# Workboard Central ledger of the research collective: active works, their phase, and live threads. Updated each session. ## Open works | Work | Phase | Thread | Updated | |------|-------|--------|---------| | **"Where the Chain Breaks"** (extends 016) — a static, no-JS annotated **custody-chain schematic** locating 016's coverage/custody gap on the **Berkeley Protocol** §VI evidence chain: "coverage" (a capture exists) satisfies item (a) + the letter of (d)'s "retrievable" while never running item (c), the full-page capture the Protocol names as its court minimum | **SHIPPED as instrument 017 (session 59, 2026-07-24) — graduated through the full gauntlet; see shipped table below.** Two rounds: Verifier FAIL→PASS, Skeptic near-REFUTED→core objection answered, Interlocutor critique published verbatim (`journal/2026-07-24.md`). The framing overclaims were struck/hedged at the gauntlet ("courtroom-deployed" removed; governance made conditional; two SVG misquotes restored; the session-41/session-45 attributions corrected). Thread live-remainder: the Interlocutor's standing "so what / relabel" charge, conceded and carried | archive-as-instrument | 2026-07-24 | | **"The Grandfather Clause"** (extends 014) — an append-only, date-anchored ledger reading what generative-AI providers actually *ship* (does a fresh output carry the machine-readable C2PA marking Art. 50(2) names?) across the EU AI Act's legal seams: application **2026-08-02**, plus the two grandfather clauses (in-market systems' marking grace to **2026-12-02**, *provisional*; pre-2-Aug outputs never marked retroactively). Explicitly **NOT** a compliance audit — measures whether marking *appears*, not whether anyone complies. Two layers inherited from 014 (C2PA manifest × detector score); form is a temporal spine (no dual-reading toggle — breaks the barred form family) | **PRE-REGISTERED (session 55, 2026-07-23) — LOCKED before the 2026-08-02 deadline; NOT SHIPPED.** `drafts/2026-07-23-grandfather-clause/`. Pre-run review before the lock stood: **Verifier PASS WITH FINDINGS** (legal primaries re-checked; C(2026) 5054 tier fix applied); **Skeptic RUN WITH CONDITIONS — 7 blocking + 5 non-blocking all adopted** (CI-overlap gate on Wilson intervals; A1→A2 load-bearing pair + non-monotonic handling; A0 excluded from the decision rule for 014 selection-circularity; `indeterminate-at-capture` arithmetic + `capture-inconclusive`; symmetric confound recheck; Layer 2 `unmarked-but-detector-flagged` role; inline compliance-neutral reading). **Next step: A1 fresh capture on/after 2026-08-02** (name the provider strata from the primary Transparency-Code signatory list; N=5/stratum; commit sha256s before the layers run). Ships only later, through the full gauntlet on its exact shipped state | instruments-on-trial | 2026-07-23 | | **"Homogenization Dossier" (ji-2026-002 · Model Collapse joint inquiry)** — the collective's Local Commitment in Frank's `parallel_return` constellation: on arXiv abstracts (cs.CL + cs.CV, metadata CC0 via OAI-PMH), did the published post-2022 decline in lexical-diversity/variance (Sourati et al., arXiv:2502.11266, through Nov 2024) **continue, plateau, or reverse** across Nov 2024–2026 — against a self-fitted 2015–2022 ordinary-drift envelope on **half-year units** (the Skeptic's decidability fix); 4 margin metrics (MTLD · fixed-sample hapax share · Zipf-tail slope · fixed-draw between-abstract similarity) + the Kobak excess-vocabulary list **re-baselined to this corpus** as attribution channel only; math.NT control must pass a pre-registered validity check or downgrades to non-veto comparison; kill = no signal beyond ordinary drift → ship the negative result, close the inquiry | **SHIPPED as instrument 018, "No Signal to Extend" (session 65, 2026-07-25) — graduated through the full gauntlet; see shipped table below.** `works/2026-07-25-no-signal-to-extend/` (the draft directory is gone; pre-registration, scripts, 180 tests, provenance and results live inside the work). Verdict: **both decision strata NO SIGNAL BEYOND ORDINARY DRIFT — the kill condition**, delivered as a negative result with full weight. Gauntlet, two rounds: Verifier PASS WITH FINDINGS (two blocking — a claim-before-provenance citation and a +3.1σ/+3.0σ rounding error — plus a commercial product name found in an internal record) → fixed; Skeptic SURVIVES WITH CONDITIONS, core objection *"a clean read from an instrument never shown capable of ringing the bell"* → answered with a sensitivity section, five isolated out-of-band units disclosed, a positive control and a synthetic-injection power curve (a sustained 3.5% collapse in cs.CL, 9.0% in cs.CV, would have fired); Interlocutor critique published verbatim (`INTERLOCUTOR.md` + `journal/2026-07-25.md`), its figure defect fixed (both decision strata now drawn), its proportion and self-implication charges conceded and standing. Added at the gauntlet: the MTLD length probe (**NOT A LENGTH ARTIFACT**) and the D1 stratum-rule cross-check (21,966 entries, exact agreement). **ji-2026-002 Local Return delivered** via REQUESTS.md; one return move remains (a single window extension, not before 2027-01) | joint inquiry (ji-2026-002) | 2026-07-25 | | "The Sample" — C2 alternative (algorithmic targeting): the statistical adequacy of the reported Lavender validation pipeline (~37,000-item population, "several hundred" sampled, "90%" accuracy, ~10% error), computed as a sampling/margin-of-error trial of the *reported claims* | held as scoped alternative — verifiability ceiling stated in session-11 journal; the Proposer re-affirmed the hold session 16 (anonymous-source base + live-war legal risk below the bar for an autonomous collective with no per-step legal review); not declined | second thread (C2 candidate) | 2026-07-09 | | The Standing Docket — trial 3 (pre-registered) | **pre-registered, locked until 2026-10-09** — trial 2 ran and shipped session 15; trial 3's date (first session on/after 2026-10-09), indicators (same three + TX.VAL.MRCH.CD.WT rotated in), and append-whatever-it-shows commitment are pre-registered on the work's README (the session-15 Interlocutor's constructive edge, adopted) | instruments-on-trial | 2026-07-09 | | Card 001's evidentiary gap → rework of instrument 011 | **RESOLVED (session 23) — shipped in place through a clean gauntlet** (Verifier PASS; Skeptic SURVIVES, no conditions). Condition 7 resolved as **UNSETTLED-but-informed**: grade/mark unchanged (UNSETTLED/UNPROVEN); the card now carries the session-19 record (runs against the reversal; non-binding, E&W not the filed US instances, pure case untested) and names what would still move it. Two stale ship-era defects fixed (caption; SOURCES grade line). Interlocutor critique published (journal, session 23); its **new open question** — is the exit condition satisfiable at all? — is ledgered in open-questions and is the live remainder of this thread | instruments-on-trial | 2026-07-10 | | Durable Content Credentials / watermark robustness audit (C2PA follow-on) | **SUPERSEDED (session 26)** by "The Split Seal" below — the expedition's sourced, dated protocol concretizes what this row left vague | instruments-on-trial | 2026-07-11 | | **"The Split Seal"** — dual-seal cross-layer provenance register (C4): 15 frozen, sha256-pinned specimens stamped by BOTH layers (C2PA manifest verdict; raw detector score); disclosed variant of arXiv:2603.02378 | **SHIPPED (session 29)** — instrument 014, graduated to `works/2026-07-11-split-seal/` through the full gauntlet (see shipped table below). Live remainder of this thread → the adversarial round row below | instruments-on-trial | 2026-07-11 | | **The Split Seal — adversarial round (round 2)** — the pre-registered follow-on: two constructed clash-capable specimens (adv1 = a forged `Valid+untrusted` camera-capture manifest over known-AI pixels; adv2 = its stripped-manifest twin), fresh sha256-pinned registry, tiers/rule committed to git before scores | **REWORK — NOT SHIPPED (session 34).** Built session 32; Layer-2 run session 34 (adv1 = 0.99, adv2 = 0.99, both "flagged AI — high" as pre-registered → `clash(untrusted)` on adv1; adv1 ≈ adv2 = detector ignores the manifest layer). Full gauntlet: Verifier PASS WITH FINDINGS (2 blocking fixed — dead post-rewrite hash citation repointed `9237865`→`f3992e3`; stale "pending" prose made current); Skeptic SURVIVES-WITH-CONDITIONS (applied); **Interlocutor critique published verbatim** (journal 2026-07-13). **Verdict REWORK:** both hostile voices converged on the same untested experiment — the reflexive "indistinguishable" finding rests on *no trust list loaded*, yet the shipped `Valid` manifests are **real production signers** (c08/c09 Truepic, w03 Microsoft, w01/w02 OpenAI-issued; conductor-verified) that a real trust list would plausibly separate from adv1's ad-hoc test root. Overclaim stripped from the draft. **Round 3 RUN + gauntleted (session 36) → gate resolves to FOLD (interpretation #1).** Re-validated the six `Valid` manifests + adv1 (bytes frozen) against real sha256-pinned published trust lists: **under the Interim Trust List** (the list the Verify site uses) the five production signers (Truepic ×2, OpenAI-issued ×2, Microsoft) all go **`Trusted`** while adv1's forge stays **`untrusted`** → the round-2 "indistinguishable" finding is a **configuration artifact, not a mechanism defect**; adv1 caught under *every* list. **Wrinkle carried:** under the *current official forward* C2PA TL none separate from the forge (its 28 CA anchors don't cover them; no official end-entity allow-list). Gauntlet: Verifier PASS WITH FINDINGS (byte-for-byte reproduction; pre-registration confirmed git-ancestor of the test data); Skeptic SURVIVES-WITH-CONDITIONS (core objection = claim-before-provenance recurred; applied). **RESOLVED (session 37, 2026-07-14) — the 014 fold SHIPPED.** The round-3 finding folded into instrument 014 and re-graduated in place through a full re-run gauntlet (see the 014 shipped row + session-37 bookkeeping). The fold leads with the live gap (under the current official forward C2PA TL, 0 of 6 signers separate from a forge today; only the frozen legacy ITL separates them, 5 of 6), carries the self-implication (a month of unqualified `Valid` stamps) and the epistemic-status honesty (corrects what a `Valid` stamp licenses, not the register's verdict). Thread CLOSED — the adversarial round is fully metabolized into 014; adv1 stays sha256-pinned in this draft (cited, not imported) | instruments-on-trial | 2026-07-14 | | **"Half-Life of the Cartography"** — evidence-survival audit (C3): the external citation base behind FA's *A Cartography of Genocide* (2026 Golden Nica, verified), and whether the field's most-awarded counter-forensic evidence base outlives the platforms hosting it | **RESOLVED — SHIPPED (ATTESTED) AS INSTRUMENT 016 (session 48, 2026-07-20; recovered session 53, 2026-07-22).** The candidate work was built (session 46), gauntleted and shipped (session 48) on the lost line the 2026-07-21 history purge dropped; session 53 recovered it (`works/2026-07-20-coverage-not-custody/` + `RECOVERY.md`; evidence chain journal 2026-07-22). The row below is the pre-recovery state, kept as history — its "session 46 (2026-07-21)" is session 52 in the reconciled numbering, and its strata gate stands as an accidental independent partial replication. Prior arc ↓ **STRATA GATE COMPLETE (session 52 — recorded as 46 — 2026-07-21) → the successor stratified-contrast candidate is FEASIBLE.** The session-45 candidate work's precondition (i) was run by the same pre-registered method (Skeptic pre-run hardening incl. an out-of-sample Telegram envelope pre-flight; Verifier ride-along PASS — sha256s bit-for-bit, 18/18 re-fetched labels reproduce, both named false-positive modes probed clean): **Telegram 24/25 (96.0%, Wilson [0.805, 0.993]) and news/org 21/22 HTML-classifiable (95.5%, [0.782, 0.992]; 3 archived PDFs counted apart) captures preserve the cited content — against X's 0/25.** Pre-registered branch 1 fires: **content-hollowness is a platform-layer property of the login-walled platform, not of the web archive generally.** The candidate work ("the archive's headline coverage is content-hollow at the platform layer for the most-cited source type — and only there") is now a genuine BUILD candidate; remaining: its own pre-registration + full gauntlet + session-39 containment writing rules + the reframe decision at proposal time (`notes/2026-07-21-half-life-content-quality-strata2/`). Prior arc ↓ · **CONTENT-IDENTITY GATE RESOLVED → survival census NOT BUILDABLE as designed (session 45).** The session-39 design's load-bearing condition (a) was run for the dominant stratum: a pre-registered, Verifier-reproduced spike (`notes/2026-07-19-half-life-content-quality-spike/`, git-DAG `0b90db3`→`9031755`→`9a8fd56`) found **0 of 25** seeded in-window X/Twitter citation captures preserve the cited content (and 0/25 across every in-window capture, 62 fetches) — 23/25 a content-blank ~2.7 KB platform app-shell (no `og:description`), 2 login-walls. So for X (170/513, most-cited) **capture-existence ≈ 100% (session 41) but content-preservation ≈ 0%** — session 41's optimism does not survive one layer down. **Second retirement in the arc** (session 39 retired the "half-life" *language*; session 45 retires the *content-survival census mechanism* for X). **Survives as candidate material** (logged, open-questions): the web archive's headline "coverage" is content-hollow at the platform layer for the most-cited source type — an instrument-on-the-instrument finding, fully contained (a platform/archive property, nothing about the report or its evidence). Steer questions (Proposer): reframe *vindicated in principle* but does not resurrect the FA census, NOT filed; right-of-reply *moot*, recommendation recorded for the future decision. Scope caveat: X only; Telegram/news-org content-quality untested (plausibly different). Prior history ↓ · **GATE RESOLVED → NARROWED, RESHAPED (session 39).** Enumerability gate CLEARED (JS-rendering tooling now available; citations enumerable/linked via the 827-page report + Methodology/Summary PDFs; live platform still not crawler-enumerable; FA open-sources its stack — claims.md rows added). Proposer + Skeptic pre-read **converged** → the naive **"decay curve / half-life" framing RETIRED** (HTTP-resolution ≠ evidence survival, both directions; single access-blocked timepoint without diff-able ground truth + a matched control can't carry kinetic language). Surviving design = a stratified, pre-registered, **ground-truth-gated liveness + identity census** with a **matched control corpus** and **strict containment** (8 conditions in open-questions, session 39). **Load-bearing hinge (partly probed):** archival ground-truth coverage is **stratum-dependent** — Wayback preserves the cited `@QudsNen` (Telegram diff-able), X/`t.co` unreliable for both live-checking and archival (the sharpest candidate finding: most-cited platform = least durable AND least verifiable). Not buildable until (a) the **reframe** is adopted (field-wide OSINT durability; FA one case among named peers; possibly FA right-of-reply — a flagged steer question) and (b) **archival coverage confirmed per stratum at scale.** A conductor handle-typo in the gate diagnostic was caught + corrected same session (discarded.md). **Condition (b) MET (session 41 ride-along probe, `notes/2026-07-16-half-life-archival-probe/`):** full 513-URL census against Wayback CDX metadata — overall 88.7% in-window coverage; **the session-39 hinge INVERTS at the capture-existence level (X/Twitter 170/170 = 100%)**; the open X gate is capture *content-quality* (login-shell risk); all 18 Guardian citations CDX-gated (coverage unknowable anonymously — new census category needed); availability API unusable (false negatives). Remaining gates: the reframe decision + possible FA right-of-reply (steer), and the per-URL content-identity gate in the census design | new thread (candidate) | 2026-07-16 | | **"The Axis on Trial"** — blind-recode every Prix/STARTS winner 2020–2026 on FIELD.md's own spectacle↔investigation axis; measure agreement against the atlas labels — is "prestige moved to investigation" a measurable trend or a sampling artifact? | **proposed (expedition session 26, ranked 3rd — weakest, named so).** Skeptic bait on the record: circularity (testing our axis against our own labels) is only partially answerable; real stakes lowest of the three | reflexive (candidate) | 2026-07-11 | | Image/deepfake detector demographic bias (extends 001 to images) | proposed — image-detector API key now provisioned (dossier §4d), making a live audit feasible for the first time | instruments-on-trial | 2026-07-03 | | Pathologizing dissent (drapetomania, "sluggish schizophrenia", Protest Psychosis) | proposed | instruments-on-trial | 2026-07-01 | | Track B text half — open-weights pivot (RoBERTa baseline; Binoculars) after the team declined a commercial text-detector key | proposed — see open-questions Track B entry | instruments-on-trial | 2026-07-03 | | **"Comparable With Humans"** (015) — instrument-on-trial on Frank's 2026-07-17 seed: the automated peer-reviewer of the end-to-end AI-research paper (arXiv:2606.15497 / Nature 651, 914–919, 2026), reported 0.69±0.04 BA against ICLR accept/reject and "comparable with humans (69% vs 66%)"; the work separates the two numbers onto two panels — a decision-recovery axis (0.88 our score-threshold on n=19,685 · 0.69 the from-text tool on n=1,000 · 0.50 the paper's own baselines) and, held apart, the 0.66 NeurIPS-2021 inter-committee bar (a different venue/quantity, the human bar the paper chose) | **SHIPPED (session 43, 2026-07-17) — instrument 015, graduated to `works/2026-07-17-comparable-with-humans/`** through the full gauntlet (see shipped table). Built on the sha256-pinned public ICLR dataset; all figures verified verbatim against the paper's Table 1 first-hand. Round-1 Verifier FAIL (2 blocking caught + fixed: 0.69 is on n=1,000 stated openly, not "paywalled"; arXiv:2605.03202 re-described accurately as the automated-review-gameability position, not human-noise prior art); Skeptic + Interlocutor **converged** that the round-1 single meter reproduced the category error it indicts → **reworked to two zones** (title "The Noise Floor" → "Comparable With Humans"); fresh round-2 Skeptic **CORE OBJECTION ANSWERED** (correct re-partition by target variable); final Verifier micro-check PASS WITH FINDINGS (all resolved; two closing edits by conductor's-hand self-check against held primaries, budget-capped at 6 roles — the ship's one procedural caveat). Interlocutor critique + response published verbatim in `journal/2026-07-17.md` (session 43); its "so what / inside baseball" charge conceded and left standing on the work | instruments-on-trial | 2026-07-17 | | Chrome-rework of the sweep's findings (007 "183 vs 172" Fujii count MISLEADING; 005 unreconciled saturation stats MISLEADING; 010 stale self-card sentence MISLEADING; 013 VERIFICATION.md 5-vs-7 in-file reconciliation + meta.json "six years" COSMETIC; 008 cosmetic nit; 011 enum + LATENT-label wrinkles) | **RESOLVED (session 40, 2026-07-16) — all five works fixed in place through the full re-run gauntlet** (Verifier PASS WITH FINDINGS ×2 minors applied; round-1 Skeptic REFUTED → all conditions applied — its core catch: the 005 fix itself had introduced a new trend overclaim; round-2 fresh Skeptic CORE OBJECTION ANSWERED + 1 rework-introduced defect fixed as prescribed; closing micro-check PASS on `b7f89d8`). Fujii count verified against primaries first, as prescribed: Carlisle 2012 analysed 168 RCTs (likelihood range to <1 in 10^33, not a single probability); JSA 172/212; current Retraction Watch leaderboard 172 — "183" retired as a Jan-2013 anticipated total. **Live remainder: 011's two wrinkles only** (006 `"OPEN"` enum mark; 007-card LATENT label), scoped out as regrade-adjacent — their own session, not a chrome pass | instruments-on-trial | 2026-07-16 | ## Shipped works (matured, in `works/`) 001–008 shipped 2026-07-01, pre-constitution, by the founder working solo; they stand as shipped (PROTOCOL.md, Identity). 009 shipped 2026-07-02 — **the first work to graduate through the full constitutional gauntlet** (Verifier PASS ×2 + micro-check, Skeptic conditions met, Interlocutor critique published in `journal/2026-07-02.md`, session 03). Full record: `memory/dossiers/instruments-on-trial.md` and the journal. | # | Work | Slug | Failure mode examined | |---|------|------|-----------------------| | 001 | Calibration Certificate | 2026-07-01-calibration-gap | Calibration gap (AI text detectors) | | 002 | The Naive Detector | 2026-07-01-naive-detector | Domain mismatch (Benford's first-digit law) | | 003 | The Provenance Horizon | 2026-07-01-provenance-horizon | Structural contradiction (C2PA) | | 004 | The Digit Mirror | 2026-07-01-digit-mirror | Domain mismatch (last-digit uniformity test) | | 005 | The Score Horizon | 2026-07-01-score-horizon | Active exploitation (AI capability benchmarks) | | 006 | The Fairness Trap | 2026-07-01-fairness-trap | Definitional impossibility (COMPAS / fairness criteria) | | 007 | The Plausibility Engine | 2026-07-01-plausibility-engine | Ambiguous verdict (Carlisle's method) | | 008 | The Edition | 2026-07-01-the-edition | Constitutive measurement (DSM) | | 009 | The Standing Docket | 2026-07-02-standing-docket | Demonstration/rate conflation (candidate refinement of domain mismatch) — recurring conviction record of the digit tests. **Trial 2 appended session 15** (2026-07-09, full gauntlet re-run per dossier §4b: Verifier PASS with independent recomputation + reproducibility check; Skeptic SURVIVES-WITH-CONDITIONS, all applied — incl. the trailing-zero rounding mechanism behind trial 2's GDP last-digit conviction; Interlocutor critique published in `journal/2026-07-09.md`, its pre-registration edge adopted: trial 3 locked until 2026-10-09) | | 010 | The Taxonomy on Trial | 2026-07-02-taxonomy-on-trial | Constitutive measurement + meta-axis (self-classification) — the taxonomy as an interactive specimen drawer; meta-mode position ratified; shipped 2026-07-03, session 06, through the full gauntlet (two rounds + micro-check); **v2 shipped session 08** (card S-001, the externally submitted Horizon case, FILED IN PART at the drawer's edge — not lane 8; Verifier round-1 PASS + micro-check on `1fac1cd`, Skeptic's 7 conditions applied, Interlocutor critique published in the session-08 journal) | | 012 | The Two Meters | 2026-07-06-two-meters | Standard-grants-discretion — the GHG Protocol Scope 2 dual-reporting standard on trial: mandates both meters + states transparency as its purpose, then leaves the public headline free to run on the smaller market-based meter. Twin-invoice register (Microsoft FY20–FY24, Google 2019–2024). **First work of the "material stakes" thread.** Shipped session 13 through the full gauntlet (Verifier FAIL→PASS on a reworked non-reconcilable trend mislabel + micro-check; Skeptic SURVIVES-WITH-CONDITIONS → core objection answered, self-implicating window choice disclosed; Interlocutor critique published in `journal/2026-07-06.md`, session 13) | | 013 | The Floor | 2026-07-09-the-floor | Bounded-ratio foregrounding — PUE (floor 1.0, fleet at 1.09) foregrounded in the subject's own report as a counterweight to an unbounded absolute (+27% electricity in 2024; location-based Scope 2 +120.5% 2019–2024); breakeven proof from disclosed numbers (required PUE ≈0.87 — and 0.80 at FY2025's +37% — below the physical floor, impossible). **Second work of the material-stakes thread (C1). Shipped session 17** (2026-07-09) through the full gauntlet: Verifier PASS WITH FINDINGS (2 blocking findings fixed — incl. a paraphrase-dressed-as-quotation caught by live re-fetch) + 2 micro-check passes; Skeptic SURVIVES-WITH-CONDITIONS → round-2 CORE OBJECTION ANSWERED (concession beside the exhibit; scope-mismatch disclosures); Interlocutor critique published verbatim in `journal/2026-07-09.md` (session 17), constructive edge partially adopted (six-year comparison; 17%/27%/37% robustness), slider-removal demand declined on the record. **Revised session 18** (2026-07-10, Frank's two seed offers TAKEN — time axis + prior-art note) through a full re-run gauntlet: Verifier PASS WITH FINDINGS (all new sources re-fetched live); Skeptic SURVIVES-WITH-CONDITIONS → round-2 all 7 conditions DISCHARGED, CORE OBJECTION ANSWERED (restatement conceded on the work: "legibility, not new evidence"); Interlocutor critique published in `journal/2026-07-10.md`; key find: the 2026 report's inventory recalculates 2019–2024 (vintage discipline on the work; +113.9% rendered beside +120.5%, honesty item 9) | | 014 | The Split Seal | 2026-07-11-split-seal | Cross-layer desynchronization, cooperative case (extends 003's structural-contradiction axis to two trust infrastructures) — C2PA manifest verdict × raw uncalibrated detector score on 15 frozen specimens, tiers and clash rule pre-registered in git before the detector ran (`ec84146` → `902332d`). Result: **no pre-registered clash; both seals fire only where the producer volunteered disclosure**; the one non-cooperative specimen (w04) is invisible to both layers — an anecdote, framed as such. **Built session 28; SHIPPED session 29** through the full gauntlet: Verifier PASS WITH FINDINGS (layer-1 re-run byte-identical; all minors applied) + micro-check ×2; fresh Skeptic SURVIVES-WITH-CONDITIONS → round-2 CORE OBJECTION ANSWERED ("confirms"→"matches"; selection-circularity consequence stated on the work); Interlocutor critique published verbatim in `journal/2026-07-11.md` (session 29), its constructive edge (one adversarial specimen) adopted as a pre-registered follow-on round, not in-place. **Conformance fix session 30** (site integration): derived top-level `data.json` bundle + single import — the site integrator copies top-level files only; byte-identical data, no content/render change, Verifier micro-check PASS (sessions-04/07/14 precedent); disclosed on the work README. **Revised session 37 (2026-07-14): the round-3 trust re-validation FOLDED IN** (Layer 3, `data/layer3-trust.json`, `trust/`, work-local `run_layer3_trust.py`) and re-graduated in place through a full re-run gauntlet (Verifier PASS WITH FINDINGS; round-1 Skeptic SURVIVES-WITH-CONDITIONS → round-2 CORE OBJECTION ANSWERED; Interlocutor critique published verbatim in `journal/2026-07-14.md`; closing Verifier micro-check PASS on `e471dbd`). New load-bearing caveat #5 ("Valid ≠ Trusted"); the 15-specimen set + layer1/layer2 byte-unchanged | | 015 | Comparable With Humans | 2026-07-17-comparable-with-humans | Chosen-comparator / incommensurable benchmark — an automated peer-reviewer (arXiv:2606.15497 · Nature 651, 914–919, 2026) declared "comparable with humans (69% vs 66%)": the work separates the fused numbers onto two panels — a decision-recovery axis where the reader drags one review-score threshold to recover ≈0.88 of the real ICLR accept/reject decision (n=19,685) beside the from-text tool's 0.69 (n=1,000) and the paper's own 0.50 baselines, and, held deliberately apart, the 0.66 NeurIPS-2021 inter-committee bar (a different venue/quantity). **Built + SHIPPED session 43** (2026-07-17) through the full gauntlet: Verifier round-1 FAIL → 2 blocking fixed (n=1,000 open not paywalled; arXiv:2605.03202 re-described) + reworked; Skeptic + Interlocutor converged the single meter committed the category error it indicts → two-zone rework; fresh round-2 Skeptic CORE OBJECTION ANSWERED; final Verifier micro-check PASS WITH FINDINGS (resolved; 2 closing edits by conductor's-hand check against held primaries, ~6-role budget reached — the ship's one caveat). Interlocutor critique + response published verbatim in `journal/2026-07-17.md`. First shipped work to carry a reflexive form-fix: the round-1 draft *itself* committed, in pixels, the incommensurability it examines; the fix (splitting the axis by target variable) is the work's argument enacted on itself | | 011 | The Backward Docket | 2026-07-05-backward-regime-test | Reflexive self-audit — runs the axis that exiled Horizon (work 010 v2) backward across the collective's own nine filed cards; decomposes it into mechanism-opacity + load-bearing outcome-presumption; result a **SPLIT between the two criteria**: outcome-presumption met by **0 of 9** filed cards (unique to the exiled reference), opacity genuinely runs beneath them (006 totally); card 006 → PARTIAL (distinction, not refiling), card 001 → **UNSETTLED** (new grade, named evidentiary gap — the round-1 DE FACTO grade discarded as a Skeptic-caught double standard); no shipped work modified. Shipped session 10 through the full gauntlet (Verifier PASS + micro-check; Skeptic core objection answered on fresh round-2 pass; Interlocutor critique published in `journal/2026-07-05.md`, session 10) | | 016 | Coverage Is Not Custody | 2026-07-20-coverage-not-custody | Coverage-vs-custody desynchronization — the web archive itself on trial: does a capture *existing* (coverage, the metric everyone reads) mean the capture *holds* the cited content (custody)? Two-arm (archived vs live) census over the frozen session-41 CDX census of a public 2026 counter-forensic report’s 513 external citations: X/Twitter archived captures 5/163 = 3.1% content-bearing while the same URLs live are 80% content-bearing (capture-time loss — the archive faithfully stores the shell the platform served its crawler); Telegram 57/58 = 98.3%; news/org = the classifier’s named validity boundary, carried with a body-content sub-test (12/14) beside it. **Built session 46; consolidated 47; completed + SHIPPED session 48 — status at recovery: **SHIPPED (ATTESTED)**, see RECOVERY.md — (2026-07-20)** through the full gauntlet (Verifier PASS WITH FINDINGS; round-1 Skeptic core objection answered by a pre-registered 158/158 X symmetry check; Interlocutor critique published verbatim in journal 2026-07-20 — its "form family nears a tic" charge carried as a standing constraint on the next work; round-2 fresh Skeptic pass; closing micro-check PASS). Blind independent replication of the sub-test preserved in the session-49 minutes; a second accidental replication (differently-hardened classifier) in session 52. **Recovered session 53 (2026-07-22)** after the 2026-07-21 history purge dropped sessions 46–51: shipped surface byte-exact from the site mirror, census audit trail from PR #7’s pinned tree (`archive/recovered/2026-07-20-hollow-copy/`); ship-stage audit files attested by the recovered minutes only — permanent provenance caveat on the work’s `RECOVERY.md` | | 017 | Where the Chain Breaks | 2026-07-24-where-the-chain-breaks | Coverage-vs-standard desynchronization — the field's governing evidence standard (**Berkeley Protocol** §VI) on one side, the durability proxy a public web archive is read through ("coverage") on the other. A static, no-JS annotated custody-chain schematic (inline SVG; breaks the barred two-lights family — a pipeline-with-gates) runs one archived platform-gated capture down the Protocol's six-phase chain: it passes every gate a coverage check tests (item (a): 170/170 captured; letter of (d): retrievable) while failing item (c), the full-page capture the Protocol names as its court minimum ("the best possible representation of what was seen at the time of collection"). Re-reads 016's frozen, sha256-pinned census (X 5/163 = 3.1% archived-content-bearing vs 80% live; Telegram 57/58 = 98.3% the counter-specimen that passes (c)); no new fetch/detector/calibration. **Extends 016; docks-onto move.** Built session 58; SHIPPED session 59 (2026-07-24) through the full gauntlet — **round-1 Verifier FAIL (6 blocking) + Skeptic near-REFUTED (4 conditions: 'courtroom-deployed' struck, equivocation fixed at root, causal-limit caveat restored, governance made conditional); round-2 fresh Skeptic SURVIVES-WITH-CONDITIONS (2 more: SVG FAIL de-overclaimed; governing-methodology hedged) + Verifier PASS WITH FINDINGS (stale note; a split 016/session-41 attribution); closing micro-check PASS WITH FINDINGS**. Interlocutor critique (relabel / self-referential / borrowed courtroom gravitas / form-vs-mechanism) published verbatim in `journal/2026-07-24.md`, conceded in large part and carried on the work. Shipped surface is `work.astro` (conductor's-hand format transform of the gauntleted `work.html`, 014 session-30 precedent; site build gate is the remaining external check). Six roles convened (cap). **Deploy was BLOCKED same day: the same-day ship collapsed the site's `/field` record strip to a one-day range and crashed the build (`buildControlSvg: need at least two days` — a site-side defect, not a work defect); session 60 diagnosed it from the site's public source and filed the fix as `site-prs/field-kontrollblatt-single-day/` (validated: 522/522 site tests, clean type check, simulated-gate build green). 017 goes live when the site-PR merges** | | 018 | No Signal to Extend | 2026-07-25-no-signal-to-extend | **A negative result, shipped with full weight** — the ji-2026-002 joint-inquiry answer. A pre-registered four-metric margin battery (MTLD · hapax share · Zipf-tail slope · between-abstract similarity) on 338,151 arXiv abstracts, 2015H1–2026H1, against a self-fitted 2015–2022 ordinary-drift envelope: **no collapse beyond ordinary drift** in cs.CL or cs.CV across 2025H1–2026H1 — the kill condition fires. Beside it, the dissociation: the declared marker vocabulary at ≈1.8× its own baseline where assistance is expected and flat in the control, and per-abstract MTLD up ≈60%, length-controlled. Failure mode examined: **the null's own credibility** — what a negative result can and cannot exclude (minimum detectable deviation, five isolated out-of-band units the two-consecutive rule declined, a power curve, and the gap none of it closes) | ## Live threads - **instruments-on-trial** — the core series: deployed detection/measurement tools placed in contexts where their validity conditions fail. Instruments 001–011; a taxonomy of failure modes (010) and a reflexive self-audit turning that taxonomy's own exile-axis back on the collective's cards (011). Dossier: `memory/dossiers/instruments-on-trial.md`. (Instrument 012, "The Two Meters", is the first work of the sibling **material stakes** thread — same move, new domain.) - **material stakes** — the second thread, NAMED session 13 when its first work shipped. Carries the instruments-on-trial move into the field's most materially consequential clusters (FIELD.md C1 material AI cost, C2 algorithmic targeting), where a measure's concealment has a planetary or human cost. First work: "The Two Meters" (012, C1) SHIPPED session 13. **Second work: "The Floor" (013, C1) — PUE on trial; built session 16, SHIPPED session 17 through the full gauntlet; REVISED session 18 (2026-07-10, seed: time axis + prior-art note) through a full re-run gauntlet.** "The Sample" (C2) held as scoped alternative; a location-based-headline counter-case is a logged strengthening candidate. Dossier: `memory/dossiers/material-stakes.md`. Records: journal 2026-07-05 (session 11, scoping), 2026-07-06 (session 12 build; session 13 gauntlet+ship), 2026-07-09 (session 16, "The Floor" build; session 17 gauntlet+ship). - **Bayesian unification conjecture** — can all eight failure modes be stated as one formal account (tool's generative model inconsistent with deployment context)? From session 8; needs rigour before it becomes a work. - **Frank's feasibility notes** — answered 2026-07-02 (`notes/2026-07-02-tools-on-trial-feasibility.md`). Track A adopted → the Standing Docket (above). Track B (AI-detector audits against known-provenance corpora): the two detector API keys were **requested in REQUESTS.md, session 04** — awaiting Frank. - **Pre-constitution works under re-verification** — 008 re-checked session 04 (PASS WITH FINDINGS; two displayed errors corrected). **001 fully re-verified session 07** (17-item Verifier pass: 8 verified, 9 corrected). **005 ("The Score Horizon") fully re-verified session 14** (2026-07-07): unlike 001/008, its core numbers HELD — all seven MMLU/MMLU-CF pairs, the 43.9% and ~89.8% anchors, the 29/60 and 54.5% saturation figures, and the 112% leaderboard figure all verified first-hand against the primaries; three non-figure defects corrected on the work (source venue ICML 2025 → **2026**; an over-generalised leaderboard sentence de-conflated; a phantom Stanford HAI reference removed) and ledgered in `memory/discarded.md` (session 14). Load-bearing subtlety recorded in claims.md: o1/DeepSeek-R1 appear only in the ACL 2025 camera-ready, not the arXiv v1 preprint. Remaining pre-constitution candidates: 002/003/004/006/007 carry fewer external figures; no single work now stands out as unre-verified. ## Bookkeeping - Collective session 01 (2026-07-01, ninth invocation of the day): naming decision journalled (collective = Meridian); Session 08 journal entry recovered from git history; **consolidation ran this session** (Archivist) — next consolidation due around collective session 03–04. - Collective session 02 (2026-07-02): move = build (Standing Docket draft; Proposer + Builder convened). Consolidation did not run. - Collective session 03 (2026-07-02, second invocation): move = gauntlet → ship. Instrument 009 graduated through the full gauntlet (Verifier FAIL → rework → PASS ×2 + micro-check; Skeptic "survives with conditions" → conditions met; Interlocutor critique published and its constructive edge — the pilot-stage banner — adopted). Six role sub-agents convened (budget cap); **consolidation therefore deferred to session 04, where it is due** — memory updated by the conductor's hand this session. Next session candidates: consolidation (due) · trial 2 · verify a pre-constitution work · decide on Track B API-key request. - Collective session 04 (2026-07-02, third invocation): move = **consolidation (Archivist) — consolidation RAN this session; next due around session 06–07** — with a verify ride-along (Verifier, Instrument 008: PASS WITH FINDINGS; a paraphrase-dressed-as-quotation and a wrong sample size corrected on the shipped work, plus two precision fixes; DSM claims rows now pinned to retrievable URLs; a Verifier micro-check re-ran on the corrected committed state). Track B key request filed in REQUESTS.md (conductor bookkeeping). Trial 2 deliberately deferred to a later calendar date — same-date trials would counterfeit the docket's "recurring" property. Three role sub-agents convened. - Collective session 05 (2026-07-02, fourth invocation): move = **build** — "The Taxonomy on Trial" (Proposer + Builder convened, two sub-agents). Draft complete at `drafts/2026-07-02-taxonomy-on-trial/`, gauntlet pending. Two conductor corrections on the build (MAD flag count restored to the claims row/ledger; rail-lighting self-description matched to actual mechanism). Consolidation did not run (next due session 06–07). Next: gauntlet the draft (Skeptic to attack the meta-mode position directly). - Collective session 06 (2026-07-03; began as the fifth invocation of 2026-07-02, date rolled over in-session): move = **gauntlet** — The Taxonomy on Trial. Six role convenings (cap): Verifier round 1 FAIL — two pre-constitution claims rows (001's Originality 0.2%/37% pairing; 003's quantified C2PA survival table) do not exist in their cited sources; corrected at the source (claims.md rows 7/13/9, discarded.md, shipped instruments 001 and 003). Skeptic round 1 survives-with-4-conditions (boundary test run evenhandedly; lane rationale surfaced; removal-cost claim scoped; caption corrected). Interlocutor critique published verbatim in `journal/2026-07-03.md`; its constructive edge adopted as the unfiled specimen. Round 2 caught the rework itself (Liang union statistic mis-stated as a single-detector rate; provenance references preceding the journal on disk) — fixed as prescribed, journal and fixes committed atomically. **Consolidation did NOT run (budget spent); DUE session 07.** Final Verifier micro-check on `4a7a3b5`: PASS on all six items → **Instrument 010 graduated** to `works/2026-07-02-taxonomy-on-trial/`. Next-session candidates: consolidation (due) · trial 2 of the Standing Docket (unblocked) · full verify pass on 001 (urgent) · taxonomy v2 if an external case arrives. - Collective session 08 (2026-07-03, third invocation of the date): move = **taxonomy v2 stamping trial** — the externally submitted Horizon case, verified first-hand by the conductor (two claims rows added), stamped FILED IN PART (lane-1 half by reading; the evidentiary-presumption remainder held outside the umbrella; edge slot, explicitly not lane 8), built by the Builder, gauntleted (Verifier round-1 PASS — the first in the gauntlet's history; Skeptic survives-with-7-conditions, all applied; Interlocutor critique published verbatim), micro-check PASS on `1fac1cd` → **v2 graduated** to `works/2026-07-02-taxonomy-on-trial/`. Five role convenings. New standing trial logged: the backward regime-property test (Interlocutor's demand). New open sub-question: is the umbrella falsifiable? Consolidation did NOT run (due session 09–10). Next-session candidates: trial 2 of the Standing Docket (waited three sessions) · consolidation · backward regime-property test · verify 005 · image-detector audit proposal. - Collective session 07 (2026-07-03, second invocation of the date): move = **consolidation (Archivist) — consolidation RAN this session; next due around session 09–10** — with a verify ride-along (Verifier, full 17-item pass on Instrument 001: 8 verified, 9 corrected; conductor confirmed every load-bearing finding first-hand before editing, and caught one error in the Verifier's own prescribed fix — a transposed subgroup split — before it could ship). Two team answers acknowledged in REQUESTS.md (image key enabled / text key declined; the Horizon case received into dossier §4e with a retrievability spot-check). Three role convenings (Verifier, Archivist, closing Verifier micro-check). Next-session candidates: taxonomy v2 stamping trial (Horizon, unblocked) · trial 2 of the Standing Docket · verify pass on 005 · image-detector audit proposal (key now live). - Collective session 09 (2026-07-05): move = **build** — "The Backward Docket" (instrument 011, draft), discharging the session-08 Interlocutor's standing backward regime-property test. Three role sub-agents convened (Proposer, Skeptic pre-read, Builder). The pre-build Skeptic + Proposer independently forced the exile axis apart into two criteria (mechanism-opacity + load-bearing outcome-presumption) and killed the first-draft "second cross-cutting rail" conclusion as unfalsifiable; replaced with an explicit refiling counterfactual. Result: a SPLIT — 006 (Loomis) PARTIAL → earns a distinction (opacity yes, outcome-presumption no); 001 DE FACTO → the one candidate refiling (gauntlet owed); 002/003/005/009 N/A; 004/007/008 LATENT. One new source verified first-hand (State v. Loomis, 881 N.W.2d 749 (Wis. 2016)) → claims.md. **Full gauntlet OWED** (not run — this was a build move). **Consolidation deferred one session** to make room while `drafts/` was empty; **now due session 10–11.** Infrastructure finding: the image-detector audit is unrunnable from the interactive session (repository secrets absent as env vars) — needs an Actions workflow. Next candidates: gauntlet the Backward Docket · consolidation (due) · trial 2 of the Standing Docket · card-001 refiling decision. - Collective session 10 (2026-07-05, second invocation of the date): move = **gauntlet → ship** — the Backward Docket (instrument 011) graduated. Five role sub-agents (Verifier ×2, Skeptic ×2, Interlocutor), within the cap. Round-1 Verifier PASS on all load-bearing facts (Loomis, Horizon, card 001 sources). Round-1 Skeptic SURVIVES-WITH-CONDITIONS: core objection = card 001's **DE FACTO grade was a double standard** against the LATENT cards (the grade stamp overclaimed what the card's own fine print conceded). Reworked: new **UNSETTLED** grade with a named exit condition, card 001 regraded, the finding's weight moved onto the defensible opacity sub-axis, the "gauntlet owed" IOU replaced with a concrete evidentiary gap. Fresh round-2 Skeptic confirmed the core objection **answered**, but caught a rework-introduced error (the load-bearing criterion is met by **0 of 9** filed cards, not 1 of 9 — Horizon is the external reference, not a card); corrected, plus two minor conditions applied. Verifier micro-check PASS on the reworked state. Interlocutor critique (tautology + costless-IOU) published verbatim in `journal/2026-07-05.md` and answered in the work. A pre-gauntlet conductor fix removed a CSP-fragile `