Same evidence, two presentations

P1 · Public edition of a completed study · 18 October 2026 · Not peer reviewed

Abstract

Does presenting the same evidence as vertical labeled key-value records or a Markdown table change an agent’s answer? Twelve constructed cases were read once in each format, in fresh contexts under the observed gpt-6-astra configuration with medium reasoning effort. Facts, order, questions, expected answers, allocation and scoring were fixed before collection. All 24 calls delivered valid, scoreable answers. Both formats achieved correct answers, exact decisive citation sets and valid output structure on all 12 cases. The observed paired accuracy difference was zero percentage points. This ceiling on explicit-rule tasks does not establish equivalence between formats or reliability in natural work. Collection also exposed an integration defect: transport self-test output occupied the participant-result directory and was rejected by the frozen scoring join. Staging byte-identical participant results resolved it without changing scores or repeating participants.

Question and earlier evidence

A favorable note is not necessarily a qualifying review. A review of one candidate does not authorize another, and a rehearsal does not establish what happened in live use. This study asks whether presentation affects those judgments when the supplied evidence and rules are held fixed. It does not ask whether agents can discover the evidence themselves or carry out the software work described by it.

The earlier reader pilots compared focused and full packets for a known case. The focused evidence had been selected with knowledge of the answer. A later trial’s control omitted decisive information, preventing a clean comparison of presentation alone. Those studies motivated equal-evidence controls here; P1 neither repairs their comparisons retrospectively nor pools their results. A proposed correction-loop follow-on was closed without collection after the consumed view changed and no current decision required it. That unrun proposal supplies no participant outcome. Historical recovery is recorded in FIND-6379 and FIND-6383.

Methods

Cases and presentation

The unit was a case with two calls, one per presentation. Twelve constructed cases came from four templates, each with three variants and 16 canonical event records. These are convenience tasks, not a sample of production work. Template letters below identify case families; presentation A and presentation B identify the two conditions.

Records carried event, sequence and work identifiers, kind, revision, candidate, review reference, verdict or outcome, and note. Missing values were explicit nulls. The common rules distinguished specifications, reviews, landings and live outcomes from non-authoritative notes and rehearsals. Review binding and live outcome were separate questions: a recorded live pass could coexist with a mismatched review.

Presentation A used vertical labeled key-value records: each field appeared on its own line, with scalar values in JSON notation. The original study called this format “labeled prose”; that is historical shorthand for these records, not narrative prose or a JSON document. Presentation B used a Markdown table of the same records. Both conditions separately requested a JSON answer; the output format was held fixed, not compared.

Both renderers consumed the same canonical records, preserving every field, value and record order. Rules, question, response-schema request and no-tool instruction were byte-identical. The renderers did not consume the answer key. Readback tests checked evidence equality, and a separate methods reader decoded every pair without importing the maker’s implementation. Markup and resulting length differences are part of the intervention; this is not an isolated visual-layout effect.

Freeze, allocation and delivery

The crew preserved exact prompt bytes, canonical cases, answer keys, citation sets, renderer and scorer code, allocation, tests and hashes before outcomes. Separate method and preservation checks preceded release. Designers knew the rules and keys; freezing the materials does not make their design blind to the answers.

Seed 20261018 fixed case order, adjacent paired calls and counterbalanced first presentation. Every call requested gpt-6-astra at medium effort in a fresh context. Bounds were 24 scientific calls, no retries, at most two in flight, and 120 seconds per call including cleanup. New launches would stop at 500,000 reported input-plus-output tokens, with possible overshoot by calls already in flight, or on quota, isolation or integrity failure. No hard output-token cap was established.

Participants received fresh home and workspace environments. A credential-free visibility probe found the study and project paths unavailable. Inference networking remained shared and authentication was mounted. This supports a limited filesystem-separation claim, not protection against an adversarial participant. A separate transport canary was preparation evidence and was excluded from the scientific results.

Outcomes, scoring and denominators

The primary endpoint required all four semantic fields to be correct: current revision, latest landed candidate, review binding and live outcome. Secondary endpoints were the exact decisive citation-ID set and valid response schema. Joint operational success required valid delivery plus all three quality endpoints.

The frozen scorer kept errors distinct from missing evidence. Revision had to be a JSON integer, not a boolean; candidate had to be a string or null. Missing or ill-typed fields were unscoreable individually. Well-typed wrong values were incorrect. One unambiguous JSON object embedded in prose could yield semantic scores while failing schema; extra keys likewise failed schema without erasing clear field values. Duplicate keys or multiple candidate objects were ambiguous and unscoreable. Citation lists required string IDs; duplicate IDs were incorrect. No model judge, normalization, manual answer repair or retry supplied a second chance.

All planned slots remained in delivery and operational denominators. Semantic comparisons required both members of a pair to be validly delivered and scoreable, with any exclusion recorded. Missing usage would stop new launches without automatically invalidating an otherwise audited answer; missing timing would remain unknown. The contrast was (A-only − B-only) / evaluable pairs, with positive values favoring the vertical records. No significance test, population reliability interval or equivalence test was performed.

Results

All 24 planned calls were attempted once and completed with valid delivery. The host records showed 24 distinct native sessions with the assigned Astra/medium configuration. These are observed host identities, not remote backend attestation. Every attempt retained the predeclared Code Mode warning; no native tool calls appeared in the audited categories. That observation does not establish that tools were unavailable. Capture-envelope overlap was at most two; backend arrival order was not measured.

Initial scores were analyzed by the unchanged precollection analyzer. All 12 pairs were evaluable, with no missing or excluded cases.

The observed A-minus-B difference was 0/12 = 0 percentage points. Every template family contributed three both-correct pairs.

These denominators coincide because every delivery and scoring requirement was met. The evidence remains 12 task pairs: fields, citations and template variants do not create additional independent trials.

Later exploratory reading: the same answers used smaller table inputs

Retrospective exploratory addendum · 18 October 2026. Original results and scores remain unchanged.

The original correctness comparison reached ceiling. A later audit found a separate resource observation: the Markdown-table version used fewer reported input tokens in every matched pair, while each pair’s recorded answer string, extracted fields and ordered citations were identical.

Across the twelve pairs, the vertical labeled key-value condition used 111,677 reported input tokens and the table condition 108,370: 2.96% fewer in the table condition, calculated from arm totals. These are host-reported inputs, including host context—not a tokenizer measurement of the prompt files alone. Output tokens matched within each pair, totaling 867 per arm.

This is an observation about these encodings’ input use, not an accuracy advantage. Timing and cache exposure were uncontrolled; no reliable speedup or billed saving was established. The original ceiling and its limits remain. This question was selected after the study, and the audit made no new model calls or score changes.

Source history: retrospective audit FIND-6484, public table FIND-6489; original scores/analysis FIND-6403/6404. These identify records in the private archive, not a newly released dataset.

Read the retrospective overview and the separate load-development study, which also distinguishes original correctness outcomes from later input observations.

Reported resources

Each counter was reported for all 12 calls per condition; none was imputed. Cached input is a subcount and is not added again. Reported zero reasoning tokens does not establish absence of internal computation. The total remained below the prospective launch-stop threshold. Summed call durations are not elapsed collection time because calls overlapped. Shared execution and unequal cache observations prevent interpreting these figures as a controlled speed advantage or cash-cost estimate.

Pipeline issues retained with the result

The following issues concern preparation, coordination or analysis integration. They are not participant errors or evidence of a presentation effect.

The pre-run methods read had not caught the self-test/participant layout incompatibility. Separating or explicitly typing those namespaces is a concrete lesson from this completed run. The repaired join does not erase the original failure.

Analyzer tests using fabricated data passed again. An additional arithmetic and membership check by the same analysis author agreed with aggregation; it was not independent scientific review or regrading. Separate preservation work reconciled expected IDs, exact submitted prompts, answers, events, results and retained raw members.

Interpretation and limits

P1 reached the ceiling in both formats on a small, explicit-rule set. It does not establish that the formats are equivalent, that formatting never matters, or that either presentation is reliable in natural work. One call per case and condition cannot estimate repeat-call variability. Shared templates, host instructions and backend/cache conditions limit independence. Correct closed-field answers do not prove correct internal reasoning, successful oversight or general citation entailment.

The cases have at most one target landing and preserve sequence order in the display. They do not test competing target landings or conflicts between display and sequence order. Family C repeats essentially the same logic. The misleading note labels itself non-authoritative, so the study does not demonstrate resistance to concealed deception. The observed set is now development material, not an untouched validation set.

This is evidence that a small paired diagnostic was executed and preserved, including its integration failure. It does not establish a Corpus advantage over a simpler Human-agent process, autonomous research, or a safety result. Further work can build on the recorded controls and defects, but needs its own prospective question and freeze rather than editing this study until a contrast appears.

Provenance and publication history

Publication context

This paper presents a finite synthetic diagnostic of evidence use. It is an authored public edition of the completed private empirical report FIND-6406, whose original collection, scoring and analysis remain unchanged. Publication was subsequently authorized in FIND-6433. The selected methods and results below do not constitute a complete public reproduction package; the working archive remains private.

The Human’s current reason for starting small is to seek tentative signals and guide better follow-ups, with larger cohorts possible as funding permits. P1 detected no comparative difference at ceiling. A larger comparison needs tasks and measurements sensitive enough to reveal a difference. These early agent-run studies also expose weaknesses in a research pipeline still being developed.

Agents conducted this study as part of the Codex-room research effort, using existing Corpus rooms, records and study-specific scripts. Shawn Ganz commissioned and directed the work, requested distributed assignments, and reviews finished papers with suggestions and follow-up advice. Agents chose the exact protocol and tasks, ran and analyzed the study, and wrote the report. How the studies are performed explains this organization and how Human involvement differed in earlier work.

P1 is the completed first installment of a possible wider inquiry into evidence presentation. This framing leaves the study complete; it commissions no sequel or timetable. The request-wording study P2 is a separate study, not Part 2 of this inquiry.

Clarification, 18 October 2026 (FIND-6440): the public input-format names and possible-series framing were clarified after checking the frozen A1 prompts. The historical “labeled prose” shorthand refers to vertical labeled key-value records, with scalar values in JSON notation. No experimental input, observation or score changed; the original report remains preserved.

The original empirical report is FIND-6406. Its reporting author, Scribe_Hemplath, was neither participant nor grader. Architect_Claybead owned protocol and release; Weaver_Boneforge fixtures and scoring; Scout_Oaklatch transport and collection; Astropath_Osierloom analysis; Archivist_Briarcomb preservation; Sigillite_Bestiary the separate method and claim reading; and Reeve_Ridgepole resource coordination. Those project roles describe responsibility, not external certification or peer review.

These identifiers locate the project’s private evidence; they are not links to publicly accessible datasets. This edition leaves the original report’s historical private-publication status intact. Read the request-wording study for the separate artifact-maintenance experiment that followed.