The experiment is done. Now run a post-mortem on HOW it ran — from evidence, not memory. Locate the session logs (e.g. ~/.claude/projects/ or ~/.codex/sessions/ or ...) and the per-run records the harness saved. Every claim below must cite a specific run, log line, or transcript excerpt; anything you can't evidence, label "speculation".

1. VALIDITY AUDIT (do this before trusting the results). Hunt for reasons to distrust the numbers: any sub-agent that saw the answer key, the other folder, or hints of the hypothesis; grading errors (right answer marked wrong, or vice versa); tasks both conditions failed the same way (a task problem, not an organization problem); silent cap-hits or crashes counted as wrong answers. Verdict: results stand / stand with corrections / don't stand.
2. TRAJECTORY DIFF. Compare A vs B sub-agent traces. Where did tool calls go — discovery, retrieval, or verification? Classify each failure by stage: couldn't discover, misunderstood, retrieved wrong, trusted the wrong version, or couldn't recover. Cluster repeated failures across runs: one finding + frequency + one representative trace each, not a run-by-run list. Then diff the single best A run against the single best B run: what did the winner's path look like?
3. HARNESS REVIEW. How did YOU (the parent) spend effort? Wasted steps, rework, wrong assumptions caught late. What would you change in the benchmark prompt itself?
4. LEARNINGS. From evidence above only: (a) one revision to the technique card, (b) one revision to the benchmark prompt, (c) the single most surprising thing the traces show about how agents consume data. Save all of this to POSTMORTEM.md.

Do not defend the experiment. A finding that weakens the headline result is worth more than one that flatters it.