Skip to content

docs(benchmarks): ATR vs Defender Garak/PINT baseline - #77

Open
hiskudin wants to merge 1 commit into
mainfrom
docs/atr-baseline-with-defender
Open

docs(benchmarks): ATR vs Defender Garak/PINT baseline#77
hiskudin wants to merge 1 commit into
mainfrom
docs/atr-baseline-with-defender

Conversation

@hiskudin

Copy link
Copy Markdown
Collaborator

Summary

Test plan

Made with Cursor

Fill the Defender Tier 1 and Tier 1+2 columns on the same corpora as the
ATR-only baseline from #66 / #75 so the comparison is in-repo.

Co-authored-by: Cursor <cursoragent@cursor.com>
Copilot AI review requested due to automatic review settings July 21, 2026 13:25
@hiskudin
hiskudin requested a review from a team as a code owner July 21, 2026 13:25

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Adds a benchmark documentation page that places Agent Threat Rules (ATR) results side-by-side with Defender Tier 1 and Tier 1+2 results on the same Garak and Lakera PINT corpora, to fill the previously blank Defender column discussed in #66 / #75.

Changes:

  • Adds benchmarks/atr-baseline.md documenting ATR vs Defender numbers for Garak (recall) and PINT (precision/recall/F1 + confusion matrix).
  • Documents reproduction steps for rerunning ATR and Defender on the referenced corpora.
  • Adds interpretation notes (“Reading the numbers honestly”) to clarify what each corpus does/doesn’t measure.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Comment on lines +18 to +20
- **Defender**: local `@stackone/defender` (post-0.7.2 workspace build).
- **Tier 1**: `enableTier1: true`, `enableTier2: false`
- **Tier 1+2**: `enableTier1: true`, `enableTier2: true` (default cascade)
Comment on lines +48 to +52
Use the same corpus JSON under the ATR checkout (`data/test-corpora/garak-full/`,
`data/pint-benchmark/pint-corpus.json`) and call `createPromptDefense` +
`defendToolResult` with `blockHighRisk: true`. Count a positive when
`result.allowed === false`. Run once with `enableTier2: false` (Tier 1) and
once with Tier 2 enabled (Tier 1+2).
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