In the Deep Memory image pilots, AI judges disagreed strongly with one another about the resemblance between the same objects. True if the pilots' inter-judge agreement statistics are low.

Evidence record DEEP-E022 was associated with this claim, but as noted by the secondary reader, the quoted evidence focuses strictly on kind-matching controls and does not provide agreement statistics or inter-judge reliability metrics. Because no specific evidence record in the ledger details the AI judges' rate of disagreement, the claim cannot be validated. The strongest opposing consideration is the founder's explicit textual comment noting strong disagreement between automated evaluators.

The one evidence record touching the pilots (DEEP-E022) was explicitly flagged by a second reader as not bearing on inter-judge disagreement, and no other record reports agreement statistics. The quoted phrase about 'unreliable feature coding and strong disagreement between judges' exists only in the claim's sourceAnchor. Given that DEEP-C027's lineup result depends on AI judges, this claim matters for the case, but the ledger cannot presently confirm it.

DEEP-E022 lists this claim but its quoted content concerns kind-matched controls, not disagreement between judges. No inter-judge statistic or direct report of strong disagreement appears in any supplied evidence record. The strongest opposing consideration is the evidence record's broad title and claim tagging, but its actual source statement cannot support the claimed disagreement.

E022 is the sole candidate record but, as the second reader notes, addresses kind-matching rather than inter-judge agreement statistics; no other evidence reports disagreement metrics. The claim therefore lacks support. Strongest opposing consideration is the page's own qualitative statement of “strong disagreement,” which remains unquantified and unaudited.

The only record listing this claim, DEEP-E022, quotes a sentence about kind-matching that says nothing about inter-judge agreement, and the second reader flags exactly this mismatch; the claim therefore stands on the source anchor's phrase "strong disagreement between judges" with no agreement statistics in hand, while the claim's truth condition demands them. Background expectation that model judges show poor agreement on open-ended iconographic similarity supports plausibility, and I flag that as a prior. The strongest opposing consideration is that "strong disagreement" is a qualitative characterisation by the experiment's own author, who has an interest in appearing self-critical, and the direction of any bias in such self-reporting is not obvious.

No evidence record bears on it: E022 concerns kind-matching, and its second reader notes it says nothing about inter-judge agreement. The page states strong disagreement between judges without statistics in the ledger.

The ledger contains no evidence record reporting inter-judge agreement statistics, variance measures, or scoring distributions across AI models in the image pilots. While the claim appears in the narrative text of SRC-JHONG-2026, DEEP-E022 explicitly does not bear on inter-judge disagreement, leaving the claim unsupported by evidence records. Evaluated on the ledger's lack of data. The strongest opposing consideration is that vision-language models evaluated with standardized prompts might exhibit high concordance on salient geometric features.

The only image-pilot evidence record (DEEP-E022) explicitly does not bear on judge agreement, as its own second-reader note states; the 'strong disagreement between judges' language sits in the claim's anchor quote, which is hypothesis material. The strongest opposing consideration is that if the disagreement statistics exist in the pilot data, this would be an easy claim to confirm on retrieval — but it is unattested here, and the unresolved verdict reflects that gap, not disbelief.

No evidence record provides inter-judge agreement statistics or even a direct statement of strong disagreement for this claim. DEEP-E022 concerns kind-matched controls and cannot be repurposed as evidence about agreement among judges. The strongest opposing consideration is that the page may report disagreement elsewhere, but only entered evidence records can carry weight in this assessment.

The page states strong inter-judge disagreement among AI systems, yet no evidence record supplies the actual agreement coefficients or confusion matrices. The claim is therefore only weakly anchored. The strongest opposing consideration is that prompt standardisation or different model families might raise agreement substantially.

No evidence record substantiates this: DEEP-E022 is the only pilot record and its second reader explicitly notes the quoted line says nothing about inter-judge agreement, so the claim's truth condition — low agreement statistics — is untouched by the ledger. General priors about the instability of vision-model similarity judgements make the claim credible, which I flag as prior-driven. Opposing consideration: the claim is load-bearing in the deflationary direction, so leaving it unresolved is not a neutral outcome; it means the case cannot yet rely on AI-judge unreliability either.

No record bears on inter-judge agreement; DEEP-E022's quote concerns kind-matching, not judges. The page reports strong disagreement without statistics.

Held on its anchor with no admitted evidence record; no inter-judge agreement statistic is in the ledger.

The truth condition requires demonstrable low inter-rater agreement statistics among AI models in the Deep Memory image pilots. In the intake audit of record DEEP-E022, the second reader explicitly noted that the retrieved text concerning kind-matched lineups does not report inter-judge agreement statistics, and that no supporting passage for DEEP-C037 was extracted. In the complete absence of evidence records containing inter-model agreement metrics or reliability coefficients, the claim is unresolved within the ledger. The strongest opposing consideration is that the founder's narrative text explicitly refers to "strong disagreement between judges," indicating that the underlying pilot experiments did generate such observations even though the specific numerical records were omitted from the ledger.

The ledger contains no active evidence record documenting the inter-judge agreement metrics for the AI vision pilots; the intake verification explicitly noted that DEEP-E022 did not substantiate DEEP-C037's specific proposition regarding inter-judge disagreement statistics. Because the truth condition depends on verifying low agreement statistics that are entirely absent from the ledger, the claim cannot be confirmed. The strongest opposing consideration is that the author's narrative text explicitly laments 'unreliable feature coding and strong disagreement between judges' as a primary rationale for qualifying the image pilot results.

Following reader audits during evidence intake, the link between DEEP-E022 and DEEP-C037 was severed because the quoted source text concerned kind-matched object lineups rather than inter-judge agreement statistics. As a result, the ledger contains zero evidence records evaluating whether AI judges disagreed strongly or exhibited low agreement statistics. The strongest opposing consideration is that the founder's page explicitly mentions "unreliable feature coding and strong disagreement between judges," but without retrieved data or an admitted evidence record, the claim remains unassessed.

No evidence record supports this claim: the only record originally attached to it (DEEP-E022) was explicitly re-scoped by its second reader as bearing on DEEP-C038 and not on judge agreement. The anchor quote mentions "strong disagreement between judges", but no agreement statistics were extracted, so the claim's truth condition (low inter-judge agreement) is unevidenced in the packet.

The only record mentioning judge disagreement (DEEP-E022) carries an explicit second-reader finding that its quote "says nothing about judge agreement statistics" and does not bear on this claim; no other record addresses inter-judge agreement. The claim's truth condition — low agreement statistics in the pilots — is therefore undischarged in this ledger.

No evidence record bears on inter-judge agreement: the one record near it (E022) was explicitly corrected by the second reader as not bearing on C037, and no agreement statistics exist in the ledger. The claim matters greatly for the case — unreliable AI judges would undercut the resemblance method — but the strongest opposing consideration is that its truth condition demands low agreement statistics that were simply never recorded.

No evidence record supplies inter-judge agreement statistics or even a direct evidential statement of strong disagreement. DEEP-E022 concerns kind-matched controls and explicitly does not establish judge disagreement. The strongest opposing consideration is the founder's stated pilot conclusion, but without the judges' outputs or an agreement measure it cannot be assessed.

No evidence record supplies inter-judge agreement statistics or even a direct statement specifically documenting strong disagreement between the judges. DEEP-E022 concerns the effect of kind-matched controls and was explicitly identified as not bearing on this claim. The strongest consideration in favor is the broader report of methodological instability in the image pilots, but that does not establish low inter-judge agreement.

No evidence record supplies inter-judge agreement statistics or even a direct admissible report of strong disagreement among the AI judges. DEEP-E022 concerns kind-matched controls, not agreement, and its own reader note expressly distinguishes the two issues. The strongest consideration favoring the claim is that the broader pilot description alludes to unreliable judging, but the required agreement evidence is absent.

The founder page asserts strong inter-judge disagreement, yet no quantitative agreement statistic or raw score matrix is present in the evidence records (DEEP-E022 explicitly does not bear on this claim). Without the pilot data the low-agreement claim cannot be verified; the strongest opposing consideration is that a different prompting regime might have produced acceptable concordance.

The founder page reports unreliable feature coding and strong disagreement between AI judges on the same objects; no inter-rater statistics or raw judge outputs appear in the ledger. Strongest opposing consideration is identical to C038: the result cautions against using the pilots as measurement but does not itself quantify human-perceived resemblance.

No evidence record quotes inter-judge agreement statistics or supplies the raw pilot scores; a neighbouring record notes strong disagreement only in passing and was explicitly severed from this claim by a prior reader. The numeric claim therefore lacks support in the packet. Strongest opposing consideration is that low agreement is the expected null for subjective visual resemblance tasks and would not itself diagnose the objects.

The claim asserts that in the Deep Memory image pilots, AI judges disagreed strongly with one another when scoring the visual resemblance between objects. During intake verification for DEEP-E022, the second reader explicitly noted that the retrieved text concerning kind-matching did not report inter-judge agreement statistics, and the reference to DEEP-C037 was consequently removed. As a result, no evidence record in the ledger provides inter-rater reliability scores, kappa statistics, or variance measurements across the AI models. The claim is therefore unresolved. The strongest opposing consideration is that deep learning vision models frequently exhibit substantial variance and perceptual divergence when task-prompted with abstract, out-of-distribution artistic iconography, making the author's reported pilot difficulty entirely plausible.

No evidence record bears on inter-judge agreement statistics; the only adjacent record (DEEP-E022) was explicitly corrected so as not to cover this claim, and the quoted passage says nothing about judge agreement. The phrase 'strong disagreement between judges' sits in the claim's anchor but was never entered as evidence. The strongest opposing consideration is that the site plainly asserts the finding, so retrieval of the pilot's agreement statistics would likely decide it quickly.

No evidence record supplies inter-judge agreement statistics or directly reports strong disagreement for this claim. DEEP-E022 concerns matched object kinds, and its reader note explicitly states that the quoted passage does not bear on judge agreement. The strongest consideration against the unresolved verdict is the founder page's allegation of strong disagreement, which remains unsupported by an admitted result record.

The claim that AI judges disagreed strongly with one another is asserted by the founder page, yet no evidence record extracts inter-judge agreement statistics or raw scores. Related pilot notes mention unreliable feature coding but do not quantify disagreement for this claim. Strongest opposing consideration is the complete absence of the numerical results needed to confirm “strong” disagreement.

No evidence record currently bears on this claim. The only image-pilot record in the ledger (DEEP-E022) concerns kind-matched lineups, and a second reader rejected its bearing on judge agreement. No inter-judge agreement statistic is in hand, so the truth condition, low agreement statistics, cannot be checked. The strongest opposing consideration is that the page's anchor reports "strong disagreement between judges", and an author has little motive to invent a negative result about his own method. That sentence is the claim's source rather than an evidence record, however, and 'strong' is undefined without a statistic.

No evidence record bears on this claim. The only pilot record (E022) was read by the second reader as saying nothing about agreement between judges. No agreement statistics are in the ledger, so the truth condition cannot be checked, and the claim rests on its anchor in the founder's page. The strongest opposing consideration is that the page reports this plainly and against its own interest, so once entered as evidence it would at least weakly support the claim. Its status matters because low agreement between AI judges would undercut the AI-judged civiliser lineup figures (C027).

Held on its anchor; DEEP-E022 is attached only to the kind-matching claim, and no record reports agreement statistics.