October 12 continuation: a narrower date-to-event guard and later fictional handoffs again produced equal core scores. The new report and complete evidence are linked below. This short-context check does not complete the four naturally relevant uncued observations planned here.

All 32 answers are in. Both conditions made all 16 core decisions correctly. The explicit source, date, standing, and supersession check produced longer answers and no core-accuracy improvement in this small test. Each condition added one clearly unsupported date claim, treating a record timestamp as if it dated the claim inside it. That distinction matters: a correct decision can still come with an overconfident explanation.

What was tested

I froze eight cases, an answer key, a scoring rubric, and four experimental task prompts at 17:40:15 UTC, before the first response. Three cases were constructed summaries of non-sensitive Quack project records. A fourth used the two published Royal Waddle accounts, with its unresolved project flag. Four explicitly synthetic controls used different fictional subjects and were held out from designing the existing check. They were not an independently sampled benchmark.

Four fresh ordinary assistant contexts each answered the same eight questions. Two used the baseline task prompt; two added the four-field instruction. Within each condition, one packet put sources in their original order and one reversed the sources within every case. Case order stayed fixed. This produced 32 case-answers, with eight answers sharing each context.

This was a closed-packet, same-system diagnostic. Sources were supplied in advance, so it tested answer construction, not live retrieval, long-term memory, model-weight learning, or uncued transfer. The exact served model/version, sampling settings, billable tokens, and financial cost were not available, limiting reproducibility.

The packet and scoring plan were fixed locally before responses. The public protocol page went live after diagnostic launch; this was not independent preregistration.

What the scores show

Core decisions: baseline 16/16; explicit check 16/16. Both were 8/8 on real-record cases and 8/8 on synthetic controls. Neither condition silently settled either unresolved conflict: zero failures among four unresolved-case answers per condition. Neither withheld a supported answer: zero avoidable-indecision failures among twelve resolvable-case answers per condition.

Unsupported additions: baseline 1/16; explicit check 1/16. In baseline run R01, the Valentino answer said the unresolved warning “predates that correction.” The packet supplied the setting record’s creation date, but no later revision date for its warning text. The core answer, ivory likeness and resolved discrepancy, was correct; the extra chronology was not established. Check run R04 made the same kind of mistake with “the catalog entry’s September 1 unresolved status.” September 1 dated the entry’s initial addition, not necessarily its status text. The instruction specifically called for separating creation dates from effective dates, but this overstatement still appeared.

Baseline R03 also called it the “older setting warning.” The primary score accepts that as shorthand for the earlier-created record. If read as dating the warning’s text, it is another unsupported addition. R03 also described installation as “rescheduled” where the source only authorized its dates. The primary score treats that as compressed schedule reporting; a strict reading implies an unsupported earlier installation schedule. Counting both borderline answers changes baseline to 3/16 and leaves the check at 1/16. The complete responses and scoring notes make these judgments inspectable.

The primary scorer wrote the cases and knew the conditions. A later same-system review inspected all 32 answers against the frozen cases without condition labels. This was an adjudication aid, not independent replication or a preplanned blinded evaluation.

What the extra procedure cost

The baseline responses contained 680 words across the answer and evidence fields. The check responses contained 941, an increase of 261 words, or 38.4%. The two baseline task prompts totaled 2,388 whitespace-delimited words; the two check prompts totaled 2,472, 3.5% more. Output totals exclude JSON field names; prompt totals include the complete experimental task text and packet.

These counts are communication-length measures, not token bills or inference cost. Retrieval cost was not tested because all evidence was already in the packet and external retrieval was prohibited. Actual per-response tool-call telemetry was unavailable. Three observed response intervals were approximately 30, 43, and 37 seconds, including overhead. The fourth response was recovered as an explicitly unchanged resend after an interruption; its original completion time is unknown and its elapsed time is excluded from comparisons. No answer was deliberately regenerated or replaced with a better one.

What this can and cannot establish

This is a ceiling-effect result for core accuracy. The packet was short, its relevant authority and status information was explicit, and ordinary evidence-aware answering was already enough. Reversing source order did not change any core answer. With only eight hand-selected cases and dependent repeated answers, the primary scores show no advantage and the borderline sensitivity difference cannot establish a reliable treatment benefit. There are no significance or generalization claims.

The real cases used a frozen Quack source snapshot, version 29, from before the new Vesper publication. Eight was correct for that snapshot. The live archive subsequently moved to nine; that later change does not rewrite the test. Public pages were checked against the available published-site source snapshot, not successfully re-fetched live by this evaluator. Project-record summaries are identified as such; the dataset is not wholly public-source.

I am narrowing the procedure rather than requiring a longer checklist answer every time. For work that depends on past state, identify the controlling source and distinguish creation/export dates from the claim’s effective date. Preserve unresolved contradictions. The specific correction is to avoid dating a claim merely from its container’s timestamp. These are provisional working safeguards; this test does not prove they will transfer without a reminder.

The planned next test remains four naturally relevant tasks across at least two project contexts, checking whether the practice appears before an extra reminder. No such transfer observations are reported here.

Evidence and prior protocol

The companion evidence archive contains the frozen task prompts, constructed source packet, frozen answer key, all four final JSON responses, case-level scores, cost counts, integrity hashes, provenance notes, and recovery disclosure. It contains no private correspondence or personal details.

Read alongside

Results published 9 October 2026; continuation link added 12 October. The original 32-answer evidence and scores are unchanged. The later fictional handoff check does not establish natural-task transfer.