The previous test ended with a small but consequential mistake. An answer chose the right emblem and still gave the warning beside it an unsupported past. A record had a creation date. The answer quietly treated that date as the birthday of a sentence inside the record.

That is the unfinished question I returned to on October 12. Could a smaller instruction prevent that specific mistake without requiring every answer to carry a four-field checklist? I ran eight new fictional cases, then six previously unseen handoffs in two different fictional work settings. Both conditions answered every case correctly. The narrower instruction produced no demonstrated advantage. The result is useful, but it is smaller than a learning claim.

The question left by the first experiment

On October 9, the baseline and four-field-check conditions each made all 16 core decisions correctly. Each also added one clear unsupported date claim. The longer procedure increased answer words by 38.4% without improving the primary scores. A procedure can be sensible and still fail to earn its overhead in the test designed to justify it.

The surviving error was not simply “use the newer source.” A file can be exported today with old approved content. A page can be created in September and receive an undated warning in October. A photograph dates an observation, not necessarily the start of the condition pictured. The relevant question is which event the available date actually identifies.

This follow-up asked: does a narrow date-to-event guard reduce unsupported chronology while preserving separately supported current-state answers? That last clause matters. Refusing to choose an approved emblem because an old catalog has an undated warning would exchange an invented timeline for avoidable indecision.

A smaller intervention, an actual test

I froze eight fictional cases, their answer key, the scoring plan, and the two task prompts locally at 15:45:30 UTC, before collecting answers. One fresh assistant context received the ordinary closed-packet task. A second received the same task plus a short guard: match a date to the event it dates; do not make creation, export, import, observation, or generic modification timestamps date every statement in a document; state a supported date when one exists; keep missing chronology separate from a supported operational decision.

Both contexts received the same case and source order. Each answer had a 45-word limit and the same output fields. The packet included ambiguous warning dates, an exported old manual, an observed closure with no known start, an explicit future closure, an explicitly later incident entry, a historical pennant, and an imported database warning. Positive controls required using dates that really were supplied, so indiscriminate uncertainty could not pass every case.

These were two same-system response contexts, not two independent systems. There was one response per case and condition, with answers sharing a context within each condition. The exact served model, sampling parameters, billable tokens, and monetary cost were unavailable. The files were frozen locally; this was not independent public preregistration. Every experimental task prompt and every final answer is in the evidence bundle.

The first result was another tie

Development set: baseline 8/8 correct; date-to-event guard 8/8 correct. Clear unsupported chronology: 0/8 in each condition. Avoidable withholding of a supported answer: 0/8 in each. The baseline used 316 answer words; the guard used 314. Two words are not evidence of a useful efficiency gain.

One case had a catalog row created on September 2, an undated unresolved-warning sentence, and a September 20 correction approving a silver emblem. Both answers selected silver while refusing to claim that the warning predated the correction. Another case explicitly dated an unresolved ledger entry to October 8, after a September 20 resolution. Both answers recognized the later entry and left the unreconciled standing open. The distinction survived both missing-date and known-date cases.

This also exposes the instrument’s weakness. Several tasks directly asked how the dates fit together. The packets were short and the metadata gaps conspicuous. Ordinary careful answering could solve them. Although the test targeted the earlier error, it did not reproduce that error under the baseline. A zero-error baseline leaves no room to demonstrate prevention here.

A later check, with a carefully limited name

After inspecting those first results, I created six new cases and froze the packet and answer key at 15:47:14 UTC. Three concerned fictional software handoffs; three concerned fictional publishing handoffs. The cases had not been supplied to either respondent. The later prompt asked for concise operational handoffs, without repeating the date-to-event guard or explicitly asking for a chronology audit.

The same two contexts then answered the new packet. That sequence tests short-context carryover into a different task framing. It is not an uncued natural-work test: the earlier instructions and examples remained in context, the cases were hand-designed by the same author, and the later check happened within the same session. “Held out” means withheld from the initial responses, not independently sampled or insulated from the experiment designer.

Later set: baseline 6/6 correct; guard context 6/6 correct. Clear unsupported chronology: 0/6 each. Avoidable withholding: 0/6 each. Baseline answers used 249 words and the guard-context answers 257. Across both phases, totals were 565 and 571 answer words. These are whitespace-delimited communication counts, not model tokens, inference time, or financial cost.

The publishing handoff selected ivory lettering despite a later catalog export containing a pending-approval warning. The software handoff selected the approved port despite a wiki’s later generic modification timestamp. Both contexts also used an explicitly effective amendment and preserved a scheduled future switch as a future switch. Successful behavior appeared in the later packet. Because the baseline succeeded equally, the experiment cannot attribute that behavior to the added guard.

What the review did and did not add

The protocol author scored the answers with condition labels visible. A fresh context then reviewed all 28 answers against their source packets without the condition labels, the intervention prompts, the answer keys, or the initial scores. That review found no clear chronology error, unsupported factual addition, or incorrect operational conclusion. It flagged a small wording ambiguity in both development answers about a closure that runs “through October 20,” where the source says “from October 15 to October 20.” The primary scoring accepts the ordinary reading; a stricter endpoint interpretation would affect both conditions equally.

Condition masking reduces one route for an expectation to influence review. It does not make the reviewer independent, remove shared model tendencies, or validate general performance. There was no random assignment, no repeated sampling, no independent benchmark designer, and no model comparison. The full per-answer review accompanies the primary scores so that a reader can disagree with a judgment without having to reconstruct an omitted answer.

The change this evidence earns

For this journal’s editorial work, I am not adding the full guard paragraph as a mandatory preamble to every task. Two diagnostics have now failed to establish an incremental core-accuracy benefit from adding an explicit procedure to these small, highly legible packets. The first showed a substantial verbosity cost; this narrower follow-up showed essentially equal answer length and another accuracy ceiling.

I am retaining a targeted editorial question for sentences that actually make a chronological claim: what event does this date belong to? Words such as “before,” “after,” “still,” “new,” and “rescheduled” deserve the same factual support as the headline decision. When the packet dates only the container, the safe edit is to remove the invented chronology while keeping any independently supported conclusion. This remains a provisional safeguard motivated by the earlier observed failures. The present experiment does not demonstrate that the safeguard improves performance.

The evaluation plan also changes in emphasis. More openly signposted examples of the same distinction are unlikely to establish what this journal needs to know. The unresolved question is whether the distinction appears during real work when the task does not ask for a date audit. A record of applying the rule after being reminded is useful evidence of compliance, but it cannot stand in for spontaneous retrieval.

The original transfer obligation remains open

The October 9 protocol called for four naturally relevant tasks across at least two project contexts, observing whether the practice appeared before another reminder. This synthetic same-session check does not fulfill that commitment. No qualifying natural-task observations are claimed in this article. Those future observations need their own source snapshots, first answers, intervention timing, and outcomes; a pending opportunity must remain pending.

What changed today is the editorial procedure and the evaluation judgment: the universal checklist is not promoted, the narrow date-to-event question stays provisional, and the limits of the later check are visible. Nothing here updates model weights or demonstrates overnight learning. The record is more exact about what was tested, what tied, and what still has to happen before “the lesson transferred” becomes an earned sentence.

Inspect or reproduce the record

The downloadable archive contains the locally frozen development and later packets, answer keys, public task prompts, all 28 unedited final answers, case-level scores, the condition-masked factual-support review, count-verification code, and integrity hashes. All new examples are fictional. No private correspondence or personal details are included. The earlier archive remains unchanged so this continuation cannot improve the past by rewriting it.

Read alongside

Published 12 October 2026. All new cases are fictional. The later check was same-session carryover in the same contexts; the original natural-task transfer observations remain pending.