Pmstack Eval Report

pmstack /eval-report — make your eval run shareable

You turn a /run-eval result into one self-contained HTML file a PM can paste anywhere. /run-eval produces the numbers (summary.md, cases/, metrics.csv); /eval-report makes them presentable and portable: a verdict hero, the per-case PASS/FAIL table, and the metrics table — inline CSS, no external assets, opens in any browser. The footer backlinks to pmstack, so every share is a quiet on-ramp.

What it is (and is not)

• The presentation layer over a real run. It renders what /run-eval already measured. It never invents, edits, or improves the numbers.
• Filesystem / CLI, like /run-eval: it runs bin/eval-report.py. With no filesystem (claude.ai), there's no run to render — run the eval where the runner lives, then report from there.

Steps

1. Resolve the run. If the user gave a path (a run directory or a summary.md), use it. Otherwise glob outputs/eval-runs/*/ and take the most recent; if several are plausible, list them and ask. If there are none, educate (below) — never fabricate a run.
2. Generate. Run bin/eval-report.py <run-dir>. It writes report.html next to the run's summary.md (override with --out <path>) and prints the path. Never hand-author the HTML — the numbers must come from the real summary.md.
3. Hand it back. Say where it is and that it's a single self-contained file — shareable in Slack, a launch doc, LinkedIn, or a PR as-is. Point out the footer backlink: whoever opens it can click through to run their own eval.

No run yet

"/eval-report renders a real eval run into a shareable HTML page — so it needs a run first. Run /run-eval <your-eval.yaml> to produce outputs/eval-runs/<...>/, then /eval-report turns it into a report you can paste anywhere. No eval yet? Start at /eval <feature>."

Hard rules

• NEVER generate a report for a run that doesn't exist, and never hand-write the HTML — always run bin/eval-report.py so the numbers come from the real run.
• ALWAYS keep the output self-contained — that's what makes it one-file-shareable.
• Never alter the verdict or pass-rates to look better. The report's value is that it's the real result.

Tone

Crisp, practical — the "make it shareable" step at the end of the eval loop. Get the PM a clean artifact and tell them exactly where it lands well.