docs(bench): record Run C, Sonnet, Run D and a held-out check on real graphs

[tonyketcham]

Summary of changes

This is layer 10 of a stack. It sits on #289.

This PR adds no code. It records four sets of paid benchmark runs, all on Haiku except where noted.

1. find against no-brief on the 10x graph: look-alike filler, and Sonnet. These runs use the same task texts and find_hash as #286. Figures are for the 245 comparable questions.

Run	find correct	No-brief correct	Only this arm right	Read tokens, find ÷ no-brief
Haiku, look-alike filler graph	233.7	214.7	19 to 1 (p < 0.0001)	0.38
Sonnet, plain graph	233.4	226.7	11 to 5 (p = 0.21)	0.56

• Look-alike filler: 30% of the filler records reuse real subsystem names, and the gain stays at 19. That is the plan's Run C, and it passes.
• Sonnet: without search Sonnet already does well, so its accuracy gap is small. Search still saves 44% of the reading.
• Questions that name no Effort: find won 7 to 0 under both models.

2. Run D: proof state added to find. This ran twice on the 97 count, status-history, blocker and dead-evidence questions of the 10x graph. findstate got 181 of 194 question-runs right and find 171 (15 to 5, p = 0.041). It read 18% fewer tokens. That meets the plan's mark for keeping state: at least 4 more per run, with no rise in tokens. The holdout half, which state was not shaped on, points the same way (5 to 1) but is too small to show on its own.

3. A held-out check on two real graphs. Two private project graphs, of 200 and 134 records, tested whether the gains hold on memory nobody tuned against. An author who never saw find or state output wrote 39 and 41 checked questions. Each set ran twice, with three arms.

Graph	find	No-brief	findstate	Read tokens, find / no-brief
A	71.7 of 78	66.3	71.3	12,488 / 27,975
B	75 of 82	71	74	12,230 / 24,389

• find against no-brief: 13 to 3 over both graphs (p = 0.021), at about half the reading.
• findstate against find: no difference, at the same tokens. The state gain does not show on graphs this small.

Privacy. The private questions, answers, graph copies and run files never enter this repository. Only the totals above are published. The diff was checked for any trace of them before each commit.

Files:

• run files for the synthetic runs in benchmarks/proof-recall/runs/;
• four README sections;
• three measurement Findings in .flatbread-proof/.

Please don't delete this checklist! Before submitting the PR, please make sure you do the following:

• I added doc comments to any new public exports, and inline comments to any hard-to-understand areas (none added)
• My changes generate no new console errors locally
• If applicable, try to include a test that fails without this PR but passes with it (results only, no code)

Does this introduce any non-backwards compatible changes?

• Yes
• No

Does this include any user config changes?

• Yes
  • If so, I have updated the relevant areas of documentation
• No

🤖 Generated with Claude Code

https://claude.ai/code/session_01DVVnhYL1Bz7mCfEKDoyLMW