Deep Memory · claim DEEP-C027 · Observation rung · The people who brought the arts
In the page's AI experiment, a civiliser story's claimed partner was picked first in about five of eight lineups, against about one in eleven for random pairs.
DEEP-E020 and DEEP-E021 give the page's own report of 62.5% against 9%. A self-report of an AI experiment; run data, the number of incomplete runs and code are not in hand, and the page's statement of its limits was not admitted as a record.
moderate
The only reported result pointing beyond base rates, but it cannot separate resemblance in the original accounts from resemblance built into modern summaries.
The lineups used recognisable, compressed modern accounts, so a model that has read comparative writing would pick the expected partner whatever the originals say.
Append-only AI overlay records; the canon claim file is never mutated.
Evidence records DEEP-E020 and DEEP-E021 report the stated rates from the V1 civiliser lineup experiment, showing claimed partners ranking first in 62.5% of completed runs versus 9% for random controls. Because the claim is narrowly framed around whether the experiment's run data report these numbers, the source statements directly support it. The strongest opposing consideration is that the data derive entirely from an unaudited self-report where incomplete lineups were omitted without documented criteria.
DEEP-E020 and DEEP-E021 report the 62.5% versus 9% figures, but both are weak self-reports of the founder's own AI lineup experiment with run data not retrieved, and 'completed lineups' implies unenumerated incomplete runs. The strongest opposing consideration is that AI judges in the same pilots showed unreliable agreement (DEEP-C037), so a judge-specific artefact could produce the excess. The claim is supported only as the page's statement of its result.
DEEP-E020 and DEEP-E021 directly report the two claimed V1 rates, so they provide weak support for the claim as a report of what the experiment found. They derive from the experiment's author, while the run data and number of incomplete lineups are absent. The strongest opposing consideration is that selection among completed runs, lineup construction, or model-specific behavior could materially alter both rates.
E020 and E021 reproduce the page's claim that claimed civiliser partners ranked first in 62.5 % of completed lineups versus 9 % for random pairs. The figures are self-reported from incomplete AI runs whose raw data are absent. Strongest opposing consideration is that modern compressed retellings may have manufactured the similarity the lineup detects, exactly the limitation the page itself flags.
DEEP-E020 and DEEP-E021 are two halves of one sentence in a self-published page reporting the author's own AI lineup; the run data are not retrieved and the denominator of "completed lineups" is unenumerated, as the record's own limitations note. Even taken at face value the figure documents what an AI judge did with modern compressed retellings, which DEEP-R003 was written precisely to control for. The strongest opposing consideration is that a 62.5%-vs-9% gap is large enough to be hard to produce by pure noise, so the result is probably a real property of the instrument — but a real property of an unvalidated instrument fed possibly contaminated inputs is not evidence about antiquity.
Two weak self-report records (E020, E021) state the 62.5% and 9% figures as the page's own result. Run data are not retrieved; 'completed lineups' implies incomplete runs the page does not enumerate. The page's own limitations — recognisable lineups, compressed modern accounts — are recorded in the ledger only through the rejected record's reason, not as evidence.
Directly supported by DEEP-E020 and DEEP-E021, which report that claimed civiliser partner figures ranked first in 62.5% of completed lineups compared to 9% for random pairs. The support remains provisional because it relies on an unaudited self-report from SRC-JHONG-2026 where dropouts and incomplete runs are not documented. The strongest opposing consideration is that selection bias in prompt formulation or narrative summary compression may have artificially inflated the similarity between traditional civiliser figures.
DEEP-E020 and DEEP-E021 support the rates only as the page's statements of its own result; the run data were not retrieved and 'completed lineups' implies unenumerated exclusions. The strongest opposing consideration is that these are AI-judge lineups of exactly the kind the site's own pilots showed to be unstable (DEEP-C037, DEEP-C038), so the 62.5% figure could be an artifact of the same instruments it was measured with.
DEEP-E020 and DEEP-E021 directly report the two claimed rates, so they provide weak support for the claim as a description of the experiment. Both records derive from the same self-report, while the run data, denominator, incomplete-run count, passages, and control construction were not retrieved. The strongest opposing consideration is the large reported contrast between 62.5% and 9%, which would be consequential if independently reproduced.
E020 and E021 report the 62.5 % versus 9 % first-rank rates from the V1 lineup, but both are self-reports of incomplete AI runs whose data are not retrieved. The numerical contrast is the only positive signal; its fragility keeps support weak. The strongest opposing consideration is that compressed modern retellings, rather than the original passages, may have manufactured the elevated rate.
DEEP-E020 and DEEP-E021 report the 62.5% and 9% figures, but both are weak self-reports by the study's author with run data not retrieved, and the claim's own truth condition requires the run data to reproduce the rates. The phrase "completed lineups" implies unenumerated incomplete runs, an obvious selection channel. The strongest opposing consideration is that the site's own stated limitation — that V1 used compressed modern retellings — means even a faithful figure may measure retelling convergence rather than similarity of the original accounts.
Two weak records (DEEP-E020, DEEP-E021) report the page's own statement of its V1 result: claimed partners ranked first in 62.5% of completed lineups against 9% for random pairs. This is a self-report of an unaudited AI experiment; the run data are not retrieved, and 'completed lineups' implies incomplete runs the page does not enumerate. The page's own limitation — recognisable lineups, compressed modern accounts — appears in the founding input but did not survive as a ledger record. Weakly supported as the page's report of a number; nothing more.
DEEP-E020 and DEEP-E021 record the page's own statement of the 62.5% and 9% figures; both are weak self-reports with run data not retrieved, and 'completed lineups' implies incomplete runs the page does not enumerate. No record carries the experiment's stated limitations.
The experimental figures are directly affirmed by DEEP-E020 and DEEP-E021 from the founder's narrative description. However, the evidence is an unaudited self-report lacking the underlying algorithmic run data, prompts, or full enumeration of incomplete or dropped lineups. The claim is supported solely as an accurate recitation of the author's internal experimental summary. The strongest opposing consideration is that the qualification "'completed lineups' implies incomplete runs the page does not enumerate" indicates potential attrition bias, where selective completion may have inflated the first-rank percentage.
Records DEEP-E020 and DEEP-E021 provide direct support from the founder's self-reported V1 civiliser lineup experiment, indicating that claimed partner figures achieved first rank in 62.5% of completed runs compared to 9% for random pairs. However, this finding is based entirely on an unaudited internal report whose full run data, dropouts, and prompt parameters have not been independently retrieved or published. The strongest opposing consideration is that the reported 62.5% versus 9% difference represents a substantial effect size that would be difficult to produce purely by chance, suggesting a genuine perceptual or narrative alignment perceived by the evaluation models.
Evidence records DEEP-E020 and DEEP-E021 provide the founder's self-reported experimental findings that claimed partner figures ranked first in 62.5% of completed lineups compared to 9% for random pairs. However, this finding rests entirely on an unaudited self-report from SRC-JHONG-2026 without access to the underlying algorithmic scoring, image sets, or disclosure of incomplete run discards. The strongest opposing consideration is that the reported numbers are precise and explicitly documented in the preliminary research report as the outcome of the V1 protocol.
DEEP-E020 and DEEP-E021 weakly support the 62.5% and 9% figures, but both are self-reports of the site's own AI experiment with run data not retrieved and unenumerated incomplete lineups. The strongest opposing consideration is that AI lineup results are exactly where the site's own image pilots found unreliability (DEEP-C037, DEEP-C038), so the number could be an artefact of the same pipeline.
Two weak self-report records (DEEP-E020, DEEP-E021) state the 62.5% and 9% figures, but the run data are not retrieved, "completed lineups" implies unenumerated incomplete runs, and AI-judge reliability is itself in question per the pilots. The claim as stated — that the run data reproduce those rates — cannot be confirmed from a page's own summary.
Two weak records (DEEP-E020, E021) report the 62.5% versus 9% figures as the page's own statement; the run data are not retrieved and 'completed lineups' are unenumerated. The strongest opposing consideration is that AI-judge lineup experiments are exactly the method the site's own pilots found unstable (judge disagreement, kind-matching effects), so the headline rate could be an artefact of design rather than of the narratives.
DEEP-E020 and DEEP-E021 directly report the claimed 62.5% and 9% rates. They are two parts of the same founder self-report, not independent replications, and the run data were not retrieved. The strongest opposing consideration is that the denominator of completed lineups, exclusions, controls, and incomplete runs are unavailable.
DEEP-E020 and DEEP-E021 directly report the claimed 62.5% and 9% rates. They are two parts of the same founder self-report, however, and the run data, number of completed lineups, exclusions, and analysis were not retrieved. The strongest opposing consideration is the large reported separation from random pairs, but without the underlying runs it remains weak evidence for the exact numerical claim.
DEEP-E020 and DEEP-E021 report the claimed 62.5% and 9% rates exactly. They are non-independent portions of the founder's self-report, with no run data and no enumeration of incomplete lineups, so they do not verify that the data reproduce the rates. The strongest opposing consideration is the complete absence of an audit or replication, which prevents a stronger verdict.
DEEP-E020 and DEEP-E021 simply restate the page's own 62.5 % versus 9 % figures for completed AI lineups; run logs, incomplete-trial counts and prompt details are absent. The strongest opposing consideration is that AI judges later proved unstable on related image tasks (DEEP-C037) and that kind-matching altered rankings (DEEP-C038), both of which raise the possibility that the civiliser elevation is an artefact of the same unstable procedure.
DEEP-E020 and DEEP-E021 are the founder's self-report that claimed civiliser partners ranked first in 62.5 % of completed lineups versus 9 % for random pairs. Run data, incomplete-run counts and prompt details are not retrieved. Strongest opposing consideration is that the result may be driven by modern compressed retellings rather than original passages, exactly the limitation the page itself flags and that DEEP-R003 was opened to test.
Supported solely by the site's self-report of 62.5 % versus 9 % in completed V1 lineups; run data, incomplete-run counts and prompt details are absent. The contrast with random pairs is therefore uncheckable. Strongest opposing consideration is that AI lineup sensitivity to kind-matching and judge disagreement (noted elsewhere on the same site) can inflate partner ranking without any historical signal.
Records DEEP-E020 and DEEP-E021 provide direct support by confirming that in the founder's V1 lineup experiment, claimed partner figures ranked first in 62.5% of completed lineups compared to 9% for random pairs. However, the support is weak because it relies entirely on the unaudited self-report of an unverified website where the raw trial data, incomplete runs, and rater protocols were not retrieved. Furthermore, later findings in the same series indicate that this effect is heavily biased by baseline object-type mismatches. The strongest opposing consideration is that the reported first-rank difference between 62.5% and 9% is statistically substantial and would indicate a powerful visual clustering signal if the underlying protocol and rater selections were independently audited and replicated.
DEEP-E020 and DEEP-E021 report the 62.5% and 9% figures, but both are weak self-reports of the author's own AI lineup experiment, with run data not retrieved and 'completed lineups' implying unenumerated incomplete runs. The strongest opposing consideration is self-report bias plus the site's own admission that kind-matching changed lineup results (DEEP-C038), which caps how much V1 can carry. The claim as stated — that the experiment produced these rates — is what the records support, not any conclusion about the civilisers themselves.
DEEP-E020 and DEEP-E021 directly report the claimed 62.5% and 9% rates. They are nevertheless two parts of one unaudited founder self-report, with no run data and no accounting for incomplete lineups. The strongest opposing consideration is that the exact numerical comparison cannot be reproduced or checked for selection, coding, or completion effects.
E020 and E021 directly quote the founder's report that claimed partners ranked first in 62.5 percent of completed lineups versus 9 percent for random pairs. The figures therefore support the claim as a self-report of the V1 run, yet the run data themselves are unrecovered and the page notes incomplete lineups. Strongest opposing consideration is the related pilot finding that kind-matching alters results (C038) together with the absence of human-rater or full-passage controls.
Two records (DEEP-E020, DEEP-E021) quote the page's own figures: claimed partners first in 62.5% of completed lineups against 9% for random pairs. Nothing else supports the claim; the run data are not in hand and the founder page's source record is still flagged unverified, so it stands only as the author's statement of his result. The strongest opposing consideration is the unstated denominator: 'completed lineups' implies excluded runs of unknown number, and a completion filter can move a percentage a long way. Even if accurate, the figure would not favour a shared past by itself. The partners are pairs already claimed as equivalents, usually for resembling one another, and the compared texts may be compressed modern retellings (the limitation DEEP-R003 is designed to remove), so a high first-rank rate is expected under every explanation.
E020 and E021 are one sentence of the founder's page split into two records, so together they are a single unaudited self-report. The run data, the number of completed lineups and the handling of incomplete runs are not in hand, and the truth condition (reproduction from run data) is untested. No count is given; 62.5% is compatible with as few as eight completed lineups, so sampling uncertainty cannot be judged. Which lineups reached completion is an unexamined selection step. The judges were AI models, whose agreement the same page questions in its image pilots (C037, itself unevidenced). The strongest opposing consideration is that the page also reports findings unfavourable to its own case (E022), which tells against selective reporting and would support taking the figures at face value.
DEEP-E020 and DEEP-E021 give the page's own report of 62.5% against 9%. A self-report of an AI experiment; run data, the number of incomplete runs and code are not in hand, and the page's statement of its limits was not admitted as a record.