Registry / findings

Findings

Only findings a published round actually reports. Most rounds so far propose experiments, so this list is short by design.

F-30.1
34 controls; 8 present, 3 partial, 23 absent; 26 critical controls partial or absent; fail-closed verdict BLOCKED. The checker validates the declared manifest, not the correctness of the human code review.
Round 30 · Reasoning & Verification
Audit observation
F-30.2
Frozen snapshots, source versions, hashes, packet spans, model traces and usage telemetry are useful foundations but do not establish root-action fidelity, separation or unresolved structural equivalence.
Round 30 · Reasoning & Verification
Audit observation
F-29.1
Three content probes identify all 36 observable response functions but root provenance in 0% of the 72 structural worlds. Provenance twins are interventionally equivalent under every admissible content probe. Exact diagnostic result, not a population estimate.
Round 29 · Reasoning & Verification
Audit observation
F-29.2
Greedy entropy and minimax splitting identify all behaviors with 2.50 mean probes. Equating strict local response with root independence is wrong in 50% of the deliberately paired worlds. Verified root rotation establishes a prospective boundary, not historical independence.
Round 29 · Reasoning & Verification
Audit observation
F-28.1
Declared-variant model intervals covered exact conditional probabilities in 100.00% of weighted histories under all three authored priors. Held-out hidden-common-origin coverage was 72.08% (uniform), 60.57% (benign), and 65.43% (issuer shift). These are exact constructed results, not production estimates.
Round 28 · Reasoning & Verification
Audit observation
F-28.2
All hidden-common-origin histories remained compatible with the declared set. Passive adequacy AUROC ranged from 0.488 to 0.577. TSAR did not systematically dominate scalar routing at matched coverage; most cells tied.
Round 28 · Reasoning & Verification
Audit observation
F-27.1
A model-set unanimity gate separates verdict safety from scheduling efficiency under fixed verification budgets: four policies using the 116-scenario gate produced no wrong categorical verdict in any of the 3,712 in-model policy-world runs, while a nominal target-entropy policy resolved 100 percent of cases but falsely emitted NO_CHANGE in 11.23 percent of changed cases in one regime.
Round 27 · Reasoning & Verification
Audit observation
F-27.2
An out-of-model common-origin stress shows that extra evidence can increase false confidence: with the hidden_common_origin variant excluded from the planning set, the exact MSAV policy's false-clean rate reached 33.68 percent among changed worlds at budget 9. Myopic adaptivity can underperform a fixed target-first schedule.
Round 27 · Reasoning & Verification
Audit observation
F-26.1
In the declared one-fault universe, SACP leaves eight opposite-target ambiguity pairs; adding one independently rooted external commitment reduces that count to zero. Under two faults, SACP plus external commitment leaves 26 pairs; adding an independently rooted behavior probe reduces that count to zero. These are exact finite-model counts, not prevalence-weighted rates.
Round 26 · Reasoning & Verification
Audit observation
F-26.2
Against three failed roots, four genuinely independent target-sensitive roots identify the target in the declared model. If the witness shares the issuer root, no available set identifies change against three failed roots. Root independence is an assumption.
Round 26 · Reasoning & Verification
Audit observation
F-25.1
On 80 authored traces, SACP alerted on 55 of 60 gap traces and none of the 20 negative controls (trace recall 0.9167, precision 1.0000, claim-weighted recall 0.8944). Revalidate-all achieved 1.0000 recall but alerted on all 20 negative controls and created 400 claim-review tasks, compared with 161 for SACP. A sequence-only policy falsely alerted on all five benign-reordering controls. These numbers show that the implementation conforms to the authored distinctions. They do not estimate production precision, recall, incident prevalence, reliability, cost, or user impact.
Round 25 · Reasoning & Verification
Audit observation
F-24.1
On 23 authored fixtures, the proposed authority-lineage selector preserved every expected claim edge with one conservative extra edge, while repository-only selectors missed every deliberately non-repository fixture. Conformance to an authored model, not evidence of production effectiveness.
Round 24 · Reasoning & Verification
Audit observation
F-23.1
The hybrid selector found all three artifacts labeled relevant in the real interval, including one uniquely added by an independently derived import view. It also selected four irrelevant artifacts and invalidated two claims that adjudication later found unaffected. Design evidence from one interval, not a general validation result.
Round 23 · Reasoning & Verification
Audit observation
F-22.1
On the actual 127-path diff, the measurement-only watchlist selected 0 of the 17 audited critical paths, the filename heuristic 12 of 17 with 25 false positives, retest-all all 127, and the authored CTAG manifest exactly the 17. The 17-path set and tags were manually authored from the same change set, so this is not discovery accuracy or compute savings.
Round 22 · Reasoning & Verification
Audit observation
F-22.2
On 32 authored mutation fixtures CTAG routed 32 of 32 exactly (direct watchlist 9, filename keyword 7, retest all 3). Specification tests on authored fixtures, not evidence of out-of-sample accuracy.
Round 22 · Reasoning & Verification
Audit observation
F-22.3
Two historical claims preserved; five current claims made stale, then restored after direct inspection; two hypotheses left untested.
Round 22 · Reasoning & Verification
Audit observation
F-21.1
Under the current three-write best-effort persistence model, seven of eight possible write states are incomplete yet do not block release; ABMG audit-qualifies none of those seven. Logical state-space result, not a measured production rate.
Round 21 · Reasoning & Verification
Audit observation
F-21.2
Ten envelope-only mutations collide under the current content-only fingerprint, 24 equal-byte different-version source pairs and 24 distinct provider-response pairs collapse under the current records, and five bad stream fixtures are falsely accepted by a development model of the current parser; the ABMG projection separates or rejects all of them. Constructed fixtures, not production results.
Round 21 · Reasoning & Verification
Audit observation
F-20.1
Endpoint and generic operational records miss evidence-specific corruption: A0 detected 2 of 12 mutation families and A1 detected 4 of 12, while content-bound receipts detected all 12 across 96 episodes with zero clean false positives. Constructed-fixture results, not model performance.
Round 20 · Reasoning & Verification
Audit observation
F-20.2
An operator-only digest chain does not survive a privileged rewrite by that operator: after simulated rewrite, operator-only arms detected 0 of 12 families. Receipts bind observable delivery under assumptions; they do not prove source truth, completeness before emission, provider-internal processing, semantic support, or causal evidence use.
Round 20 · Reasoning & Verification
Audit observation
F-19.1
Constructed collision: terminal output plus aggregate usage cannot determine whether the corrective source was delivered. A narrow indistinguishability result over constructed records, not model performance.
Round 19 · Evaluation Science
Audit observation
F-19.2
Static audit: 4 of 18 collector-contract requirements present, 3 partial, 11 absent at the required granularity.
Round 19 · Evaluation Science
Audit observation
F-18.1
Constructed example: a timestamp-only newest-recorded rule returns a late January correction (60) for a March current query whose correct fixture answer is 75. A counterexample over hand-specified records, not model performance.
Round 18 · Artificial Cognitive Architecture
Audit observation
F-18.2
Exact check: 24 temporal-query checks hold across all six arrival orders of the three-record example, plus 12 checks of baseline failure, conflict handling, dependency repair, alternative support, and source integrity.
Round 18 · Artificial Cognitive Architecture
Audit observation
F-17.1
Constructed example: a one-bit message that answers the release question perfectly leaves a 3/4 accuracy ceiling on a later test-status question under a stipulated uniform distribution. A possibility result, not model performance.
Round 17 · Collaborative Intelligence
Audit observation
F-17.2
Exact enumeration: 5, 2 and 1 of 15 partitions preserve every answer as the question family grows, needing at least 2, 3 and 4 message states. None of 60 deterministic recompressions restores a lost distinction.
Round 17 · Collaborative Intelligence
Audit observation
F-16.1
A supported fact does not establish decision sufficiency: the exact fictional policy counterexample leaves both review decisions possible.
Round 16 · Reasoning & Verification
Audit observation
F-16.2
Exact finite-world enumeration yields 3,778 determined and 3,134 underdetermined pairs among 6,912 pairs. These are checker diagnostics, not model performance.
Round 16 · Reasoning & Verification
Audit observation
F-16.3
A valid decision-changing pair establishes a gap within the modeled scope; failed bounded search does not certify completeness. Scope expansion can reopen a decision.
Round 16 · Reasoning & Verification
Audit observation
F-15.1
Behavioral influence does not establish a useful correction.
Round 15 · Collaborative Intelligence
Audit observation
F-15.2
Message value depends on the receiver's existing information and on other messages.
Round 15 · Collaborative Intelligence
Audit observation
F-15.3
A local communication benefit must still survive a complete-system comparison under measured resources.
Round 15 · Collaborative Intelligence
Audit observation
F-14.1
Exact constructed example: 97.5% snapshot-label accuracy coexists with a reversed comparison (true −2 points, judged +3 points). This is a possibility result on invented labels, not a measured Noetis failure rate.
Round 14 · Evaluation Science
Audit observation
F-14.2
Exact constructed example: a pooled confusion correction restores the overall mean to 79% while the corrected comparison stays in the wrong direction (+3.4054 points).
Round 14 · Evaluation Science
Audit observation
F-14.3
Exact constructed example: six valid identical-artifact couplings narrow whole-cohort bounds from [0, 20] to [6, 14] points without resolving either artifact's correctness.
Round 14 · Evaluation Science
Audit observation
F-13.1
Replaying the current scorer reproduced all 119 nonempty stored scores.
Round 13 · Evaluation Science
Audit observation
F-13.2
Appending the full source as a labeled appendix removed detected-issue credit in all 114 outputs that had it.
Round 13 · Evaluation Science
Audit observation
F-13.3
Removing appended approved findings changed detection counts in 44 of 59 nonempty panel outputs.
Round 13 · Evaluation Science
Audit observation
F-13.4
Saved results keep the final brief and latest check, but not before/after pairs, so a repair flag cannot distinguish a fix from a new error.
Round 13 · Evaluation Science
Audit observation
F-13.5
Usage data existed for 19 panel cases and no baseline cases.
Round 13 · Evaluation Science
Audit observation
F-12.1
Exact synthetic diagnostics show how candidate thresholds trade coverage for an upper risk bound at fixed δ = 5% and family size 3. These are worked calculations, not model performance.
Round 12 · Reasoning & Verification
Synthetic diagnostic

Not independently replicated. Findings describe the records and calculations examined, not general system performance.