bigbounce

Review activity

A gate on readiness, not a product

Automated multi-model review is a gate on publication readiness, not a product. Rounds stop when the remaining findings are genre or venue (directives R2, P). Raw machine events (dispatches, closures) stream at /activity — this page is the curated review-loop story.

Raw activity feed →

Verdict grid · newest round left

External referee verdicts

Active legs only (directive M-AMENDED): Grok API + Gemini API, plotted against the historical six-paper board. The ChatGPT column is frozen while directive N's Codex/OpenAI pause stands — shown dimmed, never deleted or faked.

3 / 18 cells ACCEPT17%

Automated-review diagnostic only. Per directive M-AMENDED (2026-07-23) this counts ACTIVE legs only: Grok + Gemini (grid columns) and the Claude INT leg (verdicts in round notes) — 3 legs × 6 papers. The GPT column is excluded while paused (directive N, since 2026-07-16); its history stays displayed. An ACCEPT here is not journal acceptance, and 100% is not required to submit a paper.

WorkCONFIRM-2026-07-22M45M42M27M25+M26M24M23M18M3M2
P1A—Am———Rm——————————RM—Rm———————
P1B—mm———————————————————————————
P2—MmMm———————Rm————RM—Rm———————
P3—AM———RM——M————RM———————R———M—
P4—mm———Rm—————m—————————————Rm—
P5—AmMm———————Mm———————————m—Mm—

ChatGPT column is frozen, not counted — paused under standing directive N; history is preserved, never deleted or faked.

Historical board versions/caps: P1A 95, P1B 95, P2 95, P3 95, P4 95, P5 95. The live-lineup works (A3, P4′, P1N) are not yet columns in this historical grid — their round-by-round evidence is in the timeline below and their readiness is on /status.

Publication status

What's left before publication

What's left before publication

0/6 signed off

0 waiting on Houston · 6 on the agents

evidence as of Sep 7, 2026

P1A95%agentv1A.0.127 has not been read by an automated review board yet — one confirm read, no new science.v1A.0.127 · board Jul 23, 2026 ↗
P1B95%agentv2B.0.23 has not been read by an automated review board yet — one confirm read, no new science.v2B.0.23 · board Jul 23, 2026 ↗
P295%agentv1.7.130 has not been read by an automated review board yet — one confirm read, no new science.v1.7.130 · board Jul 23, 2026 ↗
A3M75%agentv3M.0.24 has not been read by an automated review board yet — one confirm read, no new science.v3M.0.24 · board — ↗
P395%agentv3.2.0-r17 has not been read by an automated review board yet — one confirm read, no new science.v3.2.0-r17 · board Jul 23, 2026 ↗
4P95%agentv4P.0.7 has not been read by an automated review board yet — one confirm read, no new science.v4P.0.7 · board — ↗

How to read this. Publication readiness is science closure + evidence & reproducibility + automated review convergence + packaging, and then Houston's own final read — the last 5%. A paper marked Houston needs no further math, compute, GPU/CPU runs or new data; a paper marked agent has one named item still owned by the loop. The trailing stamp shows which exact PDF the newest automated review board actually read — ✓ means the current one, ↗ means an earlier one.

Publishing is a separate phase. arXiv endorsement, venue choice, submission clicks and journal / independent human review come after 100% and never subtract from readiness. See the publishing checklist →

Gap and skills

The review machinery, self-improving

Substantive findings only the external tier caught, and the pattern/prompt-rule catalog those findings get mined into.

03570target 0⚑EXT22 — integrity gate — loop de-biased, skills hardened⚑EXTDB — de-bias caught real self-favoring framing⚑RCEXT — 3-round grind: 0 new external findings⚑M1 — post-overhaul: 1 genuinely-new, all caught + closed⚑M2 — P4 e2e engaged-but-reflag · P3 release objection dissolved⚑M3 — first INT-API ACCEPT (Grok, P5) · P3 hinge dissolved⚑M34 — P2 streak 12→13 (cap 74) · P5 streak 1→2 + ChatGPT REJECT→MAJOR tier-lift (cap 68→74) · P3APJS mislabel caughtEXT1: 60 — 60 externally-VERIFIED findings survived six clean internal rounds (EXT1 truth-audit baseline)60EXT1 · 06-10EXT2: 32 — Genuinely-new substantive findings per EXT2 truth-audit GAP METRIC sections; P4/P5 net-new incl. PARTIAL/OPINION is 10 each (looser total 47)32EXT2 · 06-10EXT3: 27 — EXT3 truth-audits: ~27 genuinely-new, all wording/asset/policy class — zero substantive physics blockers27EXT3 · 06-11EXT4: 13 — EXT4 truth-audits: 13 genuinely-new (−52% vs EXT3), zero physics on any paper — captions, cross-refs, repo hygiene, one estimand-family item, one QC-provenance item; all closed same-day13EXT4 · 06-11EXT5: 19 — EXT5 truth-audits: ~19 verified — but ~5 are regressions/persistence failures from our own closure waves (P1A NJL + caption, P3 changelog-vs-body ×2, P5 table arithmetic); externally-sourced novel content keeps shrinking (P2: one stale sentence) — closure-agent quality became the bottleneck and got new mandatory verification rules19EXT5 · 06-12EXT6: 18 — EXT6 truth-audits: ~18 verified — TWO real self-closure regressions caught externally (P1A §IV E synthesis paragraph still said "too large" while §IV A body said "4×10⁻⁶⁹ ρ_Λ" — three prior waves missed it; P2 §V L604 arithmetic 3.5σ→3.22σ pattern-051 from R34conf OAI-E10); P1B 2 BLOCKERs (CHANGELOG + bbn_predictor YAML) closed; P4 0 scientific findings; Gemini-for-P3 dropped after 6/6 hallucinated §-numbers — Milestone external state: Gemini's first FULL ACCEPT (P1B) + Grok 4× consecutive ACCEPT18EXT6 · 06-12EXT10: 5 — EXT10: 18/18 MINOR. Full truth-audit pending (/peer-review-truth-audit). Preliminary count: ~5 likely-verified findings (P1A dimensional bookkeeping + sphaleron rate; P3 top-1% wording + Cramer's V arithmetic; P4 Shamir biblio chimera; P5 V-Web→T-Web rename). Many submission-day items expected STALE. P5 at 0 substantive external-only findings.5EXT10 · 06-13EXT11: 15 — EXT11 truth-audit: 15 VERIFIED + 4 PARTIAL across 22 total findings. Key new: P1A Eq.15 algebraic inversion (new closure regression); P5 stale V-Web figure art (figure regeneration required, text rename was done but plot titles not); P3 abstract 'catalog-grade' logic contradiction; P2 abstract r=0.75 vs r=0.84 inconsistency. P4 down to 1 VERIFIED (Shamir title text only). Internal→External gap closing: P4 now 0 substantive externals-only.15EXT11 · 06-13EXT12: 10 — EXT12 truth-audit: ~10 remaining text-only fixes across 5 papers (P4 = 0 substantive findings — confirmed 3/3 ACCEPT). P1A: 2 local wording (Sec IV/App B dimension sentence + reheating residual). P1B: 1 release-pairing harmonization across 3 locations. P2: 1 BF self-check paragraph (3 sentences). P3: 2 precision fixes (DESI validation gate type + Table IX Savage-Dickey label). P5: 4 items (3 residual V-Web tokens + Fig 8 spacing + 'Verdict.' label + DOI). Gemini NO VERDICT (synthesis-mode) — not counted as new findings; EXT11 baselines held. New pattern-057: systematic-rename-grep-body-text.10EXT12 · 06-13EXT13: 6 — EXT13-closure-wave: 5-paper text-only closure wave (P4 frozen). Remaining external-only findings closed: P1A dim-bookkeeping + reheating residual; P1B release-pairing harmonization; P2 BF self-check rewrite; P3 abstract DESI gate type + Table IX BF note; P5 pattern-057 body V-Web residuals (4 sites). P4 = 0, FROZEN at 3/3 ACCEPT.6EXT13 · 06-13EXT14: 8 — EXT14 truth-audit: 12/18 ACCEPT. ~8 verified findings. P1B+P4 at 0 (frozen ACCEPT). P1A: 3 wording items (chirality-flipping, parity-odd amplitude, local-operator-promotion framing). P2: 1 BF Eq.9 vs Eq.10 mapping. P3: 1 Table IX Savage-Dickey footnote. P5: 3 items (2 math-mode subscripts Sec IX B + Eq display). Pattern-059 encoded: math-mode subscripts require separate sweep.8EXT14 · 06-13EXT16: 4 — EXT16 truth-audit: 14/18 ACCEPT. 4 ChatGPT MINOR items remained. P1A: Sec XII.A C/P-violating thermal-scattering propagation miss. P2: CDF-tail direction corrected (raises not reduces for narrow delta-prior). P3: Table IX prior density footnote. P5: math-mode V\mbox{-}Web + nomenclature note direction + dup T-Web. P1B+P4 frozen ACCEPT — 0 findings.4EXT16 · 06-13EXT17: 0 — EXT17: 18/18 ACCEPT — zero substantive external-only findings remaining. All 4 EXT16 ChatGPT MINORs closed. 2 false positives truth-audited (version mismatch + pattern-052 fresh-reviewer). Gap reaches zero: internal tier now matches external tier quality.0EXT17 · 06-13EXT7: 14 — EXT7 truth-audits: ~14 verified — TWO real findings (P1A Fig 3 caption/code mismatch H0=67.7 claim vs H0=69.2 actual; P1B NaMaster Eq (1) sigma_b^2 divisor missing from script) + 12 polish closures. P5 CLEAN at acceptance stage (ChatGPT VoidFinder is 6th k=20 re-raise, auto-falsified; Gemini 3 MAJORs all falsified on disk). Externals running out of substantive content — closure-to-finding ratio now ~1:1.14EXT7 · 06-13EXT8: 8 — EXT8 closure-wave: honest MNRAS/PRD calibration prompt introduced — ChatGPT MAJOR→MINOR on P1B/P2/P4/P5. ~8 verified findings, mostly submission-day actions (Zenodo DOI, companion placeholders) and minor wording.8EXT8 · 06-13EXT9: 6 — EXT9 closure-wave: ChatGPT cleared P1A Fig 3 caption + P3 structural issues. ~6 verified findings remaining post-EXT9-closure. P4/P5/P1B at 0 verified external findings.6EXT9 · 06-13EXT18: 7 — EXT18 true 5-reviewer round (Claude = Claude Code sub-agent): P1B real arithmetic — Ωa relic-density subsection added post-freeze: ρ_crit,0 8.1e-11→3.7e-11 eV⁴, relic denominator 2H₀²→6H₀², H₀-marginalization ≤1%→≤3%, S8 2.5σ→2.6σ — closed v1B.0.73. P2: 3 internal-consistency fixes — closed v1.7.69. P1A/P3/P4/P5 CLEAN on truth-audit.7EXT18 · 06-14EXT19: 3 — EXT19 4-vendor confirmation (no Anthropic API key; Claude is a sub-agent now): P2 CLEAN — Fisher-invariance ESSENTIAL was a category error (sensitivity recast, not independent Fisher). P1B: 3 ALP-subsection items (anharmonic coeff O(θ²/6)→O(θ²/12), frozen-branch z_osc≤0 note, Table IV header mislabel) — closed v1B.0.74.3EXT19 · 06-14EXT20: 0 — EXT20: 6/6 ACCEPT — fresh-referee external round. 0 new substantive external-only findings. 2 trivial cosmetic micro-fixes (P2 + P5) closed in-session. Second consecutive zero-gap external round.0EXT20 · 06-18EXT22: 2 — EXT22 confirm round: 2 new-verified polish items — NV-P1A-1 (MINOR: §XII.B body-alignment; closed) + NV-P4-1 (POLISH: +3.3σ→+3.29σ; closed). All ~34 other findings already-covered/extraction-artifact/opinion/stale. Polish-tier convergence reached: 3-pass total (R52+EXT21+EXT22) → 0 MAJOR/BLOCKER. ★ integrity gate — loop de-biased, skills hardened2EXT22 · 06-26EXTDB: 2 — De-biased external-review validation: with severity-steering struck from the referee prompt, 2 GENUINE self-favoring items surfaced that the biased prompt was burying — P1A '13 logically-independent barriers'→'mechanism-class' (several share the scaling ansatz) + P3 'catalog-grade' tier was summing Gaia+eROSITA which FAILED injection-recovery (relabeled, validated ≥268,519). Both fixed. A broader real-fix wave (P1B inflated w0wa σ-distances removed; P5 L_parity operator reformulated to be SO(3)-invariant; P1A H0-artifact disclosed) closed previously-latent items. ★ de-bias caught real self-favoring framing2EXTDB · 06-28RAEXT: 0 — Round A (1 of 3) EXT: 0 genuinely-new external findings — the Round-A INT pass (12 real items closed) caught everything first. Verdicts lifted to MINOR-tier dominant; P1A drew a real Gemini ACCEPT.0RAEXT · 06-29RBEXT: 0 — Round B (2 of 3) EXT: 0 genuinely-new external findings beyond the Round-B INT closes (4 items, incl a Lesson-F self-favoring fix on P4). P4 swept all-MINOR.0RBEXT · 06-29RCEXT: 0 — Round C (3 of 3, FINAL) EXT: 0 genuinely-new external findings — neutral gate-discipline truth-audit of the harsh P1A+P3 3/3-MAJORs confirmed every one is a disclosed caveat, a structural submission feature (companion derivations, Zenodo DOI deferred), framing taste, or reviewer noise. 3-round grind: 23 real items closed across INT, 0 new surviving external findings. ★ 3-round grind: 0 new external findings0RCEXT · 06-30M1: 1 — M1 wave — first full EXT measurement after the directive-M presentation overhauls (shorter abstract + de-dup) on all 5 papers. 1 genuinely-new reader-visible finding total: P5's overhaul-introduced abstract 'pre-declared' vs §V.B 'post-hoc/exploratory primary' contradiction (both Grok+ChatGPT MAJOR#1; git-proven overhaul-reaction not oscillation) → CLOSED v0.1.125. All other findings source-cited re-flags; 0 broken refs from any overhaul (incl P3 revtex→AASTeX conversion). P4 v1.0.238 closes DP4-15 (8.47M image-level e2e injection, artifact-verified). ★ post-overhaul: 1 genuinely-new, all caught + closed1M1 · 07-12M2: 1 — M2 wave — targeted re-reads (P5 v0.1.125 fix · P4 first reads WITH the 8.47M e2e live · P3-ApJS first read with the immutable release live). 1 genuinely-new reader-visible finding total: P5's Eq.(4) prose 'the SEVEN … terms' vs the multline's EIGHT displayed/listed terms (arithmetic-label mismatch, NOT in the M1 ledger) → CLOSED-BY-EDIT one-word seven→eight in v0.1.126 (no number changed). P4: ChatGPT M2b REJECT engaged the new e2e section but re-frames the disclosed image-level injection = DP4-15 RE-FLAG; mask count 3,200,420+740=3,201,160 reconciled in tex L950 → 0 genuinely-new (streak 6→7). P3-ApJS: immutable-release objection DISSOLVED (Grok reads 'the released 22.5 M catalog … raw native scores reside on an exited pod' = pod-blocked residual DP3-15, not the missing-archive bar) → 0 genuinely-new (streak →1); ChatGPT FAILED-dead = M2c gap. ★ P4 e2e engaged-but-reflag · P3 release objection dissolved1M2 · 07-12M3: 0 — M3/M2c confirm wave — re-tests on the two just-touched papers. P5 M3 (v0.1.126): EXT Grok MINOR + INT-API OpenAI REJECT / Grok ACCEPT / Gemini MINOR / Claude MINOR — Grok's is the CAMPAIGN'S FIRST INT-API ACCEPT (verified raw body + milestone log; 4 non-blocking MINORs, central-claim endorsement); ChatGPT EXT M3 FAILED-dead (M3b gap). 0 genuinely-new → P5 streak 0→1. P3-ApJS M2c (v3.1.158-apjs): recovered ChatGPT EXT REJECT completes the M2 wave — immutable-release hinge DISSOLVED (remaining basis = DP3-15 pod-blocked re-inference OPEN-COMPUTE + DP3-16 catalog-vs-PRD venue, both Houston-gated); 0 genuinely-new → P3 streak 0→1. Caps HOLD: P5 80 · P3 56. ★ first INT-API ACCEPT (Grok, P5) · P3 hinge dissolved0M3 · 07-12M34: 0 — M34-EXT confirm wave — headed-browser Grok + ChatGPT legs on P2 + P5 (produced by the background sweep), raw verbatim + screenshot READ before every recorded verdict; 0 fabricated. P2 Grok = MINOR REVISIONS (1 MAJOR-tag + 4 MINOR) + P2 ChatGPT = REJECT (10 MAJOR + 1 MINOR) — every finding source-cited to standing D-ids (proxy ρ=−0.868 conservative-endpoint/channel-native surrogate 2.3σ HIGHER c15 → DP2-04/-07/-26/-34/-35, null-space Eq.(A4) → DP2-15/-16/-01, cubic transmission → DP2-13/-32.6, r=0.84 → DP2-14/-17/-34, Fisher → DP2-22, Bayes prior-volume → DP2-18, κ_ε → DP2-20, gauge-146 → DP2-21, compression → DP2-30); both reviewers CONCEDE −35/16 is supported (ChatGPT: 'supported by the canonical contraction formulas'); REJECT rests on disclosed survival-through-bounce + venue = structural harsh-referee floor → 0 genuinely-new → P2 clean-wave streak 12→13, cap 74 HOLDS. P5 Grok = MINOR REVISIONS (0 MAJOR + 4 MINOR) + P5 ChatGPT = MAJOR REVISIONS (9 MAJOR + 3 MINOR) — an honest REJECT→MAJOR tier-lift on byte-unchanged content; all re-flags → DP5-13/-24 (post-hoc), DP5-12/-22 (RSD), DP5-06/-19 (footprint), DP5-11 (envelope), DP5-10 (binomial OPEN-COMPUTE), DP5-08/-09 (de-attenuation), DP5-14 (T-Web), DP5-20 (App B), DP5-21 (Paper-IV venue), DP5-16/-02/-03 (minors); both reviewers affirm the qualitative null → 0 genuinely-new → P5 clean-wave streak 1→2 (re-crosses directive-K bar), cap 68→74 (latest ChatGPT REJECT→MAJOR). INTEGRITY CATCH: the sweep's 'P3APJS_chatgpt_M34' file is a MISLABELED/duplicate P5 review (internal sentinel ext_P5_M34; DESIVAST/VoidFinder content, not the P3 anomaly engine) — NOT recorded as any P3 verdict; P3's ApJS EXT re-test stays outstanding. No content bump (v1.7.116 / v0.1.127 stand). Caps below the 96 all-ACCEPT gate throughout. ★ P2 streak 12→13 (cap 74) · P5 streak 1→2 + ChatGPT REJECT→MAJOR tier-lift (cap 68→74) · P3APJS mislabel caught0M34 · 07-13
P1A 18→0 ▾P1B 11→0 ▾P2 4→0 ▾P3 10→0 ▾P4 5→0 ▾P5 12→0 ▾
0255075retro: 34 patterns · 14 prompt rules — 2026-06-02 retro baseline: 34 codified patterns34retro: 14 reviewer-prompt rulesretro (06-02)R23conf-mine: 44 patterns · 14 prompt rules — R23conf pattern-mine: catalog at 44 (incl. draft patterns 040-044)44R23conf-mine: 14 reviewer-prompt rulesR23conf (06-09)EXT1-gapmine: 48 patterns · 19 prompt rules — EXT1 gap-mine: patterns 045-048 + artifact_crosscheck.py + reviewer-prompt rules 15-1948EXT1-gapmine: 19 reviewer-prompt rulesEXT1 (06-10)EXT2-gapmine: 49 patterns · 19 prompt rules — EXT2 gap-mine: pattern-051 closure-introduced regression (5-point closure-wave protocol)49EXT2-gapmine: 19 reviewer-prompt rulesEXT2 (06-10)EXT3-gapmine: 50 patterns · 19 prompt rules — EXT3 gap-mine: pattern-052 re-raise vindication test + browser-loop completion/version gates; prompt rules unchanged50EXT3-gapmine: 19 reviewer-prompt rulesEXT3 (06-11)EXT11-gapmine: 53 patterns · 21 prompt rules — EXT11 gap-mine: 3 new auto-rules added — pattern-053 closure-arithmetic-regression-audit (Eq.15 inversion), pattern-054 figure-art-rename-verify (V-Web→T-Web in plot titles not caught), pattern-055 internal-audit-label-leak-strip ((B1)/(E*) labels in journal prose). Prompt rules +2 (figure-art-rename gate + closure-label grep).53EXT11-gapmine: 21 reviewer-prompt rulesEXT11 (06-13)EXT12-gapmine: 57 patterns · 23 prompt rules — EXT12 gap-mine: pattern-056 pdftotext-artifact-class auto-falsify (italic NS→MS rendering artifact — already in SKILL-PDFTOTEXT entry); pattern-057 systematic-rename-grep-body-text (after V-Web→T-Web rename, 3 residual tokens survived in §VIII/§IX/App C body text — figure-art gate insufficient); pattern-058 gemini-fresh-chat-verdict-format (Gemini 6/6 synthesis-mode at EXT12 — explicit ACCEPT/MINOR/MAJOR format instruction must be FIRST LINE of message). Prompt rules +2 (Gemini verdict-format gate + body-text rename grep gate).57EXT12-gapmine: 23 reviewer-prompt rulesEXT12 (06-13)R52-learning-loop: 64 patterns · 23 prompt rules — R52 learning-loop: 4 new patterns drafted (061-064). 061: dispatch-tag-vs-intext-mismatch — orchestrator brief conflicts reviewer in-text Recommendation; read the Recommendation line, not the wrapper tag. 062: stale-pdf-false-positive — served PDF lags source by 1-2 versions; pre-dispatch gate must confirm md5 match. 063: extraction-artifact-false-positive — reviewer text-layer OCR mangles math glyphs; always verify math findings against .tex source + cross-vendor full-PDF corroboration. 064: grok-harsh-outlier-false-positive — Grok REJECT/MAJOR truth-audits false-positive in 4/4 R52 papers; truth-audit each Grok reason individually, check primary/secondary inversion, disclosure-as-defect misread. Candidate not drafted: missing-released-artifact (print-only generator) — 1 finding (P2 only), below ≥3/≥2 threshold.64R52-learning-loop: 23 reviewer-prompt rulesR52 (06-26)integrity-audit-2026-06-26: 64 patterns · 24 prompt rules — Integrity-audit hardening 2026-06-26: standing integrity-audit pre-check added as mandatory first step of every R-round truth-audit (re-derive every REJECT/MAJOR dismissal independently before logging convergence); PDF-hygiene md5 pre-dispatch gate hardened into cross-vendor-r-round SKILL.md (pattern-062). EXT-prompt de-bias deferred to a separate round. Prompt-rules: 23 → 24 (integrity-audit mandate = rule 24). Pattern count unchanged at 064.64integrity-audit-2026-06-26: 24 reviewer-prompt rulesintegrity-audit (06-26)RA · de-bias + manifest-gate: 65 patterns · 26 prompt rules — Round A skill upgrades: (1) the deferred EXT-prompt DE-BIAS executed — severity-steering struck from the external referee prompt; the de-biased prompt then caught 2 genuine self-favoring items (P1A 'logically-independent'→'mechanism-class', P3 'catalog-grade' summing FAILED surveys) the biased prompt buried = reviewer-prompt rule 25. (2) pattern-067 ext-worker-manifest-inflation drafted (patterns 64→65) + its VERDICT-line anti-inflation gate = rule 26 — after a Round-A sweep-worker manifest over-counted ACCEPTs ('acceptable after revisions' ≠ ACCEPT) and was caught + corrected against the referee text.65RA · de-bias + manifest-gate: 26 reviewer-prompt rulesRA (06-29)RB/RC · referee-variance: 66 patterns · 26 prompt rules — Round B/C skill upgrade: pattern-066 llm-referee-run-to-run-variance drafted (patterns 65→66) — the SAME papers swung MINOR-dominant (Round B EXT) → MAJOR-dominant (Round C EXT) while getting slightly better; codifies that a single sweep's verdict tally is noisy, findings must recur across ≥2 sweeps or INT+EXT before closing, and convergence = '0 genuinely-new real findings on truth-audit', not one all-ACCEPT sweep. Validated by the RCEXT truth-audit (0 new real findings under the harsh 3/3-MAJOR sweep).66RB/RC · referee-variance: 26 reviewer-prompt rulesRB/RC (06-30)site-sync · staleness-gate: 67 patterns · 27 prompt rules — Site-integrity skill upgrade (Houston caught the /reviews + /papers pages showing June-26 data after 3 rounds): pattern-065 static-site-data-staleness drafted (patterns 66→67) + the static-data same-commit gate = reviewer-prompt rule 27. Root cause: the site reads BOTH the live DB AND static build-time files (papers.ts / reviewTimeline.ts / live-status.ts / hardcoded page prose) — updating the live DB alone leaves the public-facing surfaces stale. Every round now updates ALL static surfaces + verifies-after-deploy in the same commit. Folded into /bigbounce-site-sync.67site-sync · staleness-gate: 27 reviewer-prompt rulessite-sync (06-30)INT-M2 · rebuttal-hardening: 68 patterns · 28 prompt rules — INT-M2 round skill upgrade: pattern-068 preemptive-rebuttal-hardening drafted (patterns 67→68) — all 6 paper-owner agents independently converged on it. At convergence reviewers stop finding NEW defects but keep re-flagging the SAME disclosed caveats; the technique is to ADD an explicit in-paper rebuttal for any finding that recurs ≥2 rounds as STALE/FALSIFIED, so the next pass can't re-raise it = reviewer-prompt rule 28. This is how a converged review keeps producing real improvement every round (7 closures + 6 papers hardened this round) rather than flatlining. Source-grounded only; for null results, hardening makes the null MORE conservative.68INT-M2 · rebuttal-hardening: 28 reviewer-prompt rulesINT-M2 (06-30)RS5 · signpost + cross-vendor + de-biased-calibration: 71 patterns · 29 prompt rules — EXT RS5 skill upgrade (3 new patterns 069-071, count 68→71): 069 signpost-resolved-concerns — a fresh de-biased sweep re-flagged ~48 of ~52 MAJORs that were ALREADY addressed; the fix is explicit 'Response to common referee concerns' signposting (Intro box / inline pointers) so the next pass can't re-raise them (concrete technique for pattern-068). 070 cross-vendor-agreement-weighting = reviewer-prompt rule 29 — weight the truth-audit by how many independent vendors flag the same item: 2-3 vendors=real, single-harsh-vendor (ChatGPT REJECTed P1A+P3 while Grok/Gemini gave major/minor)=likely referee variance. 071 de-biased-prompt-surfaces-more — the de-biased referee prompt raises raw MAJOR counts (a feature) but is only safe paired with the source-cited audit + integrity check; the durable asset is the instrument+audit pipeline, not any single prompt. Validated: RS5's 73 raw MAJORs truth-audited down to ~4 genuinely-new items, honestly.71RS5 · signpost + cross-vendor + de-biased-calibration: 29 reviewer-prompt rulesRS5 (07-01)RS11 · convergence-floor: 71 patterns · 29 prompt rules · 0 process/tooling — RS11 convergence-floor: patterns unchanged at 71, promptRules at 29. RS7-RS11 campaign validated pattern-066 (LLM-referee run-to-run variance) as the operative convergence theory — Grok flipped minor->major on unchanged content (RS10), 2 Gemini REJECTs (RS11) truth-audited to misreads. Finding-count trend (RS8=1, RS9=0, RS10=3, RS11=0) IS the convergence signal; the terminating gate is '0 genuinely-new real findings', not literal all-vendor ACCEPT. P4+P5 at genuine convergence floor; P1A/P2/P3/P1B at the LLM-refereeing practical ceiling — human referees next. Process/tooling counter starts here at 0 — the ~10 days of self-improvement below are backfilled from git, every increment sha-cited.71RS11 · convergence-floor: 29 reviewer-prompt rulesRS11 · convergence-floor: 0 process/tooling assets (cumulative, sha-cited)RS11 (07-01)verified-review-reset · I1-I5: 71 patterns · 31 prompt rules · 1 process/tooling — Verified-review reset (bigbounce commit 6357a9aa + scistack 40fe0cc): the I1-I5 durable review-routing fix. +2 reviewer-prompt rules (29→31): rule 30 = every EXT leg saves COMPLETE raw text + screenshot, orchestrator READS+verifies before recording any verdict (a leg with no output is FAILED, not a verdict); rule 31 = INT Claude leg is the Claude Code subscription subagent NEVER the Anthropic API, never fail an INT round on Anthropic-API billing, INT-fail never stops EXT, ChatGPT never silently dropped, Perplexity optional. +1 tooling = tools/v3_native_pdf_review.py de-required ANTHROPIC/PERPLEXITY keys + routed the Claude leg to a subagent (commit 6357a9aa). Trigger: Houston caught 'converged/18-18 ACCEPT' as fabricated (unverified sub-agent sweeps, no raw text).71verified-review-reset · I1-I5: 31 reviewer-prompt rulesverified-review-reset · I1-I5: 1 process/tooling assets (cumulative, sha-cited)verified-revie… (07-03)canonical-r-round-spec · DRY: 71 patterns · 32 prompt rules · 1 process/tooling — Canonical R-round spec consolidation (scistack a82bc5f + 8a5ae11): made astrostack/bigbounce-r-round/SKILL.md the single canonical INT/EXT round spec (DRY — all other R-round skills point here) and +1 reviewer-prompt/process rule (31→32) = HEADED browser is MANDATORY before any EXT sweep ($B connect; headless can't pass Cloudflare/Google-OAuth, silently loses reviewer sessions), Houston 2026-07-05 lesson. Tooling unchanged at 1.71canonical-r-round-spec · DRY: 32 reviewer-prompt rulescanonical-r-round-spec · DRY: 1 process/tooling assets (cumulative, sha-cited)canonical-r-ro… (07-06)same-commit-board + INT-parallel: 71 patterns · 34 prompt rules · 1 process/tooling — Two loop-discipline rules codified in the canonical spec (scistack 71e4a5c + 01688957): rule 33 = every verdict round MUST hit the /reviews board in the SAME commit as its artifacts and the loop never self-idles below the bar (2026-07-08 lesson); rule 34 = INT API lanes never wait on the browser — INT closure/science runs in parallel with EXT so an INT infra stall can't stall the round (parallel-resource rule). promptRules 32→34, tooling unchanged.71same-commit-board + INT-parallel: 34 reviewer-prompt rulessame-commit-board + INT-parallel: 1 process/tooling assets (cumulative, sha-cited)same-commit-bo… (07-07)directive-J + directive-G-leak-gate + URL-at-submit: 71 patterns · 37 prompt rules · 1 process/tooling — H16/W13 lessons + Houston directive J codified (scistack 000cd25 + c40ca88 + b570c78): rule 35 = STANDING literal 0/0/0 all-reviewer bar with never-idle parallel work (Fable orchestrator + Opus subagents, Houston 2026-07-09); rule 36 = directive-G leak gate — grep for review-process/audit language before EVERY recompile so internal-audit prose can't leak into a served PDF (P1U W13 lesson); rule 37 = URL-at-submit — capture the chat URL before any polling so a died agent can never orphan a submitted EXT leg (H16 failure mode). promptRules 34→37, tooling unchanged.71directive-J + directive-G-leak-gate + URL-at-submit: 37 reviewer-prompt rulesdirective-J + directive-G-leak-gate + URL-at-submit: 1 process/tooling assets (cumulative, sha-cited)directive-J + … (07-09)H17 accel round-1 · directive_g.sh + ledgers + Convex fixes: 71 patterns · 37 prompt rules · 5 process/tooling — H17 acceleration round-1 (ACCELERATION_LOG items 1-7) shipped 4 tooling assets (tooling 1→5): (a) tools/directive_g.sh one-shot PDF-hygiene chain — leak-gate + 0-undef compile + byte-identical mirror + Convex bump/read-back, per-closure hygiene ~15min→~2min, slug drift impossible (commit 533481ae); (b) canonical disposition ledgers project-context/peer-reviews/DISPOSITIONS/*.md — 107 numbered fingerprinted entries; audits cite D<P>-NN instead of re-writing from scratch, wave audit time ~halved (commit 4a2d551d); (c) tools/int_api_review reads \paperVersion live from the tex so review headers are always truthful (commit 729165b5); (d) convex/paperVersions.ts sortVersions Date.parse fix — killed the lexicographic 'July 10 < July 9' bug that left stale 'current' chips site-wide (commit 729165b5). Also documented pattern-066 in both directions (Grok MINOR→MAJOR AND MAJOR→MINOR on unchanged content) — referee variance is symmetric; no new pattern number. The fused-owner-loop pattern (one Opus owner iterates close→INT-retest→audit internally, returns once) codified in the canonical spec (item 1). patterns/promptRules unchanged — these are process/tooling, honestly not new review-patterns or reviewer-prompt rules.71H17 accel round-1 · directive_g.sh + ledgers + Convex fixes: 37 reviewer-prompt rulesH17 accel round-1 · directive_g.sh + ledgers + Convex fixes: 5 process/tooling assets (cumulative, sha-cited)H17 accel roun… (07-10)H17 accel round-2 · ext/int wave automation + verify-only + ETA instrument: 71 patterns · 37 prompt rules · 12 process/tooling — H17 acceleration round-2 (ACCELERATION_LOG items 8-13 + the readiness-ETA instrument) shipped 7 tooling assets (tooling 5→12): tools/ext_submit.sh + tools/ext_harvest.sh + tools/post_verdict.sh (EXT wave automation — proven submit/extract recipes, dead-chat detection, schema+slug+cap-formula baked in; commits 6ca8aae7, 085ce1ae, bfc8a76d); tools/int_wave.sh + tools/ledger_match.py (all-three-INT-legs parallel with raw-save enforced + a fingerprint pre-matcher that drafts the ledger match table so Opus adjudicates only UNMATCHED; commit 576b5ef9); directive_g.sh --verify-only flag (validate without re-mirroring/re-bumping — fixed the same-date tie-break that stole the Convex 'current' row; commit 576b5ef9); the readinessMetrics Convex table + computeEta query + rigorEvents annotations + tools/record_wave.sh + tools/backfill_readiness_metrics.py — a live per-paper per-wave verdict-trajectory chart + Submission-ready ETA widget, 288 rows backfilled from REAL EXT + H17 INT-API raws, missing legs recorded 'failed' (chart gap, never a zero); commits 4af8a76a, 6f4180cf, d177d100. patterns/promptRules unchanged — pure process/tooling.71H17 accel round-2 · ext/int wave automation + verify-only + ETA instrument: 37 reviewer-prompt rulesH17 accel round-2 · ext/int wave automation + verify-only + ETA instrument: 12 process/tooling assets (cumulative, sha-cited)H17 accel roun… (07-10)site-freshness-gate · pre-push: 71 patterns · 37 prompt rules · 13 process/tooling — Site-freshness pre-push gate (tooling 12→13): tools/site_freshness_check.sh + tools/hooks/pre-push (commit 0c263178) — a hard pre-push hook that BLOCKS a push whenever a public surface (banner / this very skills chart / /reviews board / version chips) has fallen behind the newest Convex wave or tools/ commit, killing the stale-surface class that let this skills chart sit flat for ~10 days while real work shipped. It is the standing enforcement of the exact staleness Houston caught; this backfill is its first cleared run. patterns/promptRules unchanged — process/tooling.71site-freshness-gate · pre-push: 37 reviewer-prompt rulessite-freshness-gate · pre-push: 13 process/tooling assets (cumulative, sha-cited)site-freshness… (07-11)directive-K · two-clean-waves exit: 71 patterns · 38 prompt rules · 13 process/tooling — Directive K codified as a loop exit-gate rule (promptRules 37→38 = rule 38): a paper CONVERGES on TWO consecutive clean waves (0 genuinely-new real findings on truth-audit) rather than a single literal 0/0/0 sweep — the two-clean-waves streak absorbs LLM-referee run-to-run variance (pattern-066) so one lucky quiet sweep can't declare convergence, and one noisy re-flag can't reset a genuinely-converged paper (Houston 2026-07-10, bigbounce 57b3bab3). Directive J's literal 0/0/0 stays the honesty target; directive K is the operational streak-gate that decides when the loop stops re-testing a paper. This is a genuinely-new reviewer-loop rule, hence a promptRules increment; patterns/tooling unchanged.71directive-K · two-clean-waves exit: 38 reviewer-prompt rulesdirective-K · two-clean-waves exit: 13 process/tooling assets (cumulative, sha-cited)directive-K (07-11)loop-watchdog + skills-autolog: 71 patterns · 38 prompt rules · 15 process/tooling — Never-again loop durability tooling (tooling 13→15, +2 new scripts): (a) tools/loop_watchdog.sh + tools/launchd/com.bigbounce.loopwatchdog.plist — an OS-level watchdog daemon + heartbeat gate that detects a wedged/dead cron-tick and recovers it within a 60-min cap, so the review loop can never again silently stall for hours (commit 3266efe8); (b) tools/skills_autolog.sh — this very generative skill/self-improvement changelog tool that emits sha-cited paste-ready reviewTimeline drafts and a --check gate that fails on any unlogged skill/process/tooling commit, shipped alongside the cron-tick retire/repair + hourly watchdog recovery (commit 0d77ba69). The scistack canonical bigbounce-r-round SKILL.md gained the matching watchdog-recovery-cap + heartbeat/skillslog freshness-surface docs and ext_{submit,harvest,post_verdict}.sh owner pointers (scistack d0347cf, 0cefd57, d4c6a7c). patterns/promptRules unchanged — pure process/tooling. Also cites the 07-09/07-10 skill/process commits this backfill closes out: 54d7a3cf 9fb9bb29 b20b8e95 d67b6a9a 61827a1d 86a7bee3 14e4f405 8bc9a4fc 0561982f 5a123fc6 7d914020 22c4d9bb 7d1af794 8a804f4.71loop-watchdog + skills-autolog: 38 reviewer-prompt rulesloop-watchdog + skills-autolog: 15 process/tooling assets (cumulative, sha-cited)loop-watchdog … (07-11)papers-link-gate + skills-date-granularity: 71 patterns · 38 prompt rules · 16 process/tooling — Site-freshness papers-gate + a granularity bugfix (tooling 15→16, extends tools/site_freshness_check.sh). NEW papers-gate: asserts per papers.ts paper block that the version chip == the pdfMeta version == the download-href version AND that every 'Read/Download PDF' href resolves to a served file — killing the stale-download-link split-brain the old gate never checked (a reader was served paper3_draft_v3.1.149.pdf while the chip already read v3.1.155; also caught P1U/P2/P4/P5 links lagging their source versions). Negative-tested: correctly flags chip v3.1.155 / href v3.1.149 → OVERALL FAIL, then restored green. BUGFIX: the skills-freshness check compared a DATE-only skillsSeries point against an hour threshold, so any tooling commit after ~12:00 UTC on the same calendar day tripped a false '>12h behind' — changed to date-granularity (stale only if a lesson/tool commit lands on a LATER day; the sha-based skillslog check remains the precise same-day enforcement). patterns/promptRules unchanged — pure process/tooling.71papers-link-gate + skills-date-granularity: 38 reviewer-prompt rulespapers-link-gate + skills-date-granularity: 16 process/tooling assets (cumulative, sha-cited)papers-link-ga… (07-11)loop-watchdog launchd-TCC repair: 71 patterns · 38 prompt rules · 16 process/tooling — Repaired the loop-never-dies backstop that the 07-11 watchdog tooling silently relied on but that was 100% dead in production (tooling count unchanged — a fix to existing tooling, not a new script). Root cause via a launchd kickstart: the com.bigbounce.loopwatchdog agent had NO ~/Desktop TCC grant, so launchd could not even read/exec tools/loop_watchdog.sh (getcwd + exec both EPERM) — the watchdog had produced ZERO real runs since a 07:13 Terminal-context call and had never once fired recovery; separately every hourly cron heartbeat write EPERM'd because a launchd agent can create new files under ~/Desktop but cannot overwrite an existing git-tracked one (macOS App-Management/TCC). Only the independent hourly cron kept the loop alive. CLASS-KILL without a Houston System-Settings grant: moved the authoritative runtime heartbeat + watchdog log into the launchd-owned ~/Library/Application Support/bigbounce/ dir, deployed the watchdog script there, and repointed the plist ProgramArguments + WorkingDirectory off ~/Desktop; the PASS path now touches no Desktop paths and recovery delegates repo work to `claude -p` (its own grant); best-effort repo mirrors' redirects are subshell-wrapped so a failed open never leaks EPERM to launchd stderr, and the log mirror is now gitignored like the heartbeat. VERIFIED in real launchd context: reload + kickstart rc=0, /tmp launchd stderr empty, fresh PASS line written to the runtime log by the launchd-context run. Canonical source stays tools/loop_watchdog.sh with a header note that it must be re-deployed to App Support after edits. patterns/promptRules unchanged — pure durability repair.71loop-watchdog launchd-TCC repair: 38 reviewer-prompt rulesloop-watchdog launchd-TCC repair: 16 process/tooling assets (cumulative, sha-cited)loop-watchdog … (07-12)ext-submit ApJS-variant support: 71 patterns · 38 prompt rules · 17 process/tooling — ApJS-variant reviewer-submission support (tooling 16→17). tools/ext_submit.sh gained a P3APJS paper key routing to pipelines/p3_anomaly_engine/paper3_apjs.pdf, and the ApJS-variant INT review lane (int_wave_apjs.sh) was wired in — so the P3 ApJS venue-variant (the PROVEN venue-flip whose ApJS-framed reviews are legitimate reviews of the same science, directive-M) can be submitted to the headed-browser EXT reviewers and the native-PDF INT API legs from the same harvest tooling as the PRD papers, instead of a hand-built one-off. Shipped alongside the M16-EXT P1U+P4 adjudication bundle (0 genuinely-new; streaks P1U 6→7 · P4 5→6; cap P1A 62→68 · P4 74 HOLDS). patterns/promptRules unchanged — pure review-harness tooling.71ext-submit ApJS-variant support: 38 reviewer-prompt rulesext-submit ApJS-variant support: 17 process/tooling assets (cumulative, sha-cited)ext-submit ApJ… (07-13)process-audit + M20-M36 harvest hardening: 71 patterns · 38 prompt rules · 18 process/tooling — Process audit 2026-07-14 (project-context/PROCESS_AUDIT_2026-07-14.md) cataloging 12 recurring EXT-harvest failure classes across the M20→M36 wave block, 8 of them ROOT-FIXED in committed tools this block (tooling 17→18 counts the harvest raw-sanity + paper-signature provenance gate now landing this cycle): (a) ChatGPT stale pre-send URL captured as the leg chat → PRE_URL-differs guard 3fb1ffd9; (b) submit_* trailing rc leaking through the set -e dispatch → orphaned leg with no OK/FAIL → esac||true dispatch guard a08dd750; (c) three ChatGPT 'rate-limit' failures were REDIRECT LATENCY not quota (sends landed server-side) → 120s poll + sidebar content-liveness fallback 02d68a8f + standing headed-diagnostic-before-accepting-a-cause rule; (d) OK recorded with an empty URL → die guard 80914698; (e) wrong-PDF-attach ×2 (M32/M34: a stale composer chip from the prior chatgpt leg) → composer-scoped attachment-token (ext_${PAPER}_${ROUND}) verification before send, both chatgpt AND grok, 854acb99; (f) harvest trusting labels (prompt-echo stub / 0-byte / misfiled other-paper raw) → adjudicator-layer directive-I4 catches today, harvest-layer raw-sanity + paper-signature provenance gate lands this cycle; (g) post_verdict cap picked list-order not _creationTime + record_wave full-patch clobbered rich wave rows (~8 rounds hand-corrected) → _creationTime-DESC selection + non-destructive listWaves skip-guard cd02c991; (h) an INT API leg posted under a bare EXT reviewerLabel displaced the EXT cap row → <wave>-INT-<vendor> convention 029cb689. Plus wave_submit.sh per-leg subshell isolation (compound chains can't orphan a sibling) + freshness-check Convex read retry (transient blip no longer blocks a push). Cross-refs ACCELERATION_LOG_2026-07-10 + REPEATED_ASKS_AUDIT_2026-07-11, not duplicating them. patterns/promptRules unchanged — pure review-harness durability.71process-audit + M20-M36 harvest hardening: 38 reviewer-prompt rulesprocess-audit + M20-M36 harvest hardening: 18 process/tooling assets (cumulative, sha-cited)process-audit … (07-14)autolog-2026-07-15: 71 patterns · 38 prompt rules · 27 process/tooling — Auto-logged 19 skill/process/tooling commits since 2026-07-14. The durable gains are: checkpoint recovery (7d3ed7fd); append-only PDF retention and archive inventory (18a8d50e, cb6262e3, 2000eef2); exact canonical paper/version routing and immutable packets (029158b8, 2bc0e985, 1f942478, be591f45, 6a3908a1, 0f268268, 35bbe3ec, 9f898efd, bfeb2813); sanitized direct-provider receipts (1fd28b3f); subscription-only OpenAI-family review enforcement with legacy API routes failing closed (4f7be06d, 51e9c24f, 1d1f1d03, scistack 66e7ec7); isolated exact-confirmation outputs (902cb712); a freshness-gate repair that distinguishes an intentionally paused anti-loop state from a dead loop while counting exactly six paper blocks (8c1cfe7b); and safe paper-scoped Convex figure reseeding so one caption repair cannot prune every paper (1b89827b). Nine new tools increased tooling 18→27; the final two repairs change existing tooling only. No new review pattern or prompt-rule number is claimed.71autolog-2026-07-15: 38 reviewer-prompt rulesautolog-2026-07-15: 27 process/tooling assets (cumulative, sha-cited)autolog (07-15)recursive-preflight-packaging-2026-07-16: 71 patterns · 38 prompt rules · 27 process/tooling — Recursive-loop enforcement and acceleration: registry-owned release aliases (dcf84a73); proactive P5 artifact-contract regression (f3bf8398); duplicate six-paper verification removed and volatile receipt hashes excluded from deterministic packet contents (349e33dc), cutting measured P3 dry-run wall time 36.20s to 13.05s; explicit P4/P5 bundle contracts (678e93fe); line-bounded standalone undefined-reference detection with positive/negative fixtures (fb98bc03); and exact P4/P5 isolated package proofs (aaf33368). No new pattern, prompt rule, or production tool count is claimed; existing gates became faster, deterministic, and executable.71recursive-preflight-packaging-2026-07-16: 38 reviewer-prompt rulesrecursive-preflight-packaging-2026-07-16: 27 process/tooling assets (cumulative, sha-cited)recursive-pref… (07-16)directive-N-claude-stack-day: 71 patterns · 39 prompt rules · 28 process/tooling — Directive-N Claude-stack routing day (promptRules 38→39 = rule 39; UTC date — the working session ran 2026-07-16 PT). Codex paused entirely on Houston's order; orchestration moved to Claude Code with Opus referee/truth-audit subagents + direct Grok/Gemini API legs, all raw-receipted (CLAUDE.md directive N, e42292b9). The day's durable process/tooling gains: the P4 science-contract preflight now enforces the PUBLISHED provider-overlay state and forbids the stale pre-publication disclosures (21fbef5d); the freshness-gate banner check reads the newest lastUpdatedISO instead of first-match (60d262aa); packages/namaster-proof/examples/rebuild_workspace_check.py ships a deterministic workspace-regenerability recheck (tooling 27→28, ce53033f); the dead Vercel Git integration (~50GB repo) was root-caused and replaced by a trimmed CLI static-deploy path serving only site-referenced PDFs (documented in ops memory + queue); and the P1B deposit identity was reconciled from the obsolete mcmc_companion block to the canonical namaster-proof paper with an exact v2B.0.11 tarball (27aeaabd, 0cb6a99d). patterns unchanged.71directive-N-claude-stack-day: 39 reviewer-prompt rulesdirective-N-claude-stack-day: 28 process/tooling assets (cumulative, sha-cited)directive-N-cl… (07-17)g1-pod-lane-tooling: 71 patterns · 39 prompt rules · 31 process/tooling — P4 G1 pod-execution lane tooling (tooling 28→31, +3 tools; UTC date — session ran 2026-07-17 PT): tools/runpod_ctl.py (RunPod GraphQL pod status/resume/stop controller), tools/pod_ssh.sh (established SSH exec pattern for pod jobs), tools/pod_bootstrap_sshd.sh (recovers a resumed pod that comes up with no SSH host keys / empty PUBLIC_KEY via the RunPod proxy — a real failure mode hit and fixed this session). Alongside (not counted here): pipelines/p2_chirality/train_g1_manifest.py, the manifest-retained ViT retrain wrapper whose committed g1_training_manifest.json (8,637 objects, every ID per split, all seeds, data revisions) resolves the historical no-retained-manifest gap behind the 26,616-vs-26,626 record conflict (commit 1c10bf79). patterns/promptRules unchanged.71g1-pod-lane-tooling: 39 reviewer-prompt rulesg1-pod-lane-tooling: 31 process/tooling assets (cumulative, sha-cited)g1-pod-lane-to… (07-18)publication-sprint-tooling: 71 patterns · 39 prompt rules · 32 process/tooling — Publication-sprint day tooling + durability (tooling 31→32): tools/seed_paper_figures.mjs hardened to strip begin/end{comment} blocks and %-comment lines before parsing (P1A's legacy 29-page draft parked in a comment environment was seeding 6 phantom figures with superseded -35/8 captions into Convex — the prebuild extract then reverted the corrected figures.ts, the exact directive-I6 durability failure predicted by the figure-regen pass); Convex paper_figures re-seeded from the corrected current tex (37 rows, 4 stale P1A rows pruned) so regeneration can no longer revert. Same day (counted as process, not new tools): the /publish Publication Command Center page + drift-proof /reviews//architecture boards (sourced from papers.ts), the Zenodo DOI mint+embed pipeline receipts, and the wave-1/2/P5 submission kits with standalone-compile proofs. patterns/promptRules unchanged.71publication-sprint-tooling: 39 reviewer-prompt rulespublication-sprint-tooling: 32 process/tooling assets (cumulative, sha-cited)publication-sp… (07-20)d2-deposit-acceleration-tooling: 71 patterns · 39 prompt rules · 34 process/tooling — D2 deposit-acceleration tooling (tooling 32→34, +2 tools; UTC date — session ran 2026-07-21 PT): tools/zenodo_deposit.py — the previously ad-hoc Zenodo upload/publish flow (used for the P2/P3/P4 DOI mint 2026-07-20) is now a committed, fail-closed tool: creates a private DRAFT deposition from a prepare_paper_deposit.py staging dir, uploads and MD5-verifies every file against local staging, writes a machine-readable receipt, and refuses to publish without the literal --confirm PUBLISH plus a license in metadata (publication is irreversible). tools/d2_authorize_deposits.py — turns Houston's pending D2 license decision into a one-command config authorization: injects the chosen license into the fail-closed P1A/P1B/P5 deposit configs with a mandatory --authorized-by Houston provenance stamp, clearing only the license gate (P1A fully unblocks; P1B keeps its namaster-proof software-DOI gate until --p1b-software-doi; P5 keeps its Paper-IV back-patch gate). Exercised live the same day: namaster-proof 0.1.7 (already MIT-licensed in-repo, so not D2-gated) was staged commit-bound and uploaded as a reversible Zenodo DRAFT with prereserved DOI 10.5281/zenodo.21481753 — all 5 file MD5s verified, receipt committed; publish remains Houston-gated. patterns/promptRules unchanged.71d2-deposit-acceleration-tooling: 39 reviewer-prompt rulesd2-deposit-acceleration-tooling: 34 process/tooling assets (cumulative, sha-cited)d2-deposit-acc… (07-22)july-patterns-minted-and-directive-m-amended: 77 patterns · 40 prompt rules · 34 process/tooling — Patterns 71→77 (+6) and prompt-rule 39→40 — Houston's 2026-07-23 reporting-layer audit minted into the catalog so July's lessons stop living only in tooling: pattern-072 worker-stall-storms → resume-once-then-take-inline; pattern-073 the reporting layer is a first-class review surface (stale grid/banner/widget read as science drift — every wave must land on EVERY rendering surface incl. Convex FUNCTIONS, same bundle); pattern-074 a paused reviewer leg must be visibly annotated everywhere it renders (the frozen GPT column read as fresh REJECTs); pattern-075 verify deployment identity before every deploy (Vercel junk-project relink + Convex prod-vs-dev split, two real hijacks in one week); pattern-076 fix legacy embedded content at the copy SOURCE (prebuild overwrote explorer fixes); pattern-077 canonical-entity whitelist at every aggregation point (the 7/8-papers bug: junk doc-id rows + retired P1U counted, P1B excluded). New prompt rule 40 = directive M-AMENDED (Houston verbatim: amend directive M to the legs we actually run): the all-A terminal criterion counts ACTIVE legs only — Grok API + Gemini API + Claude INT — with the paused ChatGPT column excluded-but-displayed-frozen until re-enabled. tooling unchanged.77july-patterns-minted-and-directive-m-amended: 40 reviewer-prompt rulesjuly-patterns-minted-and-directive-m-amended: 34 process/tooling assets (cumulative, sha-cited)july-patterns-… (07-23)directive-p-readiness-composition: 77 patterns · 41 prompt rules · 34 process/tooling — Prompt-rule 40→41 — directive P (Houston explicit, verbatim in CLAUDE.md): PUBLICATION READINESS is recomposed as science closure (25) + evidence/reproducibility (25) + automated review convergence (25) + packaging/PDF hygiene (20) + Houston's final personal per-paper review (5). Venue/endorsement/submission and independent human peer review move OUT of the score into a separate Publishing phase (tracked on /status + /publish, never subtracting). Convergence criterion made achievable-by-construction: 0 genuinely-new-real findings outstanding across ACTIVE legs (M-AMENDED) on current exact PDFs — verdict words are feedback, never the gate; per-finding source-cited truth audits unchanged. This supersedes the 2026-07-07 verdict-derived 50+points formula (avg-68 era). Result honestly recomputed: all six papers at 95 (four agent gates complete), the last 5 per paper = Houston's sign-off; Convex caps set to 95 x6 with static mirrors synced in the same bundle. Integrity rules absolute: no gate weakened, every finding still audited, 100 still requires Houston's recorded words.77directive-p-readiness-composition: 41 reviewer-prompt rulesdirective-p-readiness-composition: 34 process/tooling assets (cumulative, sha-cited)directive-p-re… (07-23)served-surface-integrity-and-companion-status-gates: 79 patterns · 41 prompt rules · 39 process/tooling — Patterns 77→79 (+2), both minted EXECUTABLE rather than prose-only — pattern-078 a companion's status goes stale the moment it is archived (first observed 2026-07-20/21 P5→P4, re-fired 2026-07-24 P5→P2 and P2→P1A), enforced by tools/verify_companion_status.py wired into bigbounce_preflight.py as the companion-status validator against project-context/companion-status-ledger.json; pattern-079 a served PDF outside the registered mirror set is invisible to directive G (first observed 2026-07-24, four independent accidental catches in one day), enforced by tools/verify_pdf_mirror_integrity.py as the pdf-mirror-integrity validator. Both fail-close review dispatch like any other preflight failure. The reverse-direction sweep that pattern-079 made executable found 31 orphaned served PDFs across 13 documents, every one outside directive G's enforced SERVED_ROOTS — so the enforced root list was itself the blind spot; SERVED_ROOTS now includes the bare public/ and downloads/ roots. Tooling 34→39 (+5): major_completeness_check.py, verify_companion_status.py, verify_pdf_mirror_integrity.py and their two regression tests under tools/tests/. promptRules unchanged at 41 — no reviewer-prompt rule was added in this window (this counter has no derivable source in-repo; it is carried forward, not measured).79served-surface-integrity-and-companion-status-gates: 41 reviewer-prompt rulesserved-surface-integrity-and-companion-status-gates: 39 process/tooling assets (cumulative, sha-cited)served-surface… (07-27)session-closeout-2026-08-28: 79 patterns · 41 prompt rules · 41 process/tooling — Session close-out wave: pod orchestration scripts committed as run under pipelines/p1_highz_tracers/clean_rerun/pod/ (supervisor v4 identifying torch workers by /proc/<pid>/exe, 2-hourly B2+HF backup loop, range-partitioned full-scan launcher, phase-3 chain, exe-based measure) with a README of hard-won lessons; P1C R13 partial-closure record + frozen-script erratum note; linter rules E/F (physics-claim and artifact-claim consistency) recorded from the R13 cycle; SESSION_HANDOFF_2026-08-05_to_2026-08-28.md as the re-entry document. Counters unchanged: scripts live under pipelines/, not tools/; no new pattern or prompt rule.79session-closeout-2026-08-28: 41 reviewer-prompt rulessession-closeout-2026-08-28: 41 process/tooling assets (cumulative, sha-cited)session-closeo… (08-28)p1c-consistency-linter-2026-08-08: 79 patterns · 41 prompt rules · 41 process/tooling — tools/p1c_consistency_check.py (tooling 40->41): a mechanical self-consistency linter for the P1C manuscript, built after the R11 board found four internal-contradiction defects introduced by iterative editing. Four rules: (A) constraint-count agreement across abstract/text/table/figure vs the actual barrier-entry count; (B) Tier-(I) prose claims vs the real count of Tier-(I) markers in Table II; (C) extensible assert/disclaim sentence pairs (a site asserting a bound another site explicitly disclaims); (D) universal-closure claims vs entries that declare themselves non-closures. Validated adversarially: against the pristine pre-fix v1C.0.13 it exits 1 and rediscovers three of the four referee MAJORs with no reviewer in the loop; against v1C.0.14 it passes 4/4. 16 unit tests. Documented in the clean-rerun RUNBOOK; deliberately not a git hook.79p1c-consistency-linter-2026-08-08: 41 reviewer-prompt rulesp1c-consistency-linter-2026-08-08: 41 process/tooling assets (cumulative, sha-cited)p1c-consistenc… (08-08)skills-autolog-2026-08-06: 79 patterns · 41 prompt rules · 40 process/tooling — Draft-paper review infrastructure wave: auxiliary draft registry merged into the review engine (bigbounce 84d53b75), first-class draft-paper records in portfolio preflight receipts with back-compat verification (d5e247bc), and a concurrency fix — the artifact-crosscheck validator captured process-global stdout during parallel leg verification, corrupting the hashed report into false receipt-stale failures; now uses a private stream with regression tests (9b92721d). Counters unchanged: extensions to existing tools, no new standalone tool/pattern/prompt rule.79skills-autolog-2026-08-06: 41 reviewer-prompt rulesskills-autolog-2026-08-06: 40 process/tooling assets (cumulative, sha-cited)skills-autolog (08-06)skills-autolog-2026-08-05: 79 patterns · 41 prompt rules · 40 process/tooling — Directive-Q wave: standing directive text (pure-contribution framing + mandatory reproducibility manifests; bigbounce 946c6655), reproducibility manifest schema v1 (a0fac40e), JSON schemas + tools/validate_repro_manifests.py validator (44b87570 — tooling 39→40), canonical paper-lineage disposition record confirming the retired 14-barrier no-go catalog is intact with resurrection recommended (03f1fde2), and the flat All-Papers site index with plain-English purpose subtitles (30e4676c). patterns/promptRules unchanged — process/tooling wave.79skills-autolog-2026-08-05: 41 reviewer-prompt rulesskills-autolog-2026-08-05: 40 process/tooling assets (cumulative, sha-cited)skills-autolog (08-05)finalization-maintenance-autolog-2026-08-04: 79 patterns · 41 prompt rules · 39 process/tooling — Maintenance autolog for all five skill/process/tooling-matched commits since 2026-07-27: publication-finalization prompt provenance (bigbounce bd89100b); P3 directive-G disclosure correction (bigbounce a59d53c2); duplicate project skill-mirror topology record (bigbounce c4eba285); existing native-PDF provider-routing and test repair (bigbounce b75c566d); generated SciStack skill-index refresh (scistack 90eb090). Counters intentionally unchanged: no new catalog pattern, reviewer-prompt instruction rule, or standalone tool was added.79finalization-maintenance-autolog-2026-08-04: 41 reviewer-prompt rulesfinalization-maintenance-autolog-2026-08-04: 39 process/tooling assets (cumulative, sha-cited)finalization-m… (08-04)skill-improvement-2026-09-02: 79 patterns · 42 prompt rules · 42 process/tooling — P1N/P4P/A3M closure wave (promptRules 41→42, tooling 41→42): prompt rule 42 = N-AMENDED routing directive, Sonnet-body/Opus-judgment/Haiku-polling worker split for the Fable 5.1 era across both internal and external API/CLI legs (bigbounce b3c5efd9). +1 tooling: p1n_r3_checks machine-checkable closure assertions replaced prose-only regression claims for P1N's R3 final closure (bigbounce af204341), paired with the P1N R1 merge-regression lesson — 3 regressions from the P1A+P1C merge caught and restored (bigbounce 82bb7752) — and recovery-benchmark fetch hardening for the anomaly-catalog known-object benchmark (bigbounce 80dcf196). Also landed: reproducibility manifest schema v1 additive paper-code enum extensions for A2/A3/P1N/P4P (bigbounce f297cc6e, 7e420e8b, ae21546c) and Codex launchd tick retirement now that directive N pauses the Codex/OpenAI lane (bigbounce 0b3cfaba). patterns unchanged at 79.79skill-improvement-2026-09-02: 42 reviewer-prompt rulesskill-improvement-2026-09-02: 42 process/tooling assets (cumulative, sha-cited)skill-improvem… (09-02)autolog-2026-09-03: 79 patterns · 42 prompt rules · 42 process/tooling — Three process improvements from the 2026-09-03 squeezed-monopole adjudication session (patterns/promptRules/tooling unchanged — pure process fixes, no new script or catalog entry): (a) Fable/Opus science lanes must create+commit their output file within the first ~10 tool calls — two Fable lanes stalled 2026-09-02 reading with nothing written; the 2026-09-03 plan-header/classical-kernel/verdict-manifest lanes (bigbounce d0662559, 67dbe4af, f3516042) all applied the rule and completed cleanly. (b) Sonnet execution lanes must not spawn nested background agents — the phase-3 v2 landing lane delegated to a nested agent that collided with the parent on shared files; fixed with an explicit no-delegation instruction in execution-lane prompts. (c) Sample-provenance preflight (OBJTYPE=TGT / FIBERSTATUS=0 + provenance gate) is now mandatory before any GPU-billed anomaly run, per ANOMALY_SAMPLE_CONTAMINATION_2026-09-03.md. Also closes out the P3 anomaly catalogue v2 data-release doc (bigbounce 0e9e5b41, directive Q1).79autolog-2026-09-03: 42 reviewer-prompt rulesautolog-2026-09-03: 42 process/tooling assets (cumulative, sha-cited)autolog (09-03)a3m-r6-closure-2026-09-05: 79 patterns · 42 prompt rules · 42 process/tooling — A3M R6 truth-audit closure bundle (v3M.0.13->v3M.0.14): tools/a3m_convex_bump_v3M_0_14.mjs added (routine per-round Convex bump script, same pattern as prior paper bump scripts) -- lands at UTC calendar-day 2026-09-05 due to a -07:00 local-vs-UTC boundary on the commit timestamp (session date 2026-09-04 PT). No new review pattern or prompt rule this wave -- patterns/promptRules/tooling unchanged at 79/42/42; this point exists solely to keep the skills-freshness date-granularity gate current with the newest tools/ commit.79a3m-r6-closure-2026-09-05: 42 reviewer-prompt rulesa3m-r6-closure-2026-09-05: 42 process/tooling assets (cumulative, sha-cited)a3m-r6-closure (09-05)autolog-2026-09-04: 79 patterns · 42 prompt rules · 42 process/tooling — Auto-logged 9 skill/process/tooling commit(s) since 2026-09-03 (8 bigbounce, 1 scistack): skills-autolog housekeeping; P3 anomaly catalogue v2 data-release doc; A3M v3M.0.12 paperVersion bump + Fig. 1 regeneration with publication labels (directive I6); site redesign /reviews grid + six-lane pattern logging; full-reproduction pass kickoff (directive Q2); SIGW nHz reproducibility manifest (directive Q2); scistack generated skill-index refresh. patterns/promptRules/tooling unchanged — process/doc/science wave, no new catalog entry or standalone tool.79autolog-2026-09-04: 42 reviewer-prompt rulesautolog-2026-09-04: 42 process/tooling assets (cumulative, sha-cited)autolog (09-04)autolog-2026-09-07: 79 patterns · 42 prompt rules · 44 process/tooling — Auto-logged 2 skill/process/tooling commit(s) since 2026-09-05 (both bigbounce, +2 new tools/ scripts): tools/su_convex_bump_v1S_0_8.mjs and tools/a3m_convex_bump_v3M_0_24.mjs, the routine per-round Convex bump scripts for the A2 lapse-monopole reconciliation wave (paper-su v1S.0.8, A3M v3M.0.24). patterns/promptRules unchanged at 79/42; tooling 42->44.79autolog-2026-09-07: 42 reviewer-prompt rulesautolog-2026-09-07: 44 process/tooling assets (cumulative, sha-cited)autolog (09-07)brand-pass-reverted-2026-09-08: 79 patterns · 42 prompt rules · 44 process/tooling — Brand-unification pass on the lab site (2026-09-08, commits fa5a8211..eae25baf) REVERTED at Houston's direction (commit 00d769ea): synced Hubify tokens with a teal accent override, a five-act homepage recomposition, journal-style paper pages, mark/lockup components and the tools/sync_brand_tokens.sh sync script are all removed; the 2026-09-04 design he approved is restored byte-for-byte. Process learning, not a tooling delta (counters unchanged at 79/42/44): the lab site is the REFERENCE standard for the portfolio's visual language -- Hubify moves toward it (more white space, calmer minimalism, its palette), never the reverse; a cross-property brand pass must be proposed against the reference and approved before any lane touches the reference itself. The shared brand system, token file and coordination notes remain in the hubify repo for that property's own use.79brand-pass-reverted-2026-09-08: 42 reviewer-prompt rulesbrand-pass-reverted-2026-09-08: 44 process/tooling assets (cumulative, sha-cited)brand-pass-rev… (09-08)review patternsreviewer-prompt rulesprocess/tooling (sha-cited)

Round timeline · newest first · showing 60

Every round, truth-audit, closure, and skill upgrade

One line per event: date, kind, what changed, receipt link. Skill-improvement entries carry a quiet marker.

2026-09-08· skill improvementLab-site brand pass reverted -- the lab is the reference standard
2026-09-07· skill improvementA2 monopole gap RECONCILED -- paper-su v1S.0.8 + A3M v3M.0.24
2026-09-07· skill improvementS9c: dropped-evolution-term hypothesis NOT supported -- paper-su v1S.0.7 + A3M v3M.0.23
2026-09-05· skill improvementS9b: intrinsic term does not close the -55/16 gap -- paper-su v1S.0.6 + A3M v3M.0.22
2026-09-05· skill improvementS9/S10 second-order closure: paper-su v1S.0.5 + A3M v3M.0.21 -- uniform-density δN not reconciled
2026-09-05· skill improvementS7 located: Cai (2009) factor-of-2 statement corrected in paper-su v1S.0.4 and A3M v3M.0.20
2026-09-05closure waveP4′ v4P.0.7: row-16 (iv-b) DESI DR1 BGS external environment result added
2026-09-05closure waveP1B v2B.0.20 — Batch 3 value-level rule (R7) + PyMaster cross-check integrated
2026-09-05closure waveP1B v2B.0.21 — R3 truth-audit closure: batch-4 post-commitment verifier challenge (R8) answers R7's two design MAJORs
2026-09-05closure waveP1B v2B.0.22 — deferred D-R3-21 closed: all three injected-angle recovery sigmas now report mean+-SEM
2026-09-05closure waveP1B v2B.0.23 — three remaining deferred text items from the R3 audit closed
2026-09-04closure waveP4′ v4P.0.6: ledger row-16 results integrated as disclosure/robustness content
2026-09-04closure waveLedger row 16: N=20,000 pixel-level parity injection resolves the row-13 slope comparison outside the noise floor
2026-09-04internal ccPSU: R2 verification board in progress on v1S.0.2
2026-09-04closure wavePSU v1S.0.2: R1 truth-audit closure + D-PSU-1 reframe -- exact change of variable, not a 'failure'
2026-09-04closure waveP1B v2B.0.19: R2 closure -- statistics presentation corrected, rounds stopped under R2
2026-09-04internal ccP1B: R2 verification board in progress on v2B.0.18
2026-09-04closure waveP1B v2B.0.18: R1 truth-audit closure -- pre-registered batch-2 sealed blind test now primary result
2026-09-04closure waveA3M v3M.0.17: D-A3-12 -- scheme-independent tensor no-go + c_s-dependent bounce cubic term
2026-09-04closure waveA3M v3M.0.18: R8 truth-audit closure -- scope statement for the no-go (D-A3-13)
2026-09-04closure waveA3M v3M.0.19: D-A3-14 (ledger row 19) -- the no-go generalizes to the full P(X) k-essence class
2026-09-04closure wavepaper-su v1S.0.3: R2 truth-audit closure -- 20 genuinely-new-real findings, new Appendix A
2026-09-04closure waveA3M v3M.0.16: R7 truth-audit closure -- 16 genuinely-new-real findings, 0 physics errors
2026-09-04· skill improvementP1B v2B.0.17: reframed as a general verification primitive, sealed blind shortcut-detection test
2026-09-04closure waveA3M v3M.0.15: D-A3-10/D-A3-11 science reframe -- joint (r, f_NL) no-go
2026-09-04closure waveA3 v3M.0.14: R6 truth-audit closure -- 0 physics errors, rounds stopped under R2
2026-09-04closure waveA3 v3M.0.13: abstract to cap, 15 pp
2026-09-04· skill improvementSkill/process autolog 2026-09-04: 9 commits landed via tools/skills_autolog.sh
2026-09-04restructureSite redesign 2026-09-04: three tracks as the spine, flat works index, slim topbar, evidence grades, honest nulls
2026-09-04· skill improvementSkill improvement: file-disjoint six-lane redesign pattern (write-early, commit-per-page)
2026-09-04closure waveA3 v3M.0.12 -- R5 truth-audit closure C1-C7 + DA3M-R5-15/18; one verification round remains under R2
2026-09-04closure waveA3 v3M.0.11 -- science decision D-A3-9 (ledger row 9): transmitted amplitude restated as a two-scheme band
2026-09-04closure waveA3 v3M.0.10 -- R4 verification board, editorial closure, PTA channel restated as a null; rounds stopped (R2)
2026-09-04closure waveA3 v3M.0.9 -- R3 truth-audit closure (C1-C10); readiness 75
2026-09-04closure waveA3 v3M.0.8 -- three closed science-gate results integrated; readiness 75
2026-09-03restructureAnomaly catalogue → data release (ledger #8 answered, phase-3 v2 landed)
2026-09-03· skill improvementAnomaly sample provenance: S>8 sample found 84.8% sky fibers; science-target rerun launched
2026-09-03· skill improvementFable/Opus science lanes must create+commit their output file within the first ~10 tool calls
2026-09-03· skill improvementSonnet execution lanes must not spawn nested background agents
2026-09-03· skill improvementSample-provenance preflight before any GPU-billed anomaly run
2026-09-02closure waveA3 v3M.0.7 — real-KDE injection validation closed; readiness 75
2026-09-02restructureA3 v3M.0.6 — final-review REVISE (abstract); readiness 70
2026-09-02restructureA3 paper v3M.0.5: R2 CLOSED — real 30-bin injection validation, Eq. (8) σ² term, Ω_DM fix; rounds STOP per directive R2 (2/2 consumed)
2026-09-02restructureA3 paper v3M.0.4: R1 closed (NANOGrav posterior primary, handoff-conditional transmission bound, PBH ratio disclosed); R2 verification dispatched
2026-09-02restructureA3 paper v3M.0.3: PBH compaction-function channel integrated (ratio-level result); R1 board dispatched
2026-09-02· skill improvementP4′ and ECH Note abstracts trimmed to venue caps (v4P.0.5 / v1N.0.5) — final-review REVISE executed; tarballs rebuilt
2026-09-02restructureA3 multi-channel paper registered (v3M.0.2 draft) — Track A flagship submission candidate
2026-09-02internal apiECH Note R3 verification (Claude major / Grok reject / Gemini major) → v1N.0.4 with machine-checked regressions; automated review converged; final author review APPROVE; readiness 95
2026-09-02internal apiP4′ R3 verification (Claude minor / Grok reject / Gemini minor) → v4P.0.4; automated review converged; final author review APPROVE; readiness 95
2026-09-02internal apiP2′ R1 (Fable major / Grok reject / Gemini major) → v2L.0.2; scope decision: folded into the A3 paper
2026-09-02internal apiECH Note R2 closed → v1N.0.3 (23/23 incl. two errors inherited from P1C: 8π coefficient, O5 parity); R3 verification pass dispatched
2026-09-02internal apiP4′ R2 closed → v4P.0.3 (21/21; monopole disclosed; genuine 95% CL limit ≈0.75% by Neyman inversion); R3 verification pass dispatched
2026-09-02internal apiP1N v1N.0.2 and P4′ v4P.0.2 R2 Claude legs: major-revisions
2026-09-02internal apiLedger #1 closed: independent in-in reproduces −35/16; P2′ Letter drafted, R1 dispatched
2026-09-02internal apiECH Note R1 closed → v1N.0.2 (19/19), R2 board dispatched, verdicts pending
2026-09-02internal apiP4′ R1 closed → v4P.0.2, R2 board dispatched
2026-09-02internal apiR1 boards dispatched: P1N and P4′ (INT-only)
2026-09-02restructurePortfolio restructure: Track A/B/C replaces three research programs
2026-09-02· skill improvementP1N/P4P/A3M closure wave — routing directive, manifest schema, machine-checkable regressions, fetch hardening
2026-08-28· skill improvementSession close-out: pod orchestration scripts, P1C R13 partial-closure record, handoff document

Full history (append-only, 448 rounds total) in reviewTimeline.ts.