feat(evals): inline measured 2026-09-11 results across the evals pages

[devin-ai-integration]

Summary

Copies both 2026-09-11 artifact bundles (perps/predictions + spot comparison) byte-for-byte under apps/evals/src/results/2026-09-11/ and shows the measured numbers inline on the existing pages — no new route. Measured rows/panels are always labelled Measured and never mixed into illustrative aggregates.

Data boundary — apps/evals/src/measured.ts normalizes both JSONs into two MeasuredModels (gpt-5-5: Spot; gpt-5-6-sol: Spot + Perps + Predictions):

interface MeasuredFamilyResult {
  passed; failed; unscoredTimeouts; total;          // exported counts, displayed as-is ("43 / 54")
  dimensions: { routing, arguments, completion, safety };
  latencyMs: { p50, p95, max }; tokenUsage; cases;
  passRateSortKey: number;                          // passed/total*100 — sort key + bar width ONLY, never shown
}

Spot per-case rows are derived from the exported per-attempt verdicts (the spot report has no per-case aggregate); spot failed = total - passed.

Per page

• Leaderboard: measured rows appear only under Spot / Perps / Predictions (no Portfolio, no "All tasks" aggregate). Pass rate shows counts; Accuracy and Cost show — (not in the artifacts — tokens are the only cost proxy); p50 from latencyMs; Tasks = run total. Rank cell is — and illustrative ranks are computed among illustrative rows only. Measured rows sort by passRateSortKey, appear in the latency scatter, and are excluded from the cost scatter.
• Model profile: #/models/gpt-5-5 and #/models/gpt-5-6-sol render a measured profile: per-family counts, p50/p95, tokens, dimension breakdown, per-case verdicts, run ids/commits/timeout, PR Fix native OMP eval catalog compatibility and tool evidence #85 link, aborted predictions run + limitations. No ± bands/histograms or compare flow (no measured uncertainty exists).
• Task explorer: a Measured cases panel under the fixture comparison — Spot shows the 4 case prompts with a model × repetition verdict matrix and failure categories; Perps/Predictions show per-case results/notes. No tool-call bodies or transcripts (those were only in the excluded .jsonl).
• Methodology: "Measured runs · 2026-09-11" card (harness, reps, timeout, source commits, limitations).

vite.config.ts fmt ignore covers apps/evals/src/results/** so the artifacts stay byte-identical.

Verified locally: typecheck, vitest (10 tests), fmt, lint --deny-warnings, build.

Link to Devin session: https://app.devin.ai/sessions/9924c1a614b94139a2fda346027389e0
Open in Devin Desktop: https://app.devin.ai/desktop/session/9924c1a614b94139a2fda346027389e0?variant=devin
Requested by: @ericjuta

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

---

Original prompt from Eric

make a pr from the patches @askgina/plugins in the zip

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
plugins-evals	Ready	Preview	Sep 13, 2026 5:47pm UTC

[devin-ai-integration]

Browser-tested the measured-results page locally (recorded run).

• Route/title/nav highlight, dated chip, hidden Run button, provenance hashes, PR Fix native OMP eval catalog compatibility and tool evidence #85 new-tab link, footer.
• Spot: both models 11 / 12, 4 cases × 3 repetitions, prompts, expandable notes.
• Perps 43 / 11 / 0 / 54 and predictions 12 / 26 / 1 / 39 (passed / failed / timeouts / dispatched) match the artifacts; dimensions/cases tables, aborted run, limitations render.
• Leaderboard, Public results, Methodology, direct #/results reload, back/forward all fine; no console errors; 700px wraps without page overflow.

Desktop	700px attempt matrix