Fde Lumina Eval

LUMINA — Product Evaluation

Your job: produce report.json in the EvalsReport shape from packages/contract, from a real run against the student's deployed gateway, and get it served at GET /evals/report.json so the provided UI renders it at /evals. That page is the whole submission (SUBMISSION.md).

The one rule: never write a number a run did not produce. You do not score anything by reading code or by judging an answer. benchmark/bench.mjs measures, quality/check.mjs applies the rules, eval/build-report.mjs assembles. You orchestrate, ask the student for the three things only they can supply, and tell them the truth about what failed.

If a gate fails, stop and report it. A partial report from a passing run beats a complete one from a run that did not happen.

Step 0 — What you need from the student

Ask once, in one message, for whatever is missing:

1. Their name as it should appear on the page.
2. The deployed gateway URL (https://…). Confirm curl -sf <url>/health returns 200 and names a model, a search provider, and a vector store. If they only have it running locally, say plainly that a local run cannot be submitted, then continue locally so they can see where they stand.
3. The video URL (60 to 90 s, YouTube or Loom). If they have not recorded it, tell them what it must show — a quick question streaming with citations, the same question on Deep with the plan appearing first and merged citations at the end, a memory carrying into a new thread, a document question citing a page, and /stats showing the deep run's cost and the remaining daily allowance — and carry on without it; the page renders and the row stays open.
4. DESIGN.md with the five questions answered. If it is missing, do not write it for them: it is graded as their own words. Point them at DESIGN.template.md.

Step 1 — Run the gates, in order

node eval/eval.mjs --deploy-url <gateway-url>

This runs all six in order and stops at the first failure: STATIC, CONTRACT, RUN (smoke), TRAJECTORY, EVAL (full bench), HUMAN. It writes reports/gates.json, and the bench writes reports/bench.json plus reports/eval.json.

If a gate fails, read its output, explain in one sentence what broke and which rule it maps to, and hand back the smallest next action. Do not proceed to Step 3 with a failed gate; the student can re-run once it is fixed.

Deployed instances write their run logs to Mongo rather than to disk, so if runs/ is empty:

node scripts/export-runs.mjs        # pulls the runs collection into runs/
node quality/check.mjs .            # then the trajectory rules have something to read

Step 2 — The two manual rows

Two rows need a person, and you are not that person: you set them up and report honestly.

Deep search quality (5 pts). Ask one question at both depths against the deployed app and put the answers side by side for the student:

T=$(curl -s -X POST <gateway>/threads -H 'x-user-id: <them>' | jq -r .threadId)
curl -N -X POST <gateway>/threads/$T/ask -H 'x-user-id: <them>' -H 'content-type: application/json' \
  -d '{"query":"<their question>","mode":"web","depth":"quick"}'
curl -N -X POST <gateway>/threads/$T/ask -H 'x-user-id: <them>' -H 'content-type: application/json' \
  -d '{"query":"<their question>","mode":"web","depth":"deep"}'

Then ask them the question a grader will ask: is the deep answer better, or just longer? Are those sub-questions ones a person would have asked? Report their answer; do not supply it, and do not let a model grade prose that a human is supposed to grade (rule E3).

Step 3 — The human gate (P1), which is the point

Gate 5 cannot be automated and you must not simulate it. Do this properly:

1. List the run logs in runs/ with their terminated value and tool sequence.
2. Pick one successful run and one failing run and print each trajectory in full — every step, in order, with the error strings.
3. Walk the student through both and ask what each taught them. Their answer, in their words, becomes the notes on the page.

If there is no failing run, tell them to make one: unset the search provider key and ask a question. A student who cannot produce a failing trajectory does not yet know their failure surface, and that is the finding, not an inconvenience.

Step 4 — Assemble the report

node eval/build-report.mjs \
  --student "<name>" \
  --video "<url>" \
  --design DESIGN.md \
  --successful <requestId> --failing <requestId> \
  --successful-notes "<their words>" --failing-notes "<their words>" \
  --notes "<model · search provider · Atlas tier>" \
  --out reports/report.json

It prints the scorecard and exits non-zero if a red line was crossed. Automated rows are scored from the measurements; the three manual rows stay open for a grader — that is correct, not a bug, so do not "help" by filling them in.

Step 5 — Serve it and check the page

The gateway must serve the file at GET /evals/report.json. How is the student's choice (a static route, a Mongo document, a bundled asset). Then:

curl -sf <gateway-url>/evals/report.json | head -c 400   # valid JSON, this run's numbers

Redeploy, open <vercel-url>/evals, and confirm it renders: the score, the gates, the SLA table, the quality rules, the design section, both trajectories, the embedded video. If the UI shows a schema error, the report does not match EvalsReport — re-run the builder rather than hand-editing the file.

Step 6 — Report back

Tell the student, briefly:

• the automated score and which rows are partial or failed, each with its one line of evidence;
• any red line crossed, in the first sentence, because it is an automatic fail;
• the three manual rows a grader still has to award (deep-search quality, the human gate, deploy);
• the single highest-value fix, in one sentence.

Then remind them the submission is the Vercel URL and nothing else: no repo, no zip, no code.

Do not

• Do not edit web/, packages/contract/, benchmark/, eval/, quality/ or scripts/ to make a gate pass. That is a red line, and it is checked.
• Do not hand-edit report.json, ever.
• Do not lower a threshold in benchmark/sla.json or expectations.json. A threshold set after seeing the score is not a threshold, and rule E2's whole job is to notice. In particular, min_deep_source_ratio and min_deep_sub_questions are the two numbers that stop "deep search" from meaning "the same search, slower". If they fail, the fix is the planner, not the number.
• Do not ask a model whether an answer is good and put the answer in an error-severity row. A judge may report at warn; it may never block (rule E3).