Skip to content

test(projects): measure large-document paging in project evidence - #185

Open
leemour wants to merge 1 commit into
mainfrom
perf/evidence-large-document
Open

leemour wants to merge 1 commit into
mainfrom
perf/evidence-large-document

Conversation

@leemour

@leemour leemour commented Oct 20, 2026

Copy link
Copy Markdown
Contributor

Why

The frozen performance/projects/evidence test (13 short texts) cannot exercise large-document paging or the document paragraphs shown in around (RISK-1268). Owner decision NEED-1572 A: a separate test. The frozen fixture, baseline.json and results are untouched.

What

New performance/projects/evidence-large-document/: runner, baseline.json (labels written before the first run), result.json, README. Linked from performance/projects/README.md and PROCESS.md. Memo 0.82.0 (579e0af), Core 0.371.0, synthetic data only, sandboxed in one /var/tmp folder.

Headline results

  • One invocation finds all 8 readable documents; each is read once (432,246 charged bytes for 431,744 document bytes). The 284 KB document is listed under oversized.
  • A resumed run with --max-pages 2 and --cursor never completes: 4 of 9 references, then every resume reads nothing and returns the same checkpoint.
  • around for a 145 KB document is centred on p60 (the key search), not on p4, the first paragraph that names the project; three of its four matching paragraphs are not shown.
  • The name search returns a document that has "lantern" and "works" in one paragraph but never the name.

Details and the exact commands are in the README. Docs checks pass (docs:check, docs:lint, docs:spell, plus cspell and markdownlint on the new README). verify-provenance.mjs fails the same way on main (not caused by this change).

🤖 Generated with Claude Code

A separate benchmark beside the frozen project evidence test: one 145 KB
document with four matching paragraphs far apart, a 284 KB document above
the single-record limit, six medium documents that need several pages and
two decoys. Labels are in baseline.json, written before the first run.

Measures paragraphs found, paragraphs shown in `around`, pages, charged
bytes, `oversized` and `--max-pages` with `--cursor`. Observed on Memo
0.82.0 / Core 0.371.0: one-shot paging is read once and `oversized` is
reported; a resumed run stalls; `around` follows the first search, not the
first paragraph; the name search matches words, not the phrase.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

This branch has not been deployed

No deployments
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.

1 participant