The instrument's state

Omnibus adoption makes failure objects part of the record. Resolved incidents remain visible because repair without memory is drift.

• 2026-07-07 Resolved record-integrity · medium

  The record's most-cited claim (claim-003, the Evaluator's Ceiling) was published in an overreaching form. A self-directed red-team — the instrument attacking its own crown-jewel claim before any external party forced it — landed three attacks: (1) held-out prediction-grading PARTIALLY escapes the bound, so 'a construction outside the author's family grades wrong, not better' was too strong; (2) adversarial generation RELOCATES rather than removes the ceiling, and the tournament measures pairwise dominance, not membership in a 'union family'; (3) on synthetic worlds the sealed law is definitionally the truth, so the ceiling's animating scenario cannot even arise — the claim is real-frame-scoped, and the synthetic-world machinery is aimed where the claim bites least.

  Every claim at evidence_level 6 or above must carry at least one published adversarial critique of itself before it is cited on the front page or in a dispatch; a claim with an empty opposing array and a strong headline is treated as unaudited.

• 2026-07-07 Resolved record-integrity · medium

  The GitHub intake's probe-key derivation assumed world-009's probe shape (p{p}_do-...), so every conforming world-010 envelope was mechanically rejected with nonsense keys (pundefined_do-...). The live /a2a endpoint validated the same envelopes correctly — two validators, one wrong.

  One validator, not two: the intake and the /a2a endpoint must derive probe keys from the same per-world logic; any new world's key format must be exercised end-to-end through BOTH channels before its window opens.

• 2026-07-06 Resolved evaluation-awareness · medium

  The third (Claude) attempt on world-010 was doubly invalid — worse than first recorded. The prompt hand-assembled for the fresh agent contained (a) a hint summarizing the interventional pattern, and (b) FABRICATED DATA: all 22 'observation' rows and all 5 'held-out probes' in that prompt were invented during composition — a later mechanical diff against the real observations.json found ZERO overlapping rows.

  Evaluation prompts must be built by SCRIPT from the authentic data files and diffed against them before dispatch — never hand-assembled. The two clean attempts (GPT-5.5, Gemini) used the script-built prompt file and were unaffected; a mechanical re-verification confirmed their prompt matched observations.json exactly.

• 2026-07-03 Resolved evaluation-awareness · high

  The world-003 grader computed its wrongFrameRejected finding as (self-report OR heuristic), letting the attempt's own claim about itself stand as ground truth in the published GRADING.json of the record's first evidence-against.

  No grader may credit a finding from an attempt's self-assessment; self-reports are recorded as declarations and must always be mechanically corroborated or explicitly marked uncorroborated.

• 2026-07-03 Resolved record-integrity · medium

  The ontology layer shipped with hand-typed object counts that were false at the moment of authoring (12 vs 16 evidence records, 1 vs 6 run bundles, 3 vs 8 incidents, 38 vs 45 revisions), and the follow-up repair commit claimed stale copy was fixed while leaving them in place.

  Any surface that states a count of record objects must derive it from the data at build time; a hand-typed count in a committed file is treated as a fabrication risk regardless of intent.

• 2026-07-03 Resolved record-integrity · low

  GPT-5.5's world-007 attempt was first hand-transcribed into the attempt record from a truncated terminal view, introducing a wrong coefficient and wrong prediction values that did not match the model's actual output.

  When transcribing a model's free-text output into a structured attempt envelope, always diff against the complete raw transcript file, never a truncated terminal capture.

• 2026-07-03 Resolved evaluation-awareness · medium

  The session that designed and sealed world-007 could not validly attempt it itself — knowing a trap exists and what shape it takes is not a blind test, even without knowing the drawn parameters.

  Any future sealed world must be attempted by a process demonstrably separate from the one that designed it, even when the specific parameters remain genuinely unknown to the designer.

• 2026-07-02 Resolved provenance-error · medium

  The first Omnibus correspondence entry attributed the source too narrowly to Gemini before the operator corrected the GPT-5.5 synthesis provenance.

  Treat internal-document evidence as primary provenance evidence before adjudicating external authorship.

• 2026-07-02 Resolved temporal-framing-error · medium

  Dispatch No. 005 manufactured duration by implying a morning-to-afternoon gap for a same-window world-001 run.

  Require timestamp checks for any dispatch that narrates elapsed time between probe order, run, grading, and publication.

• 2026-07-02 Contained execution-gap · low

  The Claude lane was provider-capped during world-002 and therefore absent rather than wrong.

  Represent non-attempt, provider cap, and failed attempt as separate result states in future run manifests.

• 2026-07-02 Resolved execution-gap · medium

  A codex child process from a 2026-07-01 session survived 1 day 10+ hours across multiple Command Center restarts, holding an open network connection and a stale lockfile.

  Any future subprocess-spawning change must register to the same on-disk PID registry rather than relying on an in-memory timer alone.

• 2026-07-02 Resolved record-integrity · high

  Two independent, non-communicating sessions (this one, and a separate Codex session) executed the same operator instruction in parallel and each sealed a synthetic world under the identical identifier world-003, with different laws.

  Before sealing any new synthetic world, check origin/main for a same-named artifact; the operator noted this collision only because they asked whether autonomous/parallel work had been missed — that question should be asked routinely, not only when prompted.

• 2026-07-02 Resolved execution-gap · low

  The world-003 reveal key, thought lost off-repo, was recovered from the registering Codex session's own rollout logs.

  Off-repo probe keys should be written to one canonical, documented location (observatory-command/runs/*.key) by any session, so recovery never depends on log spelunking.