-way-scan escalation, the lit-review -> systematic-review transition, lit-review trigger examples), and requests to conduct a systematic review no longer route to academic-paper lit-review. - scripts/check_review_form_note_sync.py keeps the copies byte-identical and checks each note's five labels, order, neutrality sentence, quoting, ranking words, and selection marks; 34 mutation tests; wired into spec-consistency and the pytest manifest. - Routing fixtures 13-16 add the negative case, its content-independence pair, a named-form case, and a broader-synthesis case; 02 and 04 now expect the note. The fixtures have not been run; session behaviour is prompt-level and unmeasured. Cost lines are sourced in the canonical file (Borah et al. 2017; Tricco et al. 2018; Garritty et al. 2021; Whittemore & Knafl 2005). Cross-model review (gpt-6-astra, xhigh) converged in three rounds: 9 -> 2 -> 0 P1/P2. Closes #921 Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	1 个月前
deep-research	fix: block missing required inputs in the state machine; drop ranking from the Socratic example (#990, #992) (#994) The Missing Materials section of pipeline_state_machine.md let the user skip a required input at their own risk, while handoff_schemas.md rule 5 and the state tracker's gap table block the transition. The section now follows the state tracker: required inputs return HANDOFF_INCOMPLETE and are backfilled, recommended inputs do not block, and integrity checks still cannot be skipped. The #528 content lock is updated. The Socratic mentor's direct-answer example praised the user's "second idea", which the Research-Question Authorship Boundary forbids. It now asks which of the user's own directions they most want to understand. The ideation-diversity non-production variant pins the new source hash. Closes #990 Closes #992 Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	14 天前
docs	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
evals	feat(evals): Claude Opus 5.5 and Fable 5.1 pass the #1002 contrast gate (plan v10) (#1016) * feat(evals): contrast-only Claude judges for the #1002 part-two judge test (plan v10) Refs #1002 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(evals): check the Claude judge's temp folder chain before the lock (round 1) Refs #1002 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(evals): record the plan v10 contrast results for the Claude judges Refs #1002 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(changelog): plan v10 contrast test (#1002) Refs #1002 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(evals): a Claude reply needs exactly one successful result event (review round 1) Refs #1002 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(evals): state the one-result rule in plan v10 Refs #1002 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	10 天前
examples	docs(readme): correct front-page and showcase claims to match the evidence (#908) * docs(readme): correct front-page and showcase claims to match the evidence - Tagline: say that ARS can draft a whole paper in full mode, that the pipeline stops for confirmation at every stage, and that the user remains the author and answers for every submitted claim. - Showcase: add a version note (academic-pipeline v2.3, March 2026; the current gates were not measured on this paper; why the byline names Claude); drop the 10-stage label; restate the Stage 2.5 counts as the report gives them (8 bibliographic errors, 6-8 likely fabricated); describe the post-publication audit as a Claude Code + WebSearch run that found problems in 21 of the 68 final references, 19 of which were removed or corrected; call the Round 1 panel role-separated reviews generated by Claude, not independent reviews. - Calibration (README, ARCHITECTURE): the gold set has 25 synthetic tuples, the shipped test uses a stub judge that returns the gold labels, and no live-judge result is recorded. - Pipeline section: "guarantees" becomes rules that are protocol, not runtime guarantees; Schema 11 maps each reviewer concern to the author's revision claim and records whether re-review verified it. - Cost and time (QUICKSTART, README cost pointer): use the 2026-09 list-price estimate from docs/PERFORMANCE.md (about US$3-7 per run, before cache discounts); collaborative work spans hours to days. All six README locales carry the same changes, per the README sync rule in CONTRIBUTING.md. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(readme): keep the audit row's translations as neutral as the English The English row says 21 of the 68 final references still had problems after three rounds of integrity checks. The zh-TW, zh-CN, ja-JP, and ko-KR rows said the checks missed those problems, which is stronger: several of the affected references had been flagged at Stage 2.5 and marked FIXED before the audit found other errors in them. The four rows now say the problems remained after the three rounds. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(changelog): record the front-page claim corrections (#908) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	1 个月前
hooks	fix(guard): one Python process per tool call (#1015) (#1019) * fix(guard): launch Python through py -3 once per tool call, not three times (#1015) On Windows, each launch through the py App Execution Alias makes AppXSvc run an update check that leaks memory. The launcher's probe now prints sys.executable, and the guard run and the JSON check use that path. With no usable path they use the candidate as before. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(guard): #1015 review round 1 (Windows-only path, exact probe line, -I) The probe prints sys.executable only on Windows, and only when stdout can encode it, so a POSIX wrapper's environment still applies. The probe line must be exactly "<marker>\|<path>\|", so a candidate that prints more is skipped as before. The guard run and the JSON check pass -I. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(guard): #1015 review round 2 (one "\|" pair in the probe line, ASCII path only) A probe line with a "\|" beyond its two separators is skipped, as before. The path is used only when it is ASCII, since Python's stdout encoding and Git Bash can read other bytes differently. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(guard): one Python process per tool call runs the probe, guard, and check (#1015) Replace the interpreter-path approach. Each candidate now runs one process: a short runner writes the marker to fd 1, runs the guard in-process with its stdout captured and fd 1 on the null device, and writes the decision as ASCII JSON only when the guard exited 0 with a JSON object holding hookSpecificOutput. The runner uses -I and ends with os._exit. Text before the marker or after the decision skips the candidate; a failure after the marker gives a pass-through. The timeout call passes -k 1s. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(guard): #1015 review round 1 (UTF-8 stdin, end line, exit status, stderr flush) The runner reads stdin as UTF-8, since -I drops PYTHONUTF8 and PYTHONIOENCODING. The decision is followed by an end line, written in a loop, so trailing or cut-off output skips the candidate. SystemExit codes map to the status Python gives them, and stderr is flushed before os._exit. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(guard): #1015 review round 2 (marker-only output is a broken guard, wide exit codes) Only the marker with exit 0 means the guard ended the process itself, so it gives a pass-through and no other candidate runs. An exit code outside the C int range counts as a failure. The trailing-output test checks that the next candidate ran. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	10 天前
pi	fix(pi): compose ARS with structured prompt sections (#985) When Pi provides systemPromptOptions.sections and no earlier extension has set forceSystemPrompt, the wrapper edits the structured options instead of returning a rendered systemPrompt, so other extensions' sections survive. Older runners and earlier complete-prompt overrides keep the rendered fallback from #880. Closes #966.	17 天前
plugin-evals-citation-check	feat(commands): /ars-citation-check inherits the session model (#912) * feat(commands): /ars-citation-check inherits the session model Drop the `model: sonnet` pin from the citation-check command so it runs on the session model, like the three heavy commands. The Sonnet 5.5 system card (§6.3.2.1) reports a higher incorrect-answer rate on closed-book factual questions for Sonnet 5.5 (27%) than for Opus 5.5 (17%) or Mythos 5.1 (21%), and citation checking depends on factual recall. The other twelve light commands keep the sonnet pin. The SessionStart announcement and the citation-check eval README now match. Eval cases keep `model: sonnet` as a fixed eval model. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GtcbD3kmENUCYaNYEWTSNf * docs(evals): describe the eval model as a moving alias `sonnet` resolves to each new Sonnet release, so eval runs are comparable only at matching resolved agent and judge versions. Replace the "fixed, reproducible" wording in the eval README and CHANGELOG. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GtcbD3kmENUCYaNYEWTSNf --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	1 个月前
plugin-evals-citation-locale	fix: restore citation mode loading and Chinese APA checks (#857, #882) (#881) * fix: route plugin modes through core skills and bound descriptions * docs: retain citation retest failures separately from routing coverage * fix: preserve Chinese APA author and ordering rules (#882) * docs: record final Chinese citation regression outcomes	1 个月前
plugin-evals	fix(academic-paper): keep peer review out of the committee-correspondence variant (#854) (#900) The revision coach's committee guard now names the case it must not route: journal or conference reviewers, editors, area chairs, and program committees are peer review, not a committee for the #668 committee-correspondence variant, even when the user names the venue or the venue calls the role a committee. The sentence sits in the three places a run reads before it can build committee output: the academic-paper SKILL.md routing paragraph, the revision coach agent, and the committee-correspondence protocol's activation gate. A test pins the whole sentence in each file. plugin-evals case 03 (ICLR reviews) gains a with-only regex grader that fails when an answer emits a line of the committee variant's output; an answer that declines the variant in prose passes. On Claude Opus 5.5 the retest did not reproduce the misroute (0 of 7 runs with the guard, 0 of 7 on main), so the guard's effect is unmeasured. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>	1 个月前
scripts	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
shared	fix(reviewer): suggested-reference partial marker names either part (#1001) (#1018) The panel_reference_check_partial diagnostic marker joined the 'Not read:' list and the 'Not checked' result with "and", but render() prints each only when its own case occurs. The marker now says the coverage line comes with either one, or both. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	10 天前
skills	feat(sr-screener): add protocol-driven study screening as the fifth skill (#919) (#920) * feat(sr-screener): add protocol-driven study screening as the fifth skill Adds sr-screener v1.0.0: turns a proposal or PROSPERO protocol into user-confirmed eligibility rules, then screens titles/abstracts and full texts with two blinded reviewer subagents and a third-reviewer adjudicator. Eight modes; standard-library scripts for parsing, de-duplication, merging and PRISMA/RIS/literature_corpus outputs. Registers the skill on every inventory surface and lint registry (skills/ symlink, CLAUDE.md table + routing rule 8, manifests, MODE_REGISTRY, mode spectrum, announce routing, model-tiering manifest, Bucket C rosters, agent mirror, data_access pin, instruction/data boundary list), updates docs/READMEs, and adds scripts/test_sr_screener_pipeline.py. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * chore(sr-screener): keep only the skill folder; drop repo-wide registration changes Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Revert "chore(sr-screener): keep only the skill folder" The maintainer asked for the skill to be registered on the inventory, model-tiering and instruction/data-boundary surfaces (#919). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(sr-screener): address #919 review (Sonnet defaults, required QC, pilot gate, routing) - Reviewers A/B (and every screening role) default to Sonnet; no profile offers a smaller model; the cost check prints any per-role override. - The QC recheck of records both reviewers excluded is required: a reproducible sample (min 20, default 100) is drawn once screening is complete and stays pending until the senior reviewer decides it. - The full title/abstract run needs a pilot compared with the team's own labels (merge_decisions.py --pilot-labels) that missed no record the team advanced; --pilot-override records a reason for the methods text. - Routing: a Screening boundary paragraph in the routing core and its six copies (screening starts only when the user asks; no automatic handover from systematic-review mode); fixtures 13-17 incl. three negative cases. - Registration restored (inventory, model tiering, boundary block). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(sr-screener): close model, pilot and full-text QC review gaps * docs(changelog): point the sr-screener entry at renumbered fixtures 17-21 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(sr-screener): bind review revisions and close screening report gaps Freeze the first pilot's scope, require audited scope expansion, reject runtime job replacements and validate the actual dispatched IDs. Bind generated labels and pilot comparisons to review, dataset, protocol and config identities. Amendments archive prior decisions/QC state and begin a new revision. Reject foreign or obsolete results; explicitly verified legacy imports require an audited reason. Preserve prior results and decisions in the local audit trail. Derive completeness from every prepared record/report and validate full-text retrieval coverage against the current TA decision snapshot. Escape spreadsheet text, compare complete normalized titles without conflicting identifiers, reject ambiguous/prefix PDF matches, refresh empty RIS categories and diagram reports awaiting classification. Replace personal-looking examples with marked synthetic ones and advance the letter whose abstract does not establish whether it has data. Validation: 77 synthetic screening regression tests; 12 relevant spec/security/ inventory/version checks pass with repository LF content. Broader companion tests yield 476 passes and the same 71 Windows failures on both this change and the unchanged bcf394b baseline. No live model calls or screening accuracy claims. * fix(sr-screener): preserve stage decisions across reporting edits Hash the effective screening inputs for each stage, excluding reporting and preparation metadata and models/codes used only by the other stage. Preserve existing result identities when upgrading an unchanged whole-config hash; refuse to clear a legacy review whose original config cannot be verified. Quote every field in CSV exports, synchronize the screening agent mirror, and accept either preserved or normalized carriage returns in the XLSX test. Add regressions for CSV quoting, completed decisions and pilot preservation, stage-specific amendments, and safe legacy identity upgrades. Validate both openpyxl XML backends, the agent mirror, and repository consistency checks. * fix(sr-screener): neutralize formulas after alternative CSV separators Escape formula prefixes after embedded semicolons and tabs in spreadsheet exports, while preserving original bibliographic records. Keep QUOTE_ALL. Exercise all four formula prefixes with both alternative delimiters across duplicates, not-retrieved, audit, pending and screening-log exports. Verify the exported values with comma readers and semicolon/tab readers. Validation: 105 screening tests and the real-repository agent mirror test pass; all 12 repository consistency checks pass. --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Imbad0202 <132531341+Imbad0202@users.noreply.github.com>	1 个月前
sr-screener	fix(sr-screener): sniff only the CSV delimiter; read quotes by RFC 4180 (#979) (#982) * fix(sr-screener): keep RFC 4180 quote doubling after sniffing a CSV (#979) csv.Sniffer inferred doublequote from the first 5,000 characters. A sample without a doubled quote turned it off, so a later quoted field was split: comma exports lost the year and moved abstract text into the DOI or journal; semicolon exports kept the fields but added stray quotes to the abstract. Keep the sniffed delimiter and always enable quote doubling. Closes #979 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E7yiDzBvMyYSU3KY5xjw3b * fix(sr-screener): sniff only the CSV delimiter; read quotes by RFC 4180 (#979) Review of the first fix found three more misreads from trusting a 5,000-character sample: apostrophes made "'" the quote character, a quote-free sample split a later `, "a, b"` field, and a single-column export with semicolons was read as semicolon-delimited. Sniff only the delimiter, fall back to a comma when the header row lacks it, and use a fixed RFC 4180 quote dialect with a space allowed before a quoted field. Refs #979 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E7yiDzBvMyYSU3KY5xjw3b * fix(sr-screener): decide single-column CSV from the header record (#979) Checking only the first physical line misread a header whose first quoted field holds a line break and returned no records. Read the first record with the CSV reader instead. CHANGELOG now says the leading-space allowance is an extension beyond RFC 4180. Refs #979 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E7yiDzBvMyYSU3KY5xjw3b --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	19 天前
tests	fix(routing): say in the routing core that a mid-message [direct-mode] token is not the escape hatch (#892) (#961) Adds one Step 0 sentence to the routing core and its six copies, re-pins the academic-pipeline/SKILL.md content lock, and records the 2026-10 #133 fixture pass. Fixture 06 on Claude Fable 5.1 in the plugin install moved from 0 of 6 to 3 of 6; the acceptance rule (6 of 6) is not met, so #892 stays open. Fixture 16 failure tracked in #960. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	25 天前
tools	feat(ci): command-invariants gate + announce sync to 16 commands (#486) - scripts/announce-ars-loaded.sh: add missing /ars-3w and /ars-rebuttal-audit to both announce forms; stated count 14 -> 16; script version 1.0.0 -> 1.1.0 - vendor release-discipline-toolkit v0.3.2 under tools/release-discipline/ - .command-invariants.toml: per-repo lint config (ars token pattern, per-form region pattern so drift in one announce form cannot hide behind the other, count pattern, plugin.json <-> CHANGELOG version lockstep) - .github/workflows/command-invariants.yml: CI gate Claude-Session: https://claude.ai/code/session_01PVcesfz4RTLdfekQH5XcX6 Co-authored-by: Claude Fable 5 <noreply@anthropic.com>	4 个月前
.command-invariants.toml	feat(ci): command-invariants gate + announce sync to 16 commands (#486) - scripts/announce-ars-loaded.sh: add missing /ars-3w and /ars-rebuttal-audit to both announce forms; stated count 14 -> 16; script version 1.0.0 -> 1.1.0 - vendor release-discipline-toolkit v0.3.2 under tools/release-discipline/ - .command-invariants.toml: per-repo lint config (ars token pattern, per-form region pattern so drift in one announce form cannot hide behind the other, count pattern, plugin.json <-> CHANGELOG version lockstep) - .github/workflows/command-invariants.yml: CI gate Claude-Session: https://claude.ai/code/session_01PVcesfz4RTLdfekQH5XcX6 Co-authored-by: Claude Fable 5 <noreply@anthropic.com>	4 个月前
.gitattributes	fix(#454): Windows Python hook portability + graceful no-Python degradation The PreToolUse write-scope guard was wired as a bare `python3 ".../ars_write_scope_guard.py"`. On Windows `python3` is commonly a 0-byte Microsoft Store alias stub (exit 49, no output), so the hook failed before the guard's own fail-safes ran and spammed the hook log every call. A new cross-platform launcher `hooks/run_guard.sh` (POSIX sh; `hooks.json` invokes it via bash) finds a real interpreter — `py -3` / `python3` / `python`, each verified by a marker probe that must exit 0 AND print the marker (a stub is skipped) — then runs the guard as a supervised, time-bounded subprocess. Plan A graceful degradation: if no real interpreter is found or the guard subprocess misbehaves, the launcher emits a valid pass-through hook JSON and exits 0, never blocking and staying silent on stderr on the degraded paths. Hardened across a two-model dual-track review (codex + gemini, both POSIX-reproduced): the no-`timeout` watchdog fallback's stdin/stdout/timeout handling, a pid-reuse race (done-file handshake), a predictable-/tmp symlink fail-open (mktemp failure degrades to pass-through), and a gameable CI exec assertion (line-anchored). 76 tests (21 launcher + 55 guard); README documents the Git Bash prerequisite on Windows. Verified against the reporter's confirmed environment (py -3 real, python3 an exit-49 stub): the stub is skipped, the guard runs via py -3, and the decision is forwarded. Closes #454. 🤖 Generated with [Claude Code](https://claude.com/claude-code)	4 个月前
.gitignore	evals: add `claude plugin eval` suite for the academic-paper citation-check flow (#859) * evals: add `claude plugin eval` suite for the academic-paper citation-check flow Eight cases (six fire, two negative) under plugin-evals-citation-check/, graded as a with/without-plugin ablation. Fire cases carry a synthetic source pack so the four author-defined citation failures (no source on file, wrong authors, hedged finding cited as established, retracted or concern-flagged paper cited as live) are detectable offline. Styles: APA 7 (en / zh-TW mixed / es), IEEE, Vancouver (style unnamed), Chicago NB. Cases pin model: sonnet; run with --judge-model opus. Calibration (two pilots, 2026-09-13) and caveats are in the suite README. The with-plugin arm cannot load the mode prompt in the eval sandbox because the command stub references plugin files by relative path (#857), and plain-language prompts fired the skill in 3 of 6 cases (#858; the Spanish case is one data point for #850). No uplift figure is claimed. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EEbmXsHyQ3dzYoep74zSmH * evals(citation-check): apply cross-model review findings and re-calibrate Twelve of thirteen review findings applied: replace the disputable four-author "et al." planting in 04 with a year mismatch; make the clean citations in 02 and 06 supported by their abstracts; replace real Taiwan journal names in 02 with fictional ones; tie the 08 presence regexes to an "unused" statement; require metadata preservation and reject audit content on the 07 conversion negative; add no-overreach to 05; turn the 02 language check into an llm grader; exempt unchanged entries in a complete corrected list from no-false-positive and drop its DOI claim; drop the 04 style-name regex (both arms fixed the year without naming the style). The one rejected finding (08 skill check "display-only") was wrong: it carries arm: both and is scored; README says so. Pilot 3 on the revised suite: $4.65, max 142 s / 7 turns / $0.45 per run; 04 and 08 re-run clean after the last two grader fixes. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EEbmXsHyQ3dzYoep74zSmH --------- Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>	1 个月前
.gitleaks.toml	feat: add hermetic tortured-phrase screening (#699) Refs #660. Measurement and issue closure follow in the separately preregistered post-main mechanical conformance PR.	2 个月前
.gitleaksignore	test(evals): promote the adjudicated 2026-08-04/05 #610 v0.2 baseline cohort (#643) * test(evals): promote the adjudicated 2026-08-04/05 #610 v0.2 baseline cohort (#610) Nine panels at the frozen post-#638 SHA 112a869 under the #608 harness (claude-opus-5 @ xhigh, isolated per-seat two-phase claude -p --bare dispatch): seven score-eligible records — smoke ms00 r99 + the 2026-08-04 2 x 3 fleet's MS00 r1/r2, MS01 r1, MS02 r2, plus the authorized 2026-08-05 supplementary replicates MS01 r3 and MS02 r3 under the two-replicate rule — and two blocked panels (MS01 r2, MS02 r1) recorded, not replaced. All nine provenance-valid; byte-for-byte promotion verified per stem with check_e4_promotion.py. This cohort is the adjudicated v0.2 baseline: MS01 11/11 and MS02 9/9 strict recall on both replicates, clean-control false findings 0 on all three clean panels, severity agreement 0.672. The #637 decoration abort family did not recur (0 SYNTHESIS-PARSE found-0 events vs 6/6 in the attempt fleet), confirming #638; conformance aborts 0.50 -> 0.33 with the residual a new single-seat content-grammar family. README run-history gains the baseline row and retires the pending corrected-cohort row. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016oZptTA3LpqeXWSxynzSz9 * test(evals): add #610 recompute_adjudication + row metrics after dual-track round 1 (#610) Round-1 findings applied: - codex P1 (valid): both MS01 records gain the mandatory recompute_adjudication block — SD-01 grim, SD-02 n_from_df, SD-03 p_from_test_statistic, SD-11 grimmer, all VERIFIED in both replicates via prose receipts (baseline provision), maintainer-confirmed; arithmetic independently recomputed. - codex P2 (valid): README baseline row now reports receipt-backed recompute recall 4/4, class-wide statistical recall 5/5 (incl. MS02 SD-07 reporting_only), and the 0-numeric/0-narrative clean-control split. - security P3 (valid): severity_source gains a seeded-defect scope qualifier on all four defect records to stop lexical collision with the synthesizer's [SEVERITY-SOURCE: letter-fallback] markers on secondary/corroborating cells. - codex P2 #3 (rejected): nested adjudication shape without notes matches the merged 2026-08-03 precedent; notes is a no-response disclosure field and no such event occurred. - codex P2 #4 (rejected): check_e4_promotion.py is an operator-time gate by design (#616); same statement shape as the merged #639 entry. Work-dir copies updated in lockstep; check_e4_promotion.py re-PASS on the four touched stems (byte-identical). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016oZptTA3LpqeXWSxynzSz9 * test(evals): align receipt_fields_present key with the #610 spec name (#610) Security-track round-2 P2: derived_value -> derived_value_or_range in both MS01 recompute_adjudication blocks (spec docs/design/2026-08-02-610-...-spec.md line 77); the post fleet will emit under the spec name and the frozen baseline must compare key-for-key. Work-dir copies updated in lockstep; check_e4_promotion.py re-PASS byte-identical on both stems. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016oZptTA3LpqeXWSxynzSz9 * test(evals): re-adjudicate three recompute cells per spec §5 procedure rules (#610) Codex round-2 P1s (both valid, verified against the spec's own worked examples): SD-03 receipts in both MS01 replicates show only the two-tailed p ≈ .19 where §5.1 requires both tails shown when the paper states none (the silent two-tailed default §7 bans); the r3 SD-01 receipt argues from the integer product alone without the §5.2 rounding-interval reachability check (r1 carries the full check with the 3.839/3.851 neighbours and stays VERIFIED). Full-bundle sweep confirmed no other seat supplies the missing elements in either replicate. Maintainer-confirmed re-adjudication: the three cells become MISCOMPUTED (complete-looking attempt, unsupported assumption), receipt-backed recompute recall corrects to 3/4 (r1) and 2/4 (r3), README and CHANGELOG updated to the honest baseline numbers. per_defect content-level detection and all other row metrics unchanged. Work-dir in lockstep; check_e4_promotion.py re-PASS on both stems. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016oZptTA3LpqeXWSxynzSz9 * docs(design): pin MISCOMPUTED-over-CLAIM_ONLY precedence for dual-trigger receipts (#610) Security-track round-3 P3: the three downgraded baseline cells set a receipt field absent AND record an unsupported assumption; the verdict table stated no precedence between CLAIM_ONLY (absence trigger) and MISCOMPUTED (procedural trigger). One sentence pins MISCOMPUTED as the governing class before these records freeze as the classification precedent the post fleet is compared against. No record change; both verdicts carry numerator 0 so no published number moves. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016oZptTA3LpqeXWSxynzSz9 * chore(ci): fingerprint-ignore a gitleaks false positive in frozen E4 evidence (#610) Model venue-fit prose in the promoted r1 field_analysis.md ('(SoLAR, open access, Q1-equivalent in field)') trips generic-api-key entropy. The bundle is byte-frozen promotion evidence, so per the .gitleaks.toml maintenance note this is a one-off .gitleaksignore fingerprint, not a rule edit. Local re-scan: no leaks found. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016oZptTA3LpqeXWSxynzSz9 * chore(ci): reword .gitleaksignore comment that re-triggered the rule (#610) The previous comment quoted the flagged prose verbatim, so the ignore file itself matched generic-api-key on the PR merge ref. Paraphrased; the fingerprint line is unchanged. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016oZptTA3LpqeXWSxynzSz9 * chore(ci): promote frozen-E4-bundle prose to a gitleaks path allowlist class (#610) Two-layer correction of the earlier fingerprint approach, which cannot survive this repo's squash-merge flow (the squash SHA that introduces the promoted bundle to main is unknowable in advance, so a per-commit fingerprint would turn main red on merge): - .gitleaks.toml gains the frozen E4 raw-bundle .md prose as the second documented false-positive class (path entry per its own maintenance note); machine logs and JSON records in the bundles stay scanned. - .gitleaksignore shrinks to the single historical occurrence where commit 08ddaca's version of the ignore file itself quoted the flagged prose. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016oZptTA3LpqeXWSxynzSz9 * chore(ci): scope the E4-prose gitleaks exemption to the misfiring rule (#610) Security-track ack round: the E4 raw-bundle .md path entry moves from the all-rules allowlist into a [[allowlists]] block with targetRules=[generic-api-key], so private-key / PAT / token detection stays live over the frozen prose (8.30.1 forbids mixing legacy [allowlist] with [[allowlists]], so the citation-keys class converts to an untargeted [[allowlists]] block, semantics unchanged). Mutation-tested with the pinned binary: a synthetic ghp_ token planted in the exempted path still fires github-pat while generic-api-key is suppressed there; a control copy outside the path fires both. .gitleaksignore gains the post-squash removal note. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016oZptTA3LpqeXWSxynzSz9 --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>	2 个月前
CHANGELOG.md	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
CITATION.cff	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
CONTRIBUTING.md	docs(readme): keep the three most recent releases in the six READMEs; history moves to CHANGELOG.md and frozen locale archives (#870) * docs(readme): keep only the three most recent releases in the six READMEs; history moves to CHANGELOG.md and frozen locale archives README.md carried 60 release summaries (76 KB, 68% of the file), and the five translated READMEs mirrored the same list. Every release added one more paragraph to six files, and check_spec_consistency.py pinned 127 old headings across them, so the section could only grow. - Each README changelog section now keeps the three most recent releases and links to CHANGELOG.md. The 60 English one-paragraph summaries (58 of which CHANGELOG.md did not have) are merged verbatim as the leading blockquote under each CHANGELOG.md entry; the 11 releases that existed only in README.md (v1.0 .. v2.6.2, v3.0, v3.1) are backfilled as new CHANGELOG.md sections. - The five translated summary lists are frozen verbatim under docs/changelog-archive/<locale>.md (root-relative links rewritten), with a bilingual note that later releases live in the English CHANGELOG.md. - check_spec_consistency.py: the per-locale heading pins are replaced by one README_CHANGELOG_KEEP constant plus check_readme_changelog_section, which fails on any extra `### v` heading and on a missing CHANGELOG.md / archive link; the archives join the relative-link check. Fixtures and two new regression tests (fourth heading, missing archive link). - stage_capability_matrix.json: the three README.md claim anchors that pointed into the changelog now point at CHANGELOG.md (matrix re-rendered). - The Windows / real-Python prerequisite paragraph moves from README.md (and the ko-KR / es-ES translations) to docs/SETUP.md "Python (optional)" with a zh-TW counterpart; the READMEs keep a one-line pointer. - CONTRIBUTING.md release checklist records the prepend-one / drop-one rule. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CZxeK5emwXRAAqdNDfmc78 * docs(changelog): record the README changelog slimming under [Unreleased] (#870) Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CZxeK5emwXRAAqdNDfmc78 * docs(readme): retarget the three README consumers codex found, harden the changelog-section lint Codex round 1 on #870 (gpt-6-astra, xhigh) found three consumers that only ever matched README text through the release history, all confirmed locally: - shared/contracts/degradation_registry.json: two write-scope-guard authorities anchored on the Windows / real-Python paragraph now point at docs/SETUP.md, where that paragraph lives. - scripts/venue_disclosure_contract_harness.py: the five READMEs leave SURFACE_FILES; they never carried a maintained venue inventory and matched only via the v3.12.1 / v3.20 summaries. - scripts/test_reading_probe_lint.py: the ARS_SOCRATIC_READING_PROBE cross-repo check reads docs/SETUP.md and docs/SETUP.zh-TW.md (the flag table) instead of the two READMEs. Two advisory findings applied: check_readme_changelog_section strips fenced code and HTML comments before locating the section, resolves links through extract_link_targets, and rejects a kept release heading that appears more than once (two new regression tests). The five archives gain a localized H2 between the note and the copied list. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CZxeK5emwXRAAqdNDfmc78 * docs(changelog): full dates on the four backfilled early releases, from git history (#870) v1.0 and v2.0 (2026-02-26), v2.0.1 (2026-03-01) and v2.2 (2026-03-05, the README had said 2025) take the dates of the commits that introduced their README entries; the command-invariants lint requires a full YYYY-MM-DD date on every release heading. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CZxeK5emwXRAAqdNDfmc78 --------- Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>	1 个月前
GOVERNANCE.md	docs: solo-maintainer governance statement + SECURITY triage procedure (#760) (#778) * docs: solo-maintainer governance statement + SECURITY triage procedure (#760) GOVERNANCE.md: decision authority, honest cross-model scope (error-detection control, not organizational independence), release authority, EOL posture, and the operating-principles section (three distilled principles with informative ISO/IEC 42001 anchors, the #753-#760 coverage map, Annex C not-applicable assessments). SECURITY.md: 7-day promise becomes acknowledgement-only hard promise plus a written severity-tiered best-effort triage procedure. NOTICE.md: governance pointer. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AFJbYZyrJmSPFMJKQZhVHC * docs: apply dual-agent review to #760 (order-safe claims, honest ceilings, sustainable promises) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AFJbYZyrJmSPFMJKQZhVHC * docs: close codex R1 on #760 (per-operator cross-model attribution, fairness signals disclosed, EOL-scoped promises) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AFJbYZyrJmSPFMJKQZhVHC * docs: narrow fairness scoring absolute to persons/real-world allocation (codex R2) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AFJbYZyrJmSPFMJKQZhVHC --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>	2 个月前
LICENSE	fix: update LICENSE to official CC BY-NC 4.0 legalcode with personal copyright - Remove HEEACT from copyright holder, keep personal name only - Replace summary with official Creative Commons legalcode text - Enables GitHub to properly recognize the license type Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>	7 个月前
MODE_REGISTRY.md	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
NOTICE.md	docs: solo-maintainer governance statement + SECURITY triage procedure (#760) (#778) * docs: solo-maintainer governance statement + SECURITY triage procedure (#760) GOVERNANCE.md: decision authority, honest cross-model scope (error-detection control, not organizational independence), release authority, EOL posture, and the operating-principles section (three distilled principles with informative ISO/IEC 42001 anchors, the #753-#760 coverage map, Annex C not-applicable assessments). SECURITY.md: 7-day promise becomes acknowledgement-only hard promise plus a written severity-tiered best-effort triage procedure. NOTICE.md: governance pointer. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AFJbYZyrJmSPFMJKQZhVHC * docs: apply dual-agent review to #760 (order-safe claims, honest ceilings, sustainable promises) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AFJbYZyrJmSPFMJKQZhVHC * docs: close codex R1 on #760 (per-operator cross-model attribution, fairness signals disclosed, EOL-scoped promises) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AFJbYZyrJmSPFMJKQZhVHC * docs: narrow fairness scoring absolute to persons/real-world allocation (codex R2) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AFJbYZyrJmSPFMJKQZhVHC --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>	2 个月前
POSITIONING.md	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
QUICKSTART.md	feat(sr-screener): add protocol-driven study screening as the fifth skill (#919) (#920) * feat(sr-screener): add protocol-driven study screening as the fifth skill Adds sr-screener v1.0.0: turns a proposal or PROSPERO protocol into user-confirmed eligibility rules, then screens titles/abstracts and full texts with two blinded reviewer subagents and a third-reviewer adjudicator. Eight modes; standard-library scripts for parsing, de-duplication, merging and PRISMA/RIS/literature_corpus outputs. Registers the skill on every inventory surface and lint registry (skills/ symlink, CLAUDE.md table + routing rule 8, manifests, MODE_REGISTRY, mode spectrum, announce routing, model-tiering manifest, Bucket C rosters, agent mirror, data_access pin, instruction/data boundary list), updates docs/READMEs, and adds scripts/test_sr_screener_pipeline.py. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * chore(sr-screener): keep only the skill folder; drop repo-wide registration changes Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Revert "chore(sr-screener): keep only the skill folder" The maintainer asked for the skill to be registered on the inventory, model-tiering and instruction/data-boundary surfaces (#919). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(sr-screener): address #919 review (Sonnet defaults, required QC, pilot gate, routing) - Reviewers A/B (and every screening role) default to Sonnet; no profile offers a smaller model; the cost check prints any per-role override. - The QC recheck of records both reviewers excluded is required: a reproducible sample (min 20, default 100) is drawn once screening is complete and stays pending until the senior reviewer decides it. - The full title/abstract run needs a pilot compared with the team's own labels (merge_decisions.py --pilot-labels) that missed no record the team advanced; --pilot-override records a reason for the methods text. - Routing: a Screening boundary paragraph in the routing core and its six copies (screening starts only when the user asks; no automatic handover from systematic-review mode); fixtures 13-17 incl. three negative cases. - Registration restored (inventory, model tiering, boundary block). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(sr-screener): close model, pilot and full-text QC review gaps * docs(changelog): point the sr-screener entry at renumbered fixtures 17-21 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(sr-screener): bind review revisions and close screening report gaps Freeze the first pilot's scope, require audited scope expansion, reject runtime job replacements and validate the actual dispatched IDs. Bind generated labels and pilot comparisons to review, dataset, protocol and config identities. Amendments archive prior decisions/QC state and begin a new revision. Reject foreign or obsolete results; explicitly verified legacy imports require an audited reason. Preserve prior results and decisions in the local audit trail. Derive completeness from every prepared record/report and validate full-text retrieval coverage against the current TA decision snapshot. Escape spreadsheet text, compare complete normalized titles without conflicting identifiers, reject ambiguous/prefix PDF matches, refresh empty RIS categories and diagram reports awaiting classification. Replace personal-looking examples with marked synthetic ones and advance the letter whose abstract does not establish whether it has data. Validation: 77 synthetic screening regression tests; 12 relevant spec/security/ inventory/version checks pass with repository LF content. Broader companion tests yield 476 passes and the same 71 Windows failures on both this change and the unchanged bcf394b baseline. No live model calls or screening accuracy claims. * fix(sr-screener): preserve stage decisions across reporting edits Hash the effective screening inputs for each stage, excluding reporting and preparation metadata and models/codes used only by the other stage. Preserve existing result identities when upgrading an unchanged whole-config hash; refuse to clear a legacy review whose original config cannot be verified. Quote every field in CSV exports, synchronize the screening agent mirror, and accept either preserved or normalized carriage returns in the XLSX test. Add regressions for CSV quoting, completed decisions and pilot preservation, stage-specific amendments, and safe legacy identity upgrades. Validate both openpyxl XML backends, the agent mirror, and repository consistency checks. * fix(sr-screener): neutralize formulas after alternative CSV separators Escape formula prefixes after embedded semicolons and tabs in spreadsheet exports, while preserving original bibliographic records. Keep QUOTE_ALL. Exercise all four formula prefixes with both alternative delimiters across duplicates, not-retrieved, audit, pending and screening-log exports. Verify the exported values with comma readers and semicolon/tab readers. Validation: 105 screening tests and the real-repository agent mirror test pass; all 12 repository consistency checks pass. --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Imbad0202 <132531341+Imbad0202@users.noreply.github.com>	1 个月前
README.es-ES.md	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
README.ja-JP.md	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
README.ko-KR.md	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
README.md	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
README.zh-CN.md	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
README.zh-TW.md	docs(release): prepare v3.24.0 (#1029) Promote [Unreleased] to 3.24.0 (2026-10-27), add the #1027/#1028 entry, and bump the version surfaces, README summaries, lint pins, and the SKILL.md content lock. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	8 天前
SECURITY.md	docs: solo-maintainer governance statement + SECURITY triage procedure (#760) (#778) * docs: solo-maintainer governance statement + SECURITY triage procedure (#760) GOVERNANCE.md: decision authority, honest cross-model scope (error-detection control, not organizational independence), release authority, EOL posture, and the operating-principles section (three distilled principles with informative ISO/IEC 42001 anchors, the #753-#760 coverage map, Annex C not-applicable assessments). SECURITY.md: 7-day promise becomes acknowledgement-only hard promise plus a written severity-tiered best-effort triage procedure. NOTICE.md: governance pointer. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AFJbYZyrJmSPFMJKQZhVHC * docs: apply dual-agent review to #760 (order-safe claims, honest ceilings, sustainable promises) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AFJbYZyrJmSPFMJKQZhVHC * docs: close codex R1 on #760 (per-operator cross-model attribution, fairness signals disclosed, EOL-scoped promises) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AFJbYZyrJmSPFMJKQZhVHC * docs: narrow fairness scoring absolute to persons/real-world allocation (codex R2) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AFJbYZyrJmSPFMJKQZhVHC --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>	2 个月前
THIRD_PARTY.md	docs(third-party): list md-to-word as an out-of-tree DOCX layout companion (#969) Adds a row to the Listed projects table for md-to-word, a deterministic Markdown to Word (.docx) typesetting engine that pairs with ARS Stage 5 output. Per THIRD_PARTY.md eligibility: the project credits ARS and links back from its README, and the description states only what the tool does. Refs #967 Co-authored-by: 刘 <1750422899@qq.com>	22 天前
package.json	feat(platform): add community-maintained Pi wrapper (#636) * feat(platform): add pi wrapper * feat(pi): scope ARS activation and add evidence * fix(pi): gate automatic ARS skill invocation * fix(pi): preserve args and tree activation * docs(pi): scope prompt guarantees to idle runs * fix(pi): canonicalize hidden skill paths Pi preserves lexical resource paths while Node resolves extension imports through symlinks. Compare real filesystem identity after XML decoding. * fix(pi): substitute command arguments once Avoid reprocessing placeholder-like text introduced from user arguments. * fix(pi): bound inactive skill block matching Prevent alternate skill formatting from extending a match into an adjacent unrelated skill block.	2 个月前
pyproject.toml	refactor(tests): move shared helpers into tests package (#311)	4 个月前
requirements-dev.txt	feat(sr-screener): add protocol-driven study screening as the fifth skill (#919) (#920) * feat(sr-screener): add protocol-driven study screening as the fifth skill Adds sr-screener v1.0.0: turns a proposal or PROSPERO protocol into user-confirmed eligibility rules, then screens titles/abstracts and full texts with two blinded reviewer subagents and a third-reviewer adjudicator. Eight modes; standard-library scripts for parsing, de-duplication, merging and PRISMA/RIS/literature_corpus outputs. Registers the skill on every inventory surface and lint registry (skills/ symlink, CLAUDE.md table + routing rule 8, manifests, MODE_REGISTRY, mode spectrum, announce routing, model-tiering manifest, Bucket C rosters, agent mirror, data_access pin, instruction/data boundary list), updates docs/READMEs, and adds scripts/test_sr_screener_pipeline.py. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * chore(sr-screener): keep only the skill folder; drop repo-wide registration changes Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Revert "chore(sr-screener): keep only the skill folder" The maintainer asked for the skill to be registered on the inventory, model-tiering and instruction/data-boundary surfaces (#919). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(sr-screener): address #919 review (Sonnet defaults, required QC, pilot gate, routing) - Reviewers A/B (and every screening role) default to Sonnet; no profile offers a smaller model; the cost check prints any per-role override. - The QC recheck of records both reviewers excluded is required: a reproducible sample (min 20, default 100) is drawn once screening is complete and stays pending until the senior reviewer decides it. - The full title/abstract run needs a pilot compared with the team's own labels (merge_decisions.py --pilot-labels) that missed no record the team advanced; --pilot-override records a reason for the methods text. - Routing: a Screening boundary paragraph in the routing core and its six copies (screening starts only when the user asks; no automatic handover from systematic-review mode); fixtures 13-17 incl. three negative cases. - Registration restored (inventory, model tiering, boundary block). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(sr-screener): close model, pilot and full-text QC review gaps * docs(changelog): point the sr-screener entry at renumbered fixtures 17-21 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(sr-screener): bind review revisions and close screening report gaps Freeze the first pilot's scope, require audited scope expansion, reject runtime job replacements and validate the actual dispatched IDs. Bind generated labels and pilot comparisons to review, dataset, protocol and config identities. Amendments archive prior decisions/QC state and begin a new revision. Reject foreign or obsolete results; explicitly verified legacy imports require an audited reason. Preserve prior results and decisions in the local audit trail. Derive completeness from every prepared record/report and validate full-text retrieval coverage against the current TA decision snapshot. Escape spreadsheet text, compare complete normalized titles without conflicting identifiers, reject ambiguous/prefix PDF matches, refresh empty RIS categories and diagram reports awaiting classification. Replace personal-looking examples with marked synthetic ones and advance the letter whose abstract does not establish whether it has data. Validation: 77 synthetic screening regression tests; 12 relevant spec/security/ inventory/version checks pass with repository LF content. Broader companion tests yield 476 passes and the same 71 Windows failures on both this change and the unchanged bcf394b baseline. No live model calls or screening accuracy claims. * fix(sr-screener): preserve stage decisions across reporting edits Hash the effective screening inputs for each stage, excluding reporting and preparation metadata and models/codes used only by the other stage. Preserve existing result identities when upgrading an unchanged whole-config hash; refuse to clear a legacy review whose original config cannot be verified. Quote every field in CSV exports, synchronize the screening agent mirror, and accept either preserved or normalized carriage returns in the XLSX test. Add regressions for CSV quoting, completed decisions and pilot preservation, stage-specific amendments, and safe legacy identity upgrades. Validate both openpyxl XML backends, the agent mirror, and repository consistency checks. * fix(sr-screener): neutralize formulas after alternative CSV separators Escape formula prefixes after embedded semicolons and tabs in spreadsheet exports, while preserving original bibliographic records. Keep QUOTE_ALL. Exercise all four formula prefixes with both alternative delimiters across duplicates, not-retrieved, audit, pending and screening-log exports. Verify the exported values with comma readers and semicolon/tab readers. Validation: 105 screening tests and the real-repository agent mirror test pass; all 12 repository consistency checks pass. --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Imbad0202 <132531341+Imbad0202@users.noreply.github.com>	1 个月前
requirements-pdf-content-classifier.txt	feat(pdf): sandbox optional content classification (#724) * feat(pdf): sandbox optional content classification * fix(pdf): bind advisory to complete input * fix(pdf): close classifier protocol edge cases * fix(pdf): reject overflowing confidence values * fix(pdf): keep worker adapter outcomes closed * fix(pdf): close worker and publication races	2 个月前

Academic Research Skills for Claude Code

English | 繁體中文版 | 日本語版 | 한국어 | Español

一套完整的学术研究 Claude Code 技能包，涵盖从研究到论文出版的全流程。

30 秒安装（Claude Code CLI / VS Code / JetBrains，v3.7.0+）：

/plugin marketplace add Imbad0202/academic-research-skills
/plugin install academic-research-skills

安装后运行 /ars-plan，ARS 会用苏格拉底式对话帮你规划章节结构。需要前置条件或传统 symlink 安装，请看快速安装。

AI 是你的副驾驶，不是机长。 它可以起草文字，full mode 甚至会起草整篇论文；但决定权在你，pipeline 每个阶段都会停下来等你确认。它处理繁琐工作（搜文献、排格式、验数据、查逻辑一致性），让你专注在真正需要思考的事上：定义问题、选择方法、解读数据意义、决定「我认为」后面要接什么。作者是你，提交的每一个主张都由你负责。

和 humanizer 不同，这个工具不是帮你隐藏使用 AI 协作的事实，而是帮你把关文章质量。风格校准会从你过去的文章中学习你的声音，写作质量检查会识别让文字读起来像机器生成的模式。目标是质量，不是掩饰。

为什么选「人机协作」而不是「全自动」？

Lu 等人（2026，Nature 651:914-919）发表的 The AI Scientist 是第一个端到端全自动的 AI 研究系统，其生成的论文通过 ICLR 2025 workshop 的盲审（评分 6.33/10，workshop 平均 4.87）。他们自己的 Limitations 段落也列出了这类系统会遇到的结构性失败模式：实现错误、幻觉实验结果、取巧特征依赖、实现错误被包装成「意外发现」、方法论伪造、框架锁定、引用幻觉。

ARS 建立在这个前提上：人类研究者 + AI 的组合，比纯自动或纯人工更能避开这些失败模式。Stage 2.5 与 Stage 4.5 学术诚信闸门运行 7 类阻断式检查清单（见 academic-pipeline/references/ai_research_failure_modes.md），reviewer 也提供 opt-in 的 calibration mode 用用户提供的 gold set 测量 FNR/FPR。

Zhao 等人（2026-05）盘点了 arXiv、bioRxiv、SSRN、PMC 上 250 万篇论文中的 1.11 亿条引用，保守估计 2025 年单年就有 146,932 条幻觉引用，并观察到 2024 年中是上升的拐点；bioRxiv-to-PMC 这条配对的「预印本进入正式发表版本」幻觉存活率达 85.3%。他们把「真实引用被用来支撑被引文献其实没有提出的主张」描述为当前未解的问题。ARS v3.7.1 为来源 provenance 加上 trust-chain frontmatter，v3.7.3 为未来的 claim-level 审计铺设 locator 基础设施（三层引用 anchor），并在引用阶段呈现 advisory 风险信号（ARS 内部把这条 claim-faithfulness 缺口标记为「L3」，此为 ARS 的用词，不是论文的用词）。v3.7.x 的设计动机来自 Zhao 等人的 corpus-scale 发现；ARS 本身的 corpus-scale 评估仍是未来工作。

v3.8 补上 L3 缺口的另一半。v3.7.3 让每一条引用都带 locator anchor，v3.8 在这个基础上加一道 opt-in 审计（ARS_CLAIM_AUDIT=1）：获取每个 anchor 指向的原始文本，判断论文里的 claim 是否真有被该引用支撑。五类新的 HIGH-WARN annotation（claim-not-supported、negative-constraint-violation、fabricated-reference、anchorless、constraint-violation-uncited）会在 formatter terminal hard gate 直接阻止输出。Calibration runner 随 release 附一组 25 条的合成 gold set，采用 FNR<0.15、FPR<0.10 双阈值。随附的测试使用直接返回标准答案的替身裁判，所以它验证的是工具本身，不是真正的 AI 裁判；目前还没有真实裁判的 calibration 结果，正式放大投入要等这份证据（v3.8 spec §5）。

v3.3 的灵感来自 PaperOrchestra（Song, Song, Pfister & Yoon, 2026, Google）：Semantic Scholar API 验证、反泄露协议、VLM 图表验证、修订轨迹追踪。ARS 当前以分类式、证据锚定的准则轨迹实现最后一项，不计算分数差。

架构与 pipeline

👉 docs/ARCHITECTURE.md — 完整 pipeline 视图：流程图、阶段 × 维度矩阵、数据访问流、skill 依赖图、质量闸门、模式清单。

这份架构文档取代了原本散在 README 各处的 pipeline 描述。关于「哪个阶段跑什么」的所有信息都集中在一个地方。

快速安装

前置条件

Claude Code（建议最新版；plugin packaging 需要近期版本）
已导出 ANTHROPIC_API_KEY，或在第一次运行 claude 时设置
选用： Pandoc 用于 DOCX 输出，tectonic + 思源宋体 TC 用于 APA 7.0 PDF（纯 Markdown 输出不需要这两者）

Plugin 安装（v3.7.0+，推荐）：

/plugin marketplace add Imbad0202/academic-research-skills
/plugin install academic-research-skills

验证可用： 运行 /ars-plan 并描述你正在写的论文，ARS 会用苏格拉底式对话帮你规划章节结构。如果想做单次测试，可以运行 /ars-lit-review "你的主题"。

👉 docs/SETUP.md — 完整指南：安装 Claude Code、设置 API key、选用的 Pandoc/tectonic（DOCX/PDF）、跨模型验证（ARS_CROSS_MODEL），以及六种安装方式（Plugin、项目 skills、全局 skills、claude.ai Project、repo clone、Claude Science 导入）。

你的安装渠道实际启用哪些控制机制？ 可用性因安装渠道而异，请查逐渠道对照表：docs/CONTROL_AVAILABILITY.md（英文）。

👉 docs/DATA_FLOWS.md — 哪些数据会离开你的电脑（书目 resolver、可选且需明确同意的跨模型调用、更新检查）、本地缓存存什么与存多久、每条路径怎么关闭。（英文）

使用 Claude Science？ 五个 skill 可直接导入：Skills → Import from GitHub，粘贴 https://github.com/Imbad0202/academic-research-skills，点 Preview，再点 Import（需本 repo v3.14.0+——导入器读取 marketplace manifest 中显式声明的 skill 路径）。导入是一次性快照：ARS 更新后需重新导入。导入的 skill 承载 ARS 方法论（研究／写作／评审协议）；Claude Code 专属机制——slash commands、hooks、subagent 编排——不会转移。详见 docs/SETUP.md Method 5。

使用 Pi？ 运行 pi install git:github.com/Imbad0202/academic-research-skills 安装仓库内的社区维护 wrapper。它继续以原始 ARS 内容为准，并记录 Pi 在编排和 hooks 方面的限制。详见 pi/README.md。

用 Codex CLI？ 请安装姐妹版：Imbad0202/academic-research-skills-codex。同一套 workflow 内容，Codex 原生打包为单一 $academic-research-suite skill，提供 ars-* 别名。

性能与费用

👉 docs/PERFORMANCE.md — 各模式 token 预算、完整 pipeline 估算（一篇 15k 字论文，按 2026-09 牌价约 US$3–7，未计 cache 折扣），以及建议的 Claude Code 设置（Auto 模式；Agent Team 选用）。

使用指南与文章

学术写作不该是一个人的事：一套开源 AI 协作工具如何改变研究者的工作流 — 完整使用指南（繁体中文）
Academic Writing Shouldn't Be a Solo Act — Full pipeline walkthrough (English)

功能特色一览

Deep Research — 13 个 Agent 的研究团队，支持苏格拉底引导、PRISMA 系统性回顾、意图检测、对话健康度监控、可选跨模型 DA、Semantic Scholar API 验证。
Academic Paper — 12 个 Agent 的论文撰写团队，含风格校准、写作质量检查、LaTeX 输出强化、可视化、修订教练、引用格式转换、反泄露协议、VLM 图表验证。
Academic Paper Reviewer — 7 个 Agent 的多视角同行评审，采用逐准则、证据锚定的叙事判断（Journal-Fit Reviewer + 3 位动态审查者 + 魔鬼代言人），含让步门槛协议、攻击强度保持、可选跨模型 DA critique / calibration、R&R 追溯矩阵、只读约束。目前 live review 一律为 NOT_CALIBRATED；完整 calibration 只产生有界候选 profile，尚未接入 live review。
Academic Pipeline — 10 阶段全流程调度器，含自适应 checkpoint、主张验证、材料护照、可选 repro_lock、可选跨模型学术诚信验证、中途强化机制，以及逐项准则的叙事退步检查（typed trajectory 尚未实现）。
SR-Screener — 系统综述、范围综述与快速综述的研究筛选，按用户确认的筛选规则执行：两位盲审 AI 审查者加第三位裁决者、按固定顺序的排除代码、不预设任何决定、可续跑的批次、质量检查（种子研究、近似排除复查、kappa 与 PABAK）、PRISMA 2020 数字、EndNote/Zotero RIS 分组，以及交给 academic-paper 的 literature_corpus[]。AI 的决定只是辅助，最终由研究团队核查。
数据访问层级标注（v3.3.2+）— 每个 skill 声明 data_access_level（raw / redacted / verified_only），由 scripts/check_data_access_level.py 强制执行。设计灵感来自 Anthropic 的 automated-w2s-researcher（2026）。详见 shared/ground_truth_isolation_pattern.md。
任务类型标注（v3.3.2+）— 每个 skill 声明 task_type（open-ended 或 outcome-gradable）。目前 ARS 所有 skills 皆为 open-ended。
Benchmark 报告 Schema（v3.3.5+）— JSON Schema + lint script，要求诚实的 benchmark 比较报告。详见 shared/benchmark_report_pattern.md。
Artifact 可复现性 Lockfile（v3.3.5+）— Material Passport 添加可选 repro_lock 子区块。是配置文档化，不是重播保证 — LLM 输出不是逐字节可复现。详见 shared/artifact_reproducibility_pattern.md。
实验来源凭证登录（#260）— Material Passport 可选的 experiment_provenance[] 记录研究者在外部跑过的实验（ARS 从不执行实验），论文主张通过 claim_intent_manifest.planned_experiment_ids[] 与之 join。诚信 gate（Stage 2.5/4.5）逐条比对实验支撑型主张与登录凭证 — ALIGNED / OVERSTATED / NOT_SUPPORTED_BY_PROVENANCE / PROVENANCE_INSUFFICIENT — 但不判定实验本身是否正确。fail-closed 的 experiment_intake_declaration 让「有没有跑实验」成为 Stage 1 明确决定。详见 shared/handoff_schemas.md。

**诚信与验证边界：**ARS 检查的是论文与所报告的研究过程，包括引用是否存在、主张与来源是否一致、所述方法、已登记实验结果与论文主张的一致性、图表忠实度，以及报告／流程／提交包的一致性；部分检查采用抽样或由 LLM 判断。ARS 不能证明程序确实执行、原始数据真实，或结果可复现；如果捏造内容被前后一致地报告，仍可能通过这些检查。详见 POSITIONING.md〈Integrity checks and the empirical-work boundary〉。

实际产出展示

查看一次完整 pipeline 运行的实际产出，包含同行评审报告、学术诚信验证报告、完稿论文：

这是 2026 年 3 月的记录，不代表现在的表现。 这次运行（2026-03-07 至 03-08）使用的是 academic-pipeline v2.3，当时 ARS 还没有在 v3.3 加入 Semantic Scholar 核对，也还没有在 v3.11 加入确定性的四索引引用闸门。这里的数字描述的是那个版本，现行闸门还没有在这篇论文上测量过。作者栏写 Claude（Anthropic），是因为研究者在这次实验中这样要求。这篇论文不是 Anthropic 的出版物；ARS 的定位是工具不取代研究者，也不主张作者身份（见 POSITIONING.md）。

浏览所有 pipeline 产出 →

产出物	说明
完稿论文（英文）	APA 7.0 格式，LaTeX 编译
完稿论文（中文）	中文版，APA 7.0
学术诚信报告 — 审稿前	Stage 2.5：标出 15 条有问题的引用（8 条书目错误、6–8 条疑似虚构）+ 3 个统计错误
学术诚信报告 — 最终	Stage 4.5：确认零回归
同行评审第一轮	Journal-Fit Reviewer + 3 审查者 + 魔鬼代言人
再审	修订后验证审查
同行评审第二轮	跟踪审查
回复审查意见	逐点回复
出版后审计报告	另行用 Claude Code + WebSearch 审计全部引用：经过 3 轮学术诚信审查后，最后 68 条引用中仍有 21 条有问题

搭配工具：Experiment Agent

如果你的研究需要在写作前做实验（代码或人工研究），Experiment Agent 技能填补 ARS Stage 1（研究）和 Stage 2（写作）之间的空缺。

ARS Stage 1 研究      →  RQ Brief + Methodology Blueprint
        ↓
  experiment-agent     →  运行/管理实验 → 验证结果
        ↓
ARS Stage 2 写作      →  用验证过的实验结果撰写论文

功能：执行代码实验（Python、R 等）并实时监控、管理人工研究 protocol 与 IRB 伦理审查、11 种统计谬误检测、可复现性验证。

搭配使用方式：ARS pipeline 完成 Stage 1 后暂停，在另一个 experiment-agent session 中执行实验，完成后将结果（含 Material Passport）带回 ARS Stage 2。ARS 不需要任何修改。详见 experiment-agent README。

使用方式

快速开始

# 启动完整研究 pipeline
你: "我想做一篇关于 AI 对高等教育质量保障影响的研究论文"

# 苏格拉底引导模式
你: "引导我研究 AI 在教育评估中的应用"

# 引导式论文撰写
你: "引导我写一篇关于少子化影响的论文"

# 审查现有论文
你: "帮我审查这篇论文"（接着提供论文）

# 查看 pipeline 进度
你: "进度" 或 "status"

个别 Skill 使用

Deep Research（深度研究，8 种模式）

"研究 AI 对高等教育的影响"                    → full mode（完整研究）
"给我一份 X 的快速摘要"                       → quick mode（快速简报）
"帮我做 X 的系统性文献回顾，含 PRISMA"        → systematic-review mode
"引导我研究 X"                                → socratic mode（苏格拉底引导）
"帮我核查这些说法"                            → fact-check mode（事实核查）
"帮我做文献回顾"                              → lit-review mode（文献回顾）
"审查这篇论文的研究质量"                      → review mode（论文审查）

Academic Paper（学术论文撰写，11 种模式）

"帮我写一篇论文"                              → full mode（完整撰写）
"引导我写论文"                                → plan mode（引导规划）
"先帮我搭论文大纲"                            → outline-only mode（只做大纲）
"我有初稿，这是审稿意见"                      → revision mode（修订）
"帮我整理这些审稿意见成修订路线图"            → revision-coach mode
"帮我写这篇的摘要"                            → abstract-only mode（摘要）
"把这批数据写成文献回顾论文"                  → lit-review mode（文献回顾论文）
"转换成 LaTeX" / "引用格式转 IEEE"            → format-convert mode（格式转换）
"检查引用格式"                                → citation-check mode（引用检查）
"帮我生成 NeurIPS 的 AI 使用声明"             → disclosure mode（AI 使用声明）

Academic Paper Reviewer（论文审查，6 种模式）

"审查这篇论文"                                → full mode（Journal-Fit Reviewer + R1/R2/R3 + 魔鬼代言人）
"快速评估这篇论文"                            → quick mode（快速评估）
"引导我改进这篇论文"                          → guided mode（引导改进）
"检查研究方法"                                → methodology-focus mode（方法论聚焦）
"验收修订"                                    → re-review mode（再审验收）
"用我的 gold set 校准 reviewer"               → calibration mode（校准）

Academic Pipeline（全流程调度器）

"我想做一篇完整的研究论文"                    → 从 Stage 1 开始完整 pipeline
"我已经有论文，帮我审查"                      → 从 Stage 2.5 进入（先做学术诚信审查）
"我收到审稿意见了"                            → 从 Stage 4 进入

Pipeline 结束时自动产出 Stage 6：过程记录 — 含论文创建过程记录与 6 维度协作质量评估（1–100 分）。

SR-Screener（研究筛选，8 种模式）

"把我的计划书转成筛选规则"                    → protocol mode
"这篇摘要符合我的综述纳入条件吗？"            → quick mode（单一审查者初筛）
"先用种子研究试跑筛选"                        → pilot mode
"筛选这些数据库导出文件"                      → ta-screen mode
"筛选晋级记录的全文"                          → ft-screen mode
"裁决我 Rayyan 导出文件中的冲突"              → adjudicate mode
"帮我复查排除的记录"                          → audit mode
"给我这次筛选的 PRISMA 数字"                  → report mode

支持语言

繁体中文 — 用户以中文对话时默认使用
English — 用户以英文对话时默认使用
学术论文自动产出双语摘要（中文 + English）

使用其他语言？ 苏格拉底模式（deep-research）和 Plan 模式（academic-paper）采用意图匹配启动 — 检测你的请求含义，而非比对特定关键字。这代表它们支持任何语言，无需额外设置。

不过，一般的 Trigger Keywords 区块（决定 skill 是否被启动）仍以英文和繁体中文为主。如果你发现 skill 在你的语言下触发不稳定，可以在各 SKILL.md 的 ### Trigger Keywords 区块中加入你的语言的关键字，提高匹配信心。

支持引用格式

APA 7.0（默认，含中文引用规则）
Chicago（Notes & Author-Date）
MLA
IEEE
Vancouver

支持论文结构

IMRaD（实证研究）
主题式文献回顾
理论分析
个案研究
政策简报
研讨会论文

Skill 详细信息

各 agent 的职责与各阶段产出物现已移至 docs/ARCHITECTURE.md。版本号保留在此以维持 release metadata 集中管理。

Deep Research (v2.12.1)

13 个 Agent 的研究团队。模式：full、quick、review、lit-review、three-way-scan、fact-check、socratic、systematic-review。完整 agent 名单与产出物：见 ARCHITECTURE.md §3。

Academic Paper (v3.3.1)

12 个 Agent 的论文撰写 pipeline。模式：full、plan、outline-only、revision、revision-coach、abstract-only、lit-review、format-convert、citation-check、disclosure、rebuttal-audit。输出：MD + DOCX（Pandoc 可用时）+ LaTeX（APA 7.0 apa7 class / IEEE / Chicago）→ tectonic 编译 PDF。完整 agent 名单与各 phase 职责：见 ARCHITECTURE.md §3。

Academic Paper Reviewer (v1.11.1)

7 个 Agent 的多视角审查，采用 逐准则、证据锚定的叙事判断。模式：full、re-review、quick、methodology-focus、guided、calibration。目前 live review 与 Schema 6 package 一律为 NOT_CALIBRATED；完整 calibration 可产生有界候选 profile，但尚未接入 live review。不得以固定总分映射接受、小修、大修或退稿。第一轮审查面板 vs. 契约治理再审调度的分界：见 ARCHITECTURE.md §3 Stage 3 / Stage 3'。

Academic Pipeline (v3.24.0)

10 阶段调度器，含学术诚信验证、两阶段审查、苏格拉底指导、协作质量评估。Pipeline 规则（由 agent 按流程遵守，不是运行时保证）：每个阶段都需用户确认 checkpoint；学术诚信验证（Stage 2.5 + 4.5）为 MANDATORY 且没有不留记录的绕过路径（所有覆写都须记录用户理由、供 Stage 6 使用）；R&R 追溯矩阵（Schema 11）把每一项审查意见对应到作者的修订主张，并记录复审是否验证通过。v3.4 添加 Compliance Agent（PRISMA-trAIce + RAISE）于 Stage 2.5 / 4.5。v3.5 添加 协作深度观察员（collaboration_depth_agent，仅咨询性质、永不阻挡流程）于每一次 FULL/SLIM checkpoint 与 pipeline 完成时。MANDATORY 学术诚信闸门（2.5 / 4.5）明确跳过观察员，避免稀释合规检查。理论基础：Wang & Zhang (2026), IJETHE 23:11。逐阶段矩阵（agent、产出物、闸门）：见 ARCHITECTURE.md §3。

SR-Screener (v1.0.0)

4 个 agent 的研究筛选，位于 deep-research（问题、计划书、检索）与 academic-paper（撰写综述）之间。模式：protocol、quick、pilot、ta-screen、ft-screen、adjudicate、audit、report。两位盲审的审查者 subagent（只能用 Read 与 Grep）按用户确认的筛选规则审每一条记录，第三位审查者裁决"晋级 vs. 排除"的冲突；只用标准库的 Python 脚本负责解析 RIS / PubMed .nbib / Web of Science / CSV 导出文件、去重、分批、合并，并产出筛选记录表、RIS 分组、PRISMA 2020 数字、含 [TO COMPLETE] 字段的方法段草稿，以及 literature_corpus[] 文件。Agent 遵守的规则（不是运行时保证）：用户确认筛选规则前不筛任何记录、失败的调用不会被当成默认"排除"、数字发表前由研究团队核查决定。详见 sr-screener/SKILL.md。

v3.0 优化：我们发现了 AI 的哪些结构性限制

在使用 ARS 撰写一篇关于 AI 与高等教育的反思文章时，我们遇到了三个结构性问题：

框架锁定：AI 在给定框架内越来越精致，但无法质疑框架本身
谄媚倾向：每次挑战魔鬼代言人的攻击，它都让步得太快
意图检测错误：苏格拉底模式在用户仍在探索时就急着收敛

改了什么

魔鬼代言人让步门槛：反驳必须评分 1-5，≥4 才允许让步。不允许连续让步。框架锁定检测。
苏格拉底意图检测：检测用户是「探索型」还是「目标型」。探索型模式停用自动收敛。
对话健康度指针：每 5 轮后台自检，检测持续同意、回避冲突、过早收敛。
跨模型验证：设置 ARS_CROSS_MODEL 激活第二 AI 模型独立审查。详见 docs/SETUP.md。
AI 自我反思报告：Pipeline 结束后自动产出 AI 行为自评。

这些优化不能完全解决 AI 的结构性限制——它们让限制变得可见、可追踪、可被人类介入。

授权条款

本作品采用 CC-BY-NC 4.0 授权。

你可以自由：

分享 — 复制并分发本作品
改编 — 重混、转换、以本作品为基础进行创作

但须遵守以下条件：

署名 — 你必须给予适当署名
非商业性 — 你不得将本作品用于商业目的

署名格式：

Based on Academic Research Skills by Cheng-I Wu
https://github.com/Imbad0202/academic-research-skills

贡献者

吴政宜 (Cheng-I Wu) — 作者与维护者

aspi6246 — 贡献者。v3.1 优化灵感来自 Claude-Code-Skills-for-Academics：只读约束模式、Anti-Pattern 作为一等公民设计、认知框架方法（教「如何思考」而非只有步骤）、精简 skill 体量理念。

mchesbro1 — 贡献者。最初提出并撰写了 IS Basket of 8 期刊清单（Issue #5）。

cloudenochcsis — 贡献者。将 IS 章节从 Basket of 8 扩充为完整的 Senior Scholars' Basket of 11，补上 Decision Support Systems、Information & Management、Information and Organization（Issue #7、PR #8）。数据源：AIS Senior Scholars' List of Premier Journals。

eltociear（Ikko Eltociear Ashimine）— 贡献者。翻译了日文版 README（README.ja-JP.md）（PR #161）。

xpfo-go（xpfo）— 贡献者。翻译了简体中文版 README（README.zh-CN.md）（PR #181）。

ktao732084-arch — 贡献者。扩展了 academic-paper 披露系统，新增九个医学出版政策目标、按目标收集必需事实的流程，以及 fail-closed 的独立披露渲染（Issue #596、PR #599）；扩充 EQUATOR 临床报告参考，加入 CARE、STARD 和 TRIPOD+AI 的精简指引，以及 fail-closed 的研究设计路由流程（Issue #594、PR #601）；并设计及贡献独立的中文文献解析器、API 协议和合成传输 fixture 测试套件（Issue #595、PR #600）。

更新纪录

这里只列最近三个版本。完整更新记录在英文版 CHANGELOG.md。到 v3.21.2 为止的简体中文版本摘要已冻结存放于 docs/changelog-archive/zh-CN.md，之后不再更新。

v3.24.0（2026-10-27）— 审查小组推荐文献与审查意见的参考检查、write-scope guard 启动器改为单一进程，以及诚信检查规则修复

**两项审查参考检查与数项修复；没有新 skill，这两项检查在真实审查上的准确度尚未测量：**v3.24.0 在 Editorial Decision Letter 加上两份附件。第一份实时查询 Reviewer 2 与 Reviewer 3 推荐的每一篇著作，查询范围是引用验证关卡使用的四个索引，只报告有没有这条记录，不判断该著作是否支持推荐理由（#1001）。第二份把每一项 weakness，以及 Devil's Advocate 每一行 CRITICAL 或 MAJOR 意见，去掉审稿人席位标签后交给 session 模型的一个 subagent，由它判断稿件是否与该意见矛盾。只有判定者的引文通过脚本对稿件的比对检查，附件才列为矛盾；其他判定为错误的意见列为未确认（#1002）。这项检查默认开启，设置 ARS_PANEL_CLAIM_CHECK=0 可以关闭。两份附件都不影响审查决定、修订路线图或复审标准。第二项检查上线前，一项预注册的 held-out 审计检查了 805 条审查小组意见，估计错误意见的比例介于 1.5% 到 12.9%；Claude Opus 5.5 与 Claude Fable 5.1 担任判定者时都通过了它的对照关卡。write-scope guard 的启动器每次工具调用只启动一个 Python 进程，原来是三个，Windows 上 AppXSvc 的内存泄漏预计因此降为原来的三分之一左右，但 Windows 上的实际行为尚未测试（#1015）；guard 的代码注释也写明它检查哪些工具名称（#1022）。三条诚信检查规则在 Stage 5 入口、报告有效期限与 Stage 4′ 三点上取得一致（#1011），启动通知也改为只在有新版本时显示（#1027）。#1015、#1022 与 #1027 由外部用户报告。

v3.23.2（2026-10-24）— 诚信检查只读引用的版本、Stage 2.5 一律执行，以及适配器、启动通知与 Windows 修复

**修复；没有新 skill。一项提示层变更做了小规模行为检查，其余尚未测量：**v3.23.2 让 Stage 2.5 与 4.5 的声明核查只拿稿件引用的那个版本来比对（#1006）。读不到引用的版本时，核查会记录一条访问备注并请用户提供 PDF，不再改读工作论文或预印本。在一份合成测试稿上运行六次（Claude Opus 5.5 与 Claude Fable 5.1 各三次，关闭网络工具），六次都遵守这条规则。先前的诚信报告不再能让 pipeline 跳过 Stage 2.5，改为附上作为参考（#995）。缺少必要输入时，每一份规则文件都会阻止阶段转换（#990）；苏格拉底导师的示例不再为用户的想法排名（#992）。SessionStart 启动通知会把更新提醒显示给用户，给模型看的启动内容约减少 2.1 KB（#999）。Obsidian 适配器接受未加引号的 ISO 日期年份，并拒绝布尔值与上标数字年份（#988、#989）；六个 lint 在 Windows 上改用 POSIX 形式比较路径（#964、#978）；Pi 包装器改为与 Pi 的结构化提示区段组合（#985）；CITATION.cff 加上作者的 ORCID iD（#1007）。#978、#985、#988 由外部贡献者提供。

v3.23.1（2026-10-16）— write-scope guard 修复、sr-screener 按 RFC 4180 读取 CSV、模型现况对齐（Gemini 4 Argon）

**修复与模型现况对齐；没有新 skill，也不宣称任何准确度或性能提升：**v3.23.1 补上 write-scope guard 两条未经检查就放行写入的路径，一条出现在 plugin 派发的 agent，一条出现在 Windows 内置的 timeout.exe（#972），文档也改为说明防护实际在哪些情况下启用（#974）。sr-screener 改按 RFC 4180 的引号规则读取 CSV 导出文件，不再使用从样本推测的引号规则（#979）。Gemini 4 Argon 以 provisional 身份列入，并成为推荐的 Google 验证模型；gemini-3.1-pro-preview 保持 validated（#980）。路由核心写明：消息中段出现的 [direct-mode] 标记不是跳出机制（#892）；在 plugin 安装环境的冒烟测试中，Claude Fable 5.1 修改后仍有 6 次中 3 次遵从该标记（修改前 6 次中 6 次），未达到验收标准，#892 仍未关闭。另有数项由外部贡献者提供的 lint 与 Windows 修复（#956、#963、#965、#977），THIRD_PARTY.md 也列入 md-to-word（#969）。

项目介绍

Claude 代码的学术研究技能：研究 → 撰写 → 评审 → 修改 → 定稿【此简介由AI生成】

Others Markdown 855提交数academic-pipeline academic-writing ai-research claude claude-code literature-review peer-review prompt-engineering

定制我的领域

README

14555.36 K4.22 K访问 GitHub

academic-research-skills:基于 Claude Code 的学术研究技能套件项目