Triage evals

Scored against 150 Astro issues that a maintainer labeled by hand. I skipped anything the bot labeled, and kept these out of the prompt. Back to the inbox

claude-haiku-5-5 prompt v1, 2026-10-26claude-haiku-5-5 prompt v1, thinking off, 2026-10-26claude-sonnet-5-5 prompt v1, 2026-10-26Baseline
Priority, exact42.7%40.7%52.7%36.0% always P3
Priority, within one level85.3%82.7%91.3%
Packages, precision / recall80.0% / 93.6% on 44 a maintainer labeled75.4% / 97.9% on 44 a maintainer labeled70.8% / 97.9% on 44 a maintainer labeled
Needs reproduction, accuracy84.0%82.0%88.0%92.7% always no
Needs reproduction, precision / recall19.0% / 36.4%16.7% / 36.4%29.4% / 45.5%
High confidence, exact42.9% of 733.3% of 375.0% of 4
Medium confidence, exact44.0% of 13440.1% of 13750.7% of 136
Low confidence, exact22.2% of 950.0% of 1070.0% of 10
Failed0 of 1500 of 1500 of 150
Quote not in the issue110
Planted injection flaggedYesYesYes
Cost per issue$0.00054$0.00057$0.00835
Response time, typical / slow3.9s / 5.9s4.2s / 6.3s3.8s / 7.4s

These runs went through Claude Code in headless mode with its own prompt, tools and settings switched off. The model saw the same prompt as the live endpoint, with the answer's schema written into it instead of passed as a tool. Cost is what the same tokens would cost on the API at list prices. Response time is the model's own time, without the few seconds it takes to start each process.

Priority, what it confused (claude-haiku-5-5)

Maintainers saidAI said P1AI said P2AI said P3AI said P4AI said P5
P103000
P21910170
P32524230
P40119310
P500230

Where it was wrong (86)

What people decided

Loading.

Drafted replies, reviewed by hand

Not reviewed yet.

The models may have seen some of these issues in training, which would make the scores look better than they are. Using the newest issues keeps that to a minimum. The issues were opened in 2025 and early 2026 and fetched 2026-10-25.