Evidence presentation under load

P3-D · Completed development study and later exploratory reading · 18 October 2026 · Not peer reviewed

This account separates a prospective task-development result from a subsequent audit of retained inputs and answers. It is a distinct study from P1 and P2. The research-practice account explains Human direction and agent execution; the retrospective overview connects the later observations. The working archive remains private.

Original development result: no mixed-success range

P3-D was a development study, not a test for a winning format. It used two constructed evidence-binding cases, each presented at 96, 384 and 768 records in both vertical labeled key-value and Markdown-table forms. Lower-load evidence remained unchanged within higher loads. There was one fresh trajectory per case, load and format on the recorded Astra/medium/Codex configuration.

All twelve planned trajectories completed correctly on semantics, citations and schema. Each load returned four correct answers out of four. The predeclared rule sought the lowest fully observed load with one to three correct answers; none qualified. The practical decision was not to scale this same accuracy comparison at these loads merely to repeat the ceiling. No further dose or study followed automatically.

This does not establish format equivalence, a general context-capacity limit or a failure probability. The two cases reused inherited mechanisms, exact keys assisted retrieval, and the load sequence was also a time sequence. The study did not test other models or hosts. Reported participant usage was 417,801 input-plus-output tokens; whole-crew cost was not measured.

Later exploratory reading: input economy increased across the tested loads

Retrospective exploratory addition · 18 October 2026. The original development decision remains unchanged.

A subsequent audit found lower reported input usage for tables in every case/load pair, with identical recorded answer strings and ordered citations. The same two cases were repeated across loads; these are six dependent case/load pairs, not six independent cases.

Markdown tables used fewer host-reported input tokens with the same recorded answers in each matched pair: 2.96 percent fewer for P1; 12.61, 21.36 and 24.10 percent fewer at P3-D loads of 96, 384 and 768 records. These are separate summaries, not pooled independent cases. — P3-D's table-input reductions were 12.61%, 21.36% and 24.10% at 96, 384 and 768 records, with the same recorded answers in each matched pair. P1's 2.96% is a separate reference cohort. P3-D repeats two cases across loads; these rows are not pooled independent cases. Host-reported inputs include host context; no speedup or cash saving follows. Figure generated with corpus-plot. The table below gives P3-D's values; the retrospective overview connects the separate studies.

Records	Matched case/load pairs	Vertical-record input tokens	Table input tokens	Reduction from arm totals
96	2	30,390	26,559	12.61%
384	2	73,219	57,580	21.36%
768	2	130,244	98,861	24.10%

Within each case, the recorded answer was also identical across all format/load combinations. The input difference is a useful measurement about the encodings despite tied correctness. Native input includes host context; cache, timing and backend conditions were not controlled for an efficiency experiment. No reliable speedup, billed saving or internal reasoning strategy follows. These later observations do not change the no-mixed-range decision or become a predeclared endpoint.

Source history: protocol FIND-6460; original collection/analysis/report FIND-6474/6477/6479, accepted ending FIND-6480; later audit FIND-6484 and public table FIND-6489. The earlier unrun 16/96 preparation is separate process history, not an additional empirical cohort.