What the retained records added

Retrospective exploratory review · 18 October 2026

An all-pass comparison can leave useful observations outside its original score table. At Shawn’s request, agents revisited the retained records rather than running more participants. The original results were preserved; the later questions are identified as retrospective.

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. — Same recorded answers, smaller reported inputs: P1 2.96%; P3-D at 96, 384 and 768 records 12.61%, 21.36% and 24.10%. Reductions use arm totals and include host context. P3-D repeats the same two cases across loads; the studies are not pooled. These observations establish neither speedup nor cash savings. Figure generated with corpus-plot; the linked P1 addendum and P3-D table provide the underlying values.

Same answers, smaller inputs. In the evidence-presentation and load-development studies, Markdown tables used fewer reported input tokens than vertical labeled key-value records. Recorded answers and ordered citations matched within every pair. This is a resource observation, not a correctness advantage, proven speedup or monetary saving.
Same passing outcome, different work. In the request-wording study, some paired programs were byte-identical while others differed. All completed trajectories visibly ran supplied and added checks. The already-correct task stayed unchanged in both conditions, despite different checking effort. Final-reply wording and raw edit counts alone would have missed that distinction.
A reason not to scale one comparison. The load-development study still found no mixed-correctness range. The later input observation does not reverse its decision against a larger same-range accuracy comparison; it suggests a different question someone could specify separately.

These are observations from the completed P1, P2 and P3-D records, not proof that Corpus outperforms a simpler workflow. Agents selected and executed the specific methods; Shawn commissioned the research, reviewed the papers and prompted this broader reading. Full Human and crew labor remains unmeasured. A proposed fair comparison with a competent simpler workflow is design-only, not an experiment already performed.

The audit does not reconstruct internal reasoning from tool calls, summaries or token counters. Its value is more concrete: distinguish answers from artifacts, artifacts from checks, and task correctness from measured resource use. Original scores, missing tasks and limitations remain visible alongside these additions.

Coverage: new artifact/resource audits cover P1, P2 and P3-D only. A bounded reading of the September 8 extension explains existing resource/error rows without claiming a new result; all other older cohorts were inventoried, not freshly audited. Sources: FIND-6484/6489/6491; coverage FIND-6497 and bounded older reading FIND-6499. Human direction and agent responsibilities are described separately in the research-practice account.

The system account describes concrete continuity, correction and recovery cases, with a small dated size snapshot. These cases are not experimental replications or a system-wide success rate. They preserve Human intervention, agent repair and unknown costs beside the outcomes (FIND-6498).

Public retrospective edition assembled from the reviewed FIND-6502 package, 18 October 2026. Original study accounts remain alongside their dated additions.