docs(bench): record proof find against no-brief on the 10x graph

.flatbread-proof/findings/fnd-search-cards-raise-accuracy-and-cut-reading-on-t--ydqm2ybyqjedzw9y.md

@@ -0,0 +1,19 @@
+---
+id: fnd-search-cards-raise-accuracy-and-cut-reading-on-t--ydqm2ybyqjedzw9y
+effort: eff-effort-graph-memory-and-agent-wedge--szeqvmgqjqnhd002
+title: Search cards raise accuracy and cut reading on the 10x graph
+kind: measurement
+created_at: '2026-10-19T07:39:20.029Z'
+derives_from:
+  - fnd-a-graph-ten-times-larger-hides-answers-from-both--f9h4wcrg4jpzczzg
+---
+
+`proof find` searches every record by what it says and returns short cards: the current version of each hit, its best passages, any attached Blob passage, and one hop of links. On the benchmark graph ten times larger, Haiku answered the 245 comparable questions twice with `find` and twice without it (no-brief), one fresh session per question.
+
+- Accuracy rose by about 20 questions per run: 231.1 and 229.1 with `find`, against 210.9 and 210.7 without. Over both runs, `find` alone was right 48 times and no-brief alone 12 times (sign test p < 0.0001).
+- Agents read 61 percent fewer tokens (13,019 against 33,839 on average) and made 9 tool calls instead of 21.
+- "Unknown" answers fell from 32 and 35 per run to 14 and 12.
+- No question category lost ground. The gain held on the half of the questions the cards were not tuned on (26 wins, 6 losses), and with the seven timed-out sessions set aside (45 wins, 9 losses).
+- On 24 questions that name no Effort, `find` got 46 of 48 and no-brief 31 of 48.
+
+So the limit at scale was finding a record by what it says, as the 10x Finding suggested, and a search read lifts it. Open: the `brief` arm was not rerun, the graph is fictional and its filler easy to tell apart, and only Haiku was tested.

benchmarks/proof-recall/README.md

@@ -767,3 +767,60 @@ lookups, though, so they are easier than the full set, and they test
 subsystem names cost the gate one point of evidence and no answers, so the
 10x result is not only a sign that filler is easy to tell apart. The paid
 runs (Runs A to C in the plan) still have to show that agents gain from it.
+
+### `proof find` against no-brief on the 10x graph
+
+The plan's Run A and Run B, on Haiku, twice: `run.mjs --strategies find,nobrief --only default,ambiguous --graph <x10 build>`, 540 sessions per
+run. The `find` arm's task text starts with `proof find`, then `get` for a
+hit's full text, and `records` only to count. Both runs used the cards above
+(`find_hash` `e5be6532…`). Run files: `runs/x10-find-r1-*.json` and
+`runs/x10-find-r2-*.json`.
+
+On the 245 original questions (all but `world-077`, whose count includes
+filler), pooled over both runs:
+
+| Measure                                    | `find` | no-brief |
+| ------------------------------------------ | ------ | -------- |
+| Correct, run 1                             | 231.1  | 210.9    |
+| Correct, run 2                             | 229.1  | 210.7    |
+| Runs of a question only this arm got right | 48     | 12       |
+| Mean read tokens                           | 13,019 | 33,839   |
+| Median read tokens                         | 8,888  | 29,680   |
+| Mean tool calls                            | 9      | 21       |
+
+The sign test over the pooled disagreements gives p < 0.0001. Counting each
+question once, `find` did better on 37 and worse on 10 (p = 0.0001). It read
+fewer tokens than no-brief in 447 of 490 sessions.
+
+Against the plan's pass marks, on its 232 comparable questions (the 245
+less the 13 retraction questions):
+
+| Mark                                       | Result                                             | Pass |
+| ------------------------------------------ | -------------------------------------------------- | ---- |
+| (a) Pooled sign test p < 0.05              | 46 wins, 9 losses, p = 4e-7                        | yes  |
+| (b) Read tokens at most 0.65 of no-brief's | 13,260 against 34,313: 0.39                        | yes  |
+| (c) "Unknown" falls by at least 10 per run | 32 to 14 and 35 to 12                              | yes  |
+| No category loses more than 2 questions    | No category loses any; `negative` 30 of 32 in both | yes  |
+
+Seven sessions hit the 10-minute limit, six of them in the no-brief arm.
+Setting those questions aside leaves the result in place: 45 wins, 9
+losses. The gain also holds on the questions the cards were not tuned on:
+26 wins and 6 losses on the holdout half (p = 0.0005), against 23 and 6 on
+the tune half. The largest gains are in `lookup` (+13.8 over two runs) and
+`truncation` (+13), where the fact sits past the 600-character cut that
+`records` applies; the smallest are in `temporal`, `negative`, `rationale`,
+and `dead-evidence` (0 to +1).
+
+**The ambiguous-Effort set (Run B).** On the 24 questions that name no
+Effort, pooled: `find` 46 of 48, no-brief 31 of 48. `find` did better on 10
+questions and worse on none (p = 0.002), and read 9,914 tokens to
+no-brief's 42,017.
+
+What this does not show:
+
+- The `brief` arm was not run, so this says nothing new about compaction.
+- Haiku only, and on a fictional graph whose filler is easier to tell apart
+  than real records. Run C, on the `--filler-overlap 0.3` graph, tests that.
+- The ambiguous questions come from tiers 0 to 2. They show that search
+  finds the right Effort, not that agents handle a hard question without a
+  named Effort.