Skip to content

docs(bench): record Phase 2 on the scaled test bed - #296

Draft
tonyketcham wants to merge 1 commit into
claude/agent-memory-context-potency-6sniud-p2-scalefrom
claude/agent-memory-context-potency-6sniud-p2-scale-results
Draft

tonyketcham wants to merge 1 commit into
claude/agent-memory-context-potency-6sniud-p2-scalefrom
claude/agent-memory-context-potency-6sniud-p2-scale-results

Conversation

@tonyketcham

Copy link
Copy Markdown
Collaborator

Summary of changes

This is layer 16 of a stack. It sits on #295.

This PR adds no code. It records the paid rerun of the Phase 2 trial on the scaled test bed from #295. Haiku answered the same 44 code and branch questions twice, using findcode (proof find --path and --refs all) or read-only git alone, with a 20-minute timeout.

Test bed findcode correct No-brief correct Read tokens, findcode / no-brief Tool calls
Small (524 records) 85 of 88 84 of 88 11,642 / 26,283 7 / 13
Scaled (5,131 records) 85 of 88 85 of 88 13,422 / 22,829 7 / 15

What it shows:

  • Accuracy: scale does not separate the arms. They disagree on two question-runs, one each way.
  • Cost: findcode reads 41% fewer tokens and makes half the tool calls.
  • Session checks: no session timed out. One no-brief session was flagged for reading outside its folder. It was reading its own background task's output, nothing from the benchmark.

This differs from search over record text, which did gain accuracy on the 10x graph. Git is already a precise index for code. git log -- <path> and git diff main...origin/<branch> stay small however large the graph grows, and the filler never touches a real file. So --path and --refs earn their place by cost, not accuracy. A test that could show more would ask about code that no single path names, or about many branches at once.

Files:

  • four run files;
  • a README section, "Phase 2 at scale";
  • a measurement Finding 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 by Claude Code

Haiku answered the 44 code and branch questions twice on the 10x test
bed, with `findcode` (proof find --path and --refs all) or with
read-only git alone:

- correct: 85 against 85 of 88 question-runs (1 to 1), as on the small
  bed (85 against 84);
- read tokens: 13,422 against 22,829 (41% fewer); tool calls 7 against 15.

Scale does not separate the arms, because git log and git diff stay
precise as the graph grows. The README and a measurement Finding say
the reads earn their place by cost, and name the test that could show
more.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DVVnhYL1Bz7mCfEKDoyLMW
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants