capture-evidence

Capture Evidence

Use this worker when the developer has not yet produced current local evidence for the workload, or when any core artifact is missing, stale, ambiguous, or untrusted.

The OSS loop does not require registration, auth, provider keys, an Understudy account, or hosted gateway access. Do a sufficient local pass that turns the workload into auditable artifacts and can answer the named decision.

Safety Gates

Default to the evidence plan most likely to resolve the decision under the developer's constraints, not the cheapest or smallest pass. Make cost, time, scope, and expected confidence visible; use ../understudy/reference.md → Outcome-first spend posture. A developer action that launches a named bounded evidence plan authorizes its declared uploads, hosted calls, model evaluation, receipts, and cleanup. Ask again only if the plan expands its displayed envelope.

Follow the repo public boundary in ../../docs/privacy-and-data-boundaries.md. Prefer metadata, paths, hashes, counts, schemas, and representative examples. Dropped workload content is available to the active analyst, and workload data may move through the destination named by an activated plan. Never print, commit, or transmit secrets; do not send data beyond the activated destination.

Goal

Create or refresh these artifacts under .understudy/capture-evidence/:

workload-profile.md
harness.json
environment.json
metric.json
splits.json
baseline.json

Each artifact must include a creation timestamp, source refs or path refs, and enough provenance for another agent to repeat the step without guessing.

Required Checks

0. Confirm the workload profile. If .understudy/capture-evidence/workload-profile.md is missing or stale, route to ../understand-workload/SKILL.md before building metrics. The profile should summarize the task purpose, data or trace distribution, prompt/request structure, code path, tool/action surface, output contract, failure taxonomy, and user-confirmed success criteria. It may cite local paths and hashes, but should not contain raw private payloads unless explicitly approved.
1. Attach the harness. Capture the local runner, command, fixture path, entrypoint, timeout, dependency notes, input schema, output schema, and validator invocation in harness.json.
2. Attach the environment. Record language/runtime versions, package manager, relevant lockfile status, model/provider route used by the incumbent, local hardware notes when relevant, and required env var names without values in environment.json. If the harness needs a local proxy, bootstrap repair, or provider-key presence check, route to the existing public setup skill for that recovery path before claiming the baseline is runnable. Model preflight: record whether the intended candidate/student model supports the workload's required capabilities (tool-calling, structured-output, vision, reasoning toggle) and whether its context window fits the workload's longest input. A context-window mismatch is a silent failure mode that later surfaces as confusing zero scores.
3. Confirm the scoring metric and validator (the load-bearing step). Write metric.json with the primary metric, pass/fail threshold, tie-breakers, validator, failure taxonomy, and approved: true only after a human confirms it. The metric is the real game: optimizing a proxy metric instead of the real validator is how prior runs scored 0/12. Record the validator kind and follow its rule:
   • unit-test / golden / custom-command — runs a deterministic check; the feedback is the assertion or diff that failed.
   • schema (e.g. Zod/JSON-schema safeParse) — keep schema_pass separate from quality_pass; a valid-shape, valid-enum output must not be failed merely for not matching a teacher trace verbatim. Two grounding checks a shape-only schema misses, both observed in real workloads: (a) verbatim evidence — when a field claims to quote a source (transcript, doc, log), verify the quote appears verbatim in the source; smaller open models hallucinate correctly-formatted quotes; (b) conditional requireds — when an optional signal (date, flag, risk) is present, its evidence subfield must be present too; models that include the signal but drop its evidence are the dominant residual failure and usually prompt-fixable.
   • rubric — a confirmed criteria list (each criterion: id, description, review type); auto-generated rubrics need human approval.
   • llm-judge — must debias position with a swapped two-pass score ((r_ab − r_ba + 2) / 4); never single-pass.
   • human-review — a blind, order-randomized packet; report judge-vs-human agreement separately from candidate preference. Whatever the kind, the metric must emit natural-language feedback that diagnoses why an output failed and what to change — not just a scalar. For API workflow benchmarks, record final-state correctness, policy compliance, data accuracy, endpoint discovery, required-write completion, forbidden-write avoidance, unnecessary calls/retries, schema validity, and recoverable errors as separate axes before collapsing to an overall score. If the metric or validator is unclear, stop and ask one concrete question.
4. Pass the evaluation evidence gates. Build and confirm the coverage matrix, run harness-conformance sentinels, and inspect actual scored rows before interpreting aggregate results. Follow references/evaluation-evidence-gates.md. An uncovered important stratum blocks a whole-workload conclusion; a failed read-then-write sentinel is a harness bug until proven otherwise. Keep a redacted local review packet with the coverage matrix, representative rows, counterexamples, scorer rationale, the data-sufficiency plan and stopping evidence, and exact artifact/log refs. Treat pilot sizes as minimums, never caps: when conclusions remain unstable or important strata are underfilled, collect more rows instead of increasing confidence in the prose.
5. Freeze splits. Write splits.json with train/dev/holdout names, sizes, source refs, deterministic split seed or frozen row ids, the per-stratum counts from the coverage matrix, uncovered strata, and an explicit "no holdout mutation" note.
6. Rerun the incumbent baseline. Use the frozen harness, metric, validator, and splits to rerun the current incumbent route. Write baseline.json with command, timestamp, split used, sample size, score, latency basis, cost basis if available, failures, and caveats. It must also include harness_sha256, metric_sha256, and splits_sha256 for the exact artifacts used by the rerun. Record the per-row (or per-cluster) pass/fail set, not just an aggregate score, so the next step can see whether optimization headroom exists — i.e. rows the incumbent fails that a stronger model could fix. Record each per-row result as an understudy.eval_result.v1 row — the required row format for eval evidence across every Understudy surface (schemas/understudy.eval_result.v1.schema.json): run_id, task_id, split, score (0..1 or null — a 0 is a scored failure, never a missing value), status (ok/error/skipped/unscored), model, route, cost/tokens/latency when known, and a provenance block whose harness_sha256/split_sha256 carry the same hash chain as baseline.json.
7. Run the trivial-agent floor. Alongside the incumbent, run a do-nothing agent — empty output, or for agentic workloads an immediate finish with no writes — through the same frozen harness, metric, and split, and record its score in baseline.json as null_floor. This is the Agentic Benchmark Checklist's trivial-agent baseline (uiuc-kang-lab/agentic-benchmarks); a do-nothing agent scores 38% on tau-bench, so an unmeasured floor makes every downstream number uninterpretable. If the null agent scores above ~5%, the metric or gold set is miscalibrated — fix it before freezing. Every later claim packet cites this floor: a savings or quality claim is invalid if the null agent also clears the bar.

Flow

Inspect the repo first to find where LLM calls happen and the current model/provider/harness/eval state, then surface that inventory before building anything. The inventory includes evidence that already exists outside the repo: the Understudy desktop app exports benchmark comparison packets (understudy.fusion_benchmark_comparison.v1) under ~/.understudy/exports/, each carrying understudy.eval_result.v1 rows plus a packet-level provenance block (rows in a sibling JSONL file; verify shasum -a 256 <provenance.eval_results_path> equals provenance.eval_results_sha256 before admitting it). Surface any verified packets in the inventory so the developer isn't asked to re-measure what the app already measured — the admission checklist lives in ../ramp-and-verify/SKILL.md. If the request/response path, dataset/trace shape, prompt purpose, or success criteria are not already clear, route to ../understand-workload/SKILL.md first and use its workload profile as the narrative source of truth. The deep inspection checklist (call sites by SDK family, env vars, tracing, CI) and the eval-harness discover-then-build playbook live in reference.md. For cross-cutting objective/constraint framing, read ../understudy/reference.md. For multi-turn / tool-use / agentic workloads — both read-only search loops and multi-step REST/API workflows that mutate state — route the eval to ../optimize-agentic-workload/SKILL.md instead of building a single-output harness; its state-mutating lens records reset/seed state, API schemas, policy docs, request logs, and final-state validators as part of the harness.

Start from the real local workload:

• app route, eval suite, trace export, benchmark fixture, prompt set, dataset, report, or existing runner;
• no traces yet? Start from a public benchmark — the golden-path ladder (AutomationBench, Harvey LAB) in references/public-benchmark-path.md runs this same evidence loop against public fixtures;
• otherwise, create only a synthetic fixture and label it clearly as synthetic.

Do not optimize, tune prompts, choose replacement models, mutate splits, or claim savings in this worker. Its job is to make the next validation step possible.

If any artifact cannot be created, write down the missing input, the attempted local command or inspection, and the next action most likely to unblock the decision, with its cost and scope.

Output Standard

End with:

• workload source inspected;
• workload profile status and whether the task understanding was confirmed;
• artifact paths created or refreshed;
• metric, validator, split boundary, and incumbent baseline status;
• result type: evidence-capture or blocked;
• one recommended next local command or action.