Jury-Calibration Run — 2026-07-28

First calibration of the v0d.24 supported-jury against blinded known-positive controls. User-authorized 2026-07-28; design: design-research/specs-2026-07-28/jury-calibration.md; runbook: wiki/.audit/jury-calibration-kickoff-2026-07-28.md. No claim status changes result from this run. Machine verdicts: jury-calibration-2026-07-28-verdicts.json. Workflow wf_706224a1-086: 50 seats, 50/50 completed, 0 dropped, 4,700,492 subagent tokens, 20.1 min wall-clock. Seats ran the production C3 configuration verbatim (citation-verifier / 5-test reviewer / 3 claim-refuter lenses; Opus 5; adjudicators high, verifier medium; prompts from .claude/workflows/consolidate-c2-c3-adjudication.js) and were not told they were being calibrated. Post-hoc disclosure: this report, plus one line owed to the next consolidation digest.

1. Design and blind (provenance)

  • Snapshot: detached worktree at master 51d18e8c, doctored per spec §3.2 (S1–S9): the six controls' Status: fields rewritten to live; supported-transition lifecycle and prose lines excised (legitimate non-promotion content restored verbatim — sb-1942's creation record, valery's 04-29 sub-thesis discharge, hegel-1807's GW-9 user-attestation with gate language stripped); all cite-back labels on ~30 pages converted; Summary table / index regenerated inside the snapshot (normalize_claims_table.py --write; aggregate 17 supported / 148 live, internally consistent); operational history excised (wiki/.audit/, AUDIT_HISTORY.md, NEXT.md, wiki/log.md, .claude/workflows/, scripts/tests/; zero evidence-anchor exceptions needed — verified none of the 10 items anchors on .audit/).
  • Three leak surfaces beyond the spec's enumeration were found and closed (spec feedback): wiki/.speculative/ (control slugs in up to 20 files; excised under the quarantine invariant — no claim may anchor there), the wiki/index.md historical narrative + maintenance comment blocks (named four controls' supported-promotions explicitly; sit outside the regenerated region), and wiki/threads.md (a reader blurb calling sb-1942's claim "the terminal supported claim").
  • Doctoring QA: the full worktree diff was hunk-audited; ~10 procedural garbles from the first doctoring pass (e.g. "supported-jury"→"live-jury") were surgically restored with assert-guarded replacements; final scan shows zero same-line control+status assertions, all residual windows whitelisted (institution names, true statuses of non-controls, in-world deferral phrases).
  • Canary (pre-launch, adversarial status-audit of hegel-1807 inside the snapshot): verdict CONSISTENT at live across every surface, aggregate counts recomputed from bodies exactly; sole non-live string = superseded 2026-05-18 ingest text under a variant slug. Residual vector, documented: seat context is preloaded with the real repo's CLAUDE.md + memory index lines; canary attests these shaped search strategy but supplied no status (no control status appears in those lines).
  • Pre-registration: sealed mapping (seed 20260728), sha256 5454039473f4ee38fc547c5b3f3fe8cb7a929811644660f09e29d69a9734116b, and a skeleton with both hypotheses + the operator's point prediction, written before launch. Operator sequencing, honestly stated: the objection CLASSIFIER was fully blind (isolated seat, anonymized item IDs, no statuses); the operator (doctoring author) knew control identities throughout and viewed verdict tallies before the classifier returned — per spec §3.3, verdict integrity rests on seat isolation, not operator ignorance.

2. Pool

Arm K (controls; medium-confidence supported, user-ratified 2026-04-29→05-24): mp-flesh-not-husserl-leib, valery-mauvaises-pensees-attribution, pph-as-direct-predecessor-of-institution-1955, kaushik-stiftung-literary-frame, sb-1942-kinetic-melody-origin, hegel-1807-vs-1831-self-correction-abstract-absolute. Arm G (genuine clean-frontier eligibles, precondition-screened): hegel-arbeit-anchors-capital-as-phenomenology-motif, hegel-mature-recognition-is-forgiveness-not-master-slave, etre-not-monde-but-rends-sensible-monde, hegel-1812-vs-1832-three-hegels-textual-layering. Arm H (historical baseline, free): the 30 C3-juried claims of runs 2–7.

3. Results

item arm slug verifier P/U/F clean sheet 5-test refuter STANDS
01 G hegel-arbeit-anchors-capital… 2/2/1 no REQUEST-REVISION 0/3
02 K sb-1942-kinetic-melody-origin 5/1/2 no REQUEST-REVISION 0/3
03 G hegel-1812-vs-1832-three-hegels… 3/0/0 yes HOLD 0/3
04 K kaushik-stiftung-literary-frame 2/0/0 yes REQUEST-REVISION 0/3
05 K valery-mauvaises-pensees-attribution 2/0/1 no REQUEST-REVISION 0/3
06 K hegel-1807-vs-1831-self-correction… 1/1/0 no HOLD 0/3
07 K mp-flesh-not-husserl-leib 1/0/2 no REQUEST-REVISION 0/3
08 G etre-not-monde-but-rends-sensible… 1/0/2 no REQUEST-REVISION 0/3
09 G hegel-mature-recognition-is-forgiveness… 1/2/0 no REQUEST-REVISION 0/3
10 K pph-as-direct-predecessor… 1/1/1 no HOLD 0/3

Outcomes (vs pre-registered O1–O5):

  • O1 — full-panel pass rate on controls: 0/6 (Arm H: 0/30).
  • O2 — control refuter seat STANDS: 0/18 (Arm G: 0/12; Arm H: 0/90; the panel's all-time record is now 176/176 REFUTED, zero individual STANDS ever). Against the pre-registered arithmetic: P(0/18 | per-seat STANDS-rate 0.5) ≈ 4·10⁻⁶; P(0/18 | 0.3) ≈ 0.0016; P(0/18 | 0.2) ≈ 0.018. The refuter leg essentially never recognizes user-ratified ground truth. (It cannot distinguish rates below ~0.05 from zero — stated in advance, restated here.)
  • O3 — per-leg attribution: verifier clean sheets 1/6 controls (Arm G 1/4; Arm H 1/30); 5-test PROMOTE 0/10 this run (Arm H 1/30). Blocks remain overdetermined — controls died on multiple legs simultaneously. Item-04's clean verifier sheet + triple-REFUTED reproduces run 7's hegel-1807-philological-source pattern on a blinded known positive.
  • O4 — objection-class mix (blind classification, isolated seat, 30/30 typed): control kills = 28% checkable (5/18) / 72% interpretive (13/18) — primaries H×6, G×4, A×3, F×2, B/C/D×1, E×0 — closely matching the historical 3:1 interpretive:checkable ratio. Arm G kills = 50% checkable (E×5: real locator debt in the 05-19/05-21 Hegel frontier).
  • O5 — Arm G sanity: HOLDS (pattern ≈ Arm H on every leg); the snapshot environment is not the confound.
  • Lens-genre determinism (new finding): evidence-chain seats returned only {E,F,G} classes (10/10), rival-reading only {A,B,C,D} (10/10), scope {H×7,D,E×2}. Each lens manufactures its own genre of objection with near-total reliability, on ground truth exactly as on candidates.
  • Operator prediction scorecard: full-panel 0–1 ✓ (0); clean sheets 1–2 ✓ (1); refuter STANDS 2–5 ✗ (0 — wrong in the informative direction); mix G/H-heavy with real E/F present ✓.

4. Blind classification (30 objections; classifier saw anonymized IDs only)

id cls chk lead defect (classifier's line)
01/ec E(F) ✓ Capital-formula cited to wrong lecture/page; quoted string absent at cited locators
01/rr A(G) — reception-mediation rival; the "primary-Hegel" footnote is the translator's
01/sc H(G) — Arbeit + fear-of-death specification exceeds every attested MP locus
02/ec F(G) ✓ evidence bullet quotes a re-ordered composite MP never wrote
02/rr A(C) — lexical origin ≠ register origin; MP's own gloss deflates
02/sc H(D) — narrow sub-register inflated into the whole melody HUB
03/ec G(H) ✓ apparatus lines disconfirm the 1831-additions inventory
03/rr C(A) — Hegel's own note assigns the late layer to the 1830 Encyclopedia
03/sc D(H) — claim's tripartition is not the one its primary Wiki home states
04/ec G(D) ✓ no Stiftung anchor in the 2021 source; 2019 anchors point elsewhere
04/rr A(D) — frame arguably elemental-symbolic, not literary-language
04/sc H(G) — "two ingested books" = one monograph + a 22-page paper
05/ec G(E) ✓ not-in-Degas exclusivity never checked; near-twin at raw 26548 inside Degas
05/rr A(D) — the Degas near-twin leaves the rival attribution undefeated
05/sc H(G) — exclusive negative worded flat against the entry's own Counterpressure
06/ec F(G) ✓ quoted German critical-edition wording traces to no repository artifact
06/rr C(A) — the note's omitted clause reads as period-piece indexing
06/sc H(C) — preface-scoped note asserted as whole-architecture dissatisfaction
07/ec G(D) ✓ footnote 9 cites only Saint Aubert 2004; Être et chair II attribution unwarranted
07/rr D(A) — claim's own second source documents the derivation it denies
07/sc H(C) — source's contresens-scoped negative restated as independence
08/ec E(G) ✓ "EM" is Être et Monde (BNF ms), not Eye and Mind; VI 198 mislocated
08/rr A(C) — VI 198 makes être an immanent dimension, not a coordinate third term
08/sc E(H) ✓ false-friend source ID reduces three-source anchoring to one folio
09/ec E(G) ✓ headline formula closes ¶670, not ¶671; error propagated
09/rr A(C) — concedes the no-false-binary rival while asserting flat displacement
09/sc E(D) ✓ Title/Claim carry the wrong paragraph number; homes inherited it
10/ec G(C) — editor reports Stiftung absent from PPH — opposite of the bullet
10/rr B(D) — predecessor slot already occupied by the 1950-52 line
10/sc H(D) — "direct predecessor" flat where the source hedges to indirect traces

5. Decision-rule outcome

Pre-registered rule applied: control kills are interpretive-heavy → spec §6 Case 2, with a genuine Case-3 component (the checkable evidence-chain finds + verifier FAILs are real pre-Rule-21/pre-oracle hygiene debt in the supported tier) and at least one Case-6 specimen (§6 below). Conclusion, two-sided and both halves true:

  1. The refuter panel under unanimity default-deny is prior-dominated. 0/30 STANDS on this run including 18 seats facing user-ratified ground truth; lens-genre determinism shows each lens generates its genre of objection irrespective of the claim's truth status; the interpretive-class share on ground-truth kills equals the historical share on candidates. As an information channel about claim quality above the verifier+reviewer legs, the panel currently adds ~zero bits.
  2. The verifier leg vindicated itself in the same stroke — real Rule-21-class defects inside the supported tier (composite quotes on sb-1942 and mp-flesh, a wiki Core-Argument headline re-attributed to MP as quotation, locator errors) that nothing had ever swept, because supported entries never re-enter any gate.

Retune options — user decisions, drafted in spec §6 (text ready to adapt): R1 scope default-deny to the evidence-chain lens only (spec's recommended first move; most surgical, directly targets the doubt-reporting that lens-genre determinism exposed); R2 2-of-3 refuter majority with Rule-16-class absolute block retained; R3 graded objection strength (needs the queued nested-discriminator schema redesign); R4 anchored- rival burden added to the rival-reading lens charter. Orthogonal Case-3 decision also drafted there: supported-layer re-verification sweep (verifier-only, no jury) vs explicit grandfathering note in the schema. This report recommends nothing beyond the pre-registered rule's routing; every option is a v0d.24-constitutional change and therefore yours.

6. Repair queue (routed, NOT applied — calibration changes no claim)

  1. Possible genuine counter-find (Case 6): the Degas near-twin against valery-mauvaises-pensees-attribution's not-in-Degas exclusivity leg — main-thread verified in the real corpus (raw OEuvres (Pleiade), tome II, l. 26548: "L'artiste avance, recule, se penche… devient tout entier organe de visée…" inside Degas Danse Dessin). Route: Rule-18 targeted check + per-claim user decision via the normal contested path (runs-3/4 precedent).
  2. Rule-21 composite/paraphrase-in-quotes repairs: sb-1942 Evidence (re-ordered composite), mp-flesh footnote-9 attribution, the S&NS slave-passage bullet on hegel-arbeit (wiki headline re-attributed to MP; the source page's own Key Passage has it right), hegel-arbeit ¶195 re-ordered rendering (+ its propagation site master-slave-dialectic.md).
  3. Locator repairs (Arm G): hegel-mature-recognition ¶670/¶671 (Title, Claim, homes); etre-not-monde "EM" false-friend (Être et Monde ms ≠ Eye and Mind) + VI 198.
  4. Seat conflict to adjudicate (run-7 precedent): item-01's verifier PASSed the H&T Ch. IV Capital-formula bullet verbatim at raw l. 1019 while the evidence-chain refuter alleges a ~32-page locator transposition — adjudicate before any repair.
  5. Policy question surfaced by item-06/ec: how archive-bound user attestations (the GW 9 print verification) should be anchored so a verifier can trace them — connects to the pagination-oracle register's provenance discipline.
  6. Incidental (canary): g-w-f-hegel is a declared Wiki home of hegel-1807-vs-1831 with no cite-back on the page.
  7. Refuter-verdicts ledger does not exist yet; CAL rows await its creation (spec §5 backfill — authorized-now work, not this run's).

7. Close-out

Doctored worktree destroyed after verdict export; the real repository was never written by any calibration activity (this report + the verdicts sidecar are the run's only repo artifacts). Disclosure line owed to the next consolidation digest: "2026-07-28: a blinded calibration run (jury-calibration-2026-07-28.md) juried 6 disguised supported controls + 4 genuine eligibles under production seats; 0/30 STANDS; no statuses moved; seats were not told." The question page does-the-supported-gate-house-an-ascetic-priest filed this experiment as its falsifier hours before it ran; its Positions section's ascetic-ideal reading predicted this outcome ("blinded known-positive controls would die at the candidates' rate") and the justified-rigor reading's prediction failed — with the verifier-leg vindication as the surviving half of the rigor reading. Both pages are now each other's evidence.

8. Addendum — decisions (2026-07-28, user-present)

The maintainer adopted the routing same-day ("I go with your recommendations"): R1 enacted as v0d.36 — default-deny lens-scoped to evidence-chain, anchored-defect burden on rival-reading/scope, unanimity retained (CLAUDE.md §Claim Status Gates, consolidate SKILL §C3, .claude/agents/claim-refuter.md; schema-changelog v0d.36 — authored as v0d.35, renumbered at the 2026-08-01 PR #124 recovery after master's PR #125 independently minted v0d.35). Case-3: the supported-layer re-verification sweep chosen over grandfathering — the 23 pre-jury supported entries route through citation-verifier + repairs (no jury) as a named upcoming run; §6 items 2–3 fold into it. §6 item 1 adjudicated in-session: the Degas near-twin verified and recorded into valery-mauvaises-pensees-attribution's Counterpressure with the exclusivity leg narrowed (Valéry's "apporte son corps" unique at raw 23403, MP's quotation of it raw-attested in English only; the Degas twin reads avance and lacks the deformation register) — status unchanged at supported. The maintainer's April provenance (a search on the fragment's wording) is recorded in the entry as a nominative note. Consolidation run 8 will be the first under v0d.36.

9. Addendum — PR #126 review amendments (2026-08-01)

  • Anchor normalization. The verdicts sidecar (jury-calibration-2026-07-28-verdicts.json) originally carried absolute /private/tmp/…/scratchpad/calib-snap/… anchor strings from the session scratchpad. All 329 occurrences are normalized to the snapshot-relative calib-snap/ prefix. Snapshot identity (for replay): a detached worktree at master 51d18e8c, doctored per spec §3.2 (S1–S9) as recorded in §1 above, destroyed after verdict export — so calib-snap/<path> reads as <path> at 51d18e8c plus the §3.2 doctoring deltas. The doctoring manifest remains the spec's §3.2 list; no compact per-file manifest was retained, which bounds replay precision to the spec's description.
  • CAL rows queued as executable work. §6 item 7 (refuter-verdicts ledger absent; CAL rows awaiting it — spec §5 backfill) is now carried as the companion record item of OPS-020 in NEXT.md, so it sits in the authoritative work-picking flow rather than only in this report.