Skip to content

Benchmark page: PrecisionMemBench 77 of 77 - #46

Merged
chrisjleal merged 4 commits into
mainfrom
docs/benchmark-77-of-77
Nov 10, 2026
Merged

chrisjleal merged 4 commits into
mainfrom
docs/benchmark-77-of-77

Conversation

@chrisjleal

Copy link
Copy Markdown
Contributor
  • PrecisionMemBench is now 77 of 77 (precision 0.98, recall 1.00) with its own section, a new public record and the per-case results of all three full runs (77, 77, 74) under public/evidence/precisionmembench/2026-11-10-77/. The 76-of-77 folder and record stay published.
  • The static chart images and their CSS are removed. The charts live on proofable.me/benchmarks; this page links it in a note and uses its share image.
  • Run notes appear once, in Details for reviewers. Two run-date stamps and a duplicated "Run the test" link are gone from the body.
  • The changelog entry for today now reads 77 of 77.

Merge after proofable/neus#257, which serves the share image at proofable.me/og/benchmarks.

🤖 Generated with Claude Code

chrisjleal and others added 4 commits November 10, 2026 01:39
…me/benchmarks

The memory-precision row is now 77 of 77 (precision 0.98, recall 1.00) with its own section, a new public record and the per-case results of all three full runs. Run notes live once in Details for reviewers. The static chart images are gone; the charts and comparisons are on proofable.me/benchmarks, which this page links and uses as its share image.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Mintlify serves repository files under /public, so every docs.proofable.me/evidence/... file link returned 404, including the folder each signed benchmark record names. /evidence/<path> now redirects to the same path in the repository, and the benchmark page and changelog link the evidence folder directly.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Injection rates depend on the model, so the results table now sets 6.2% beside 16.4% for the same agent and model without Proofable, not beside a defense measured on a different model.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s as cards

The page opens with the two perfect scores and four headline cards instead of a paragraph. The eleven results are cards rather than a three-column table, so nothing is cut off on a phone. Remaining tables use tighter type on small screens, and the memory table drops to five columns.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@chrisjleal
chrisjleal merged commit 961ea5c into main Nov 10, 2026
2 checks passed
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