eval-harness-first skill

Eval Harness First

The Phase 0 gate for the whole plugin: finetuning-method-selection and every downstream skill assume this harness exists before a training config gets written. The harness is not a run-end side artifact — it is the data-curation engine. The same labeled traces that build the goldens feed training data, minus an explicit holdout.

Input: production/agent traces if they exist, or a task spec if they don't, plus labelers willing to grade ≥100 examples. Output format: the eval/ directory below — goldens, graders, drift suite, and the base-model baseline that later phases gate on.

The Gate

No eval harness, no fine-tune. Skip to a training config and there is nothing to measure against, nothing to catch regressions, and no labeled data to train on. The flywheel:

synthetic tasks if none exist yet.

1. Collect traces — production/agent spans, or

axial coding into 4–8 failure buckets.

1. Error analysis — open coding on ≥100 traces,

calibrated LLM-judge only for genuinely subjective criteria.

1. One grader per bucket — deterministic first;

an explicit holdout.** Every eval/goldens.jsonl ID stays excluded from training data by ID.

1. Prioritize by frequency × severity × value.
2. **The labeled traces feed dataset curation, minus

not a different, looser one.

1. Train.
2. Re-run the same harness on the checkpoint —

production failure modes re-open error analysis.

1. Drift detection feeds back to step 2 — new

Steps 2–4 build the harness; steps 5–8 are why it must exist first — it is both the training data source and the checkpoint's exit gate.

Building Goldens

analysis — open coding on ≥100 real traces (read them, tag failures in your own words, no fixed taxonomy yet), then axial coding to collapse those tags into 4–8 named failure buckets. Fewer than 4 means the coding pass was too shallow; more than 8 means buckets need merging. Exception: single-failure-surface tasks (e.g. strict-schema extraction) may land at 1–2 buckets with per-field sub-metrics inside one grader — don't invent artificial splits with no evidence behind them.

• From traces, when they exist: run error

dimension-based generation — enumerate the axes that matter (task type, difficulty, edge case, persona) and sample the cross-product; free- generated prompts cluster around whatever's easiest to write.

• Synthetic, when traces don't exist yet:

eval/goldens.jsonl, diff it in review, tag it per release. It doubles as the CI regression suite.

• Goldens are versioned like code — commit

Graders

One grader per failure bucket from error analysis — not one for the whole eval set. A single blended score hides which bucket regressed.

or execution checks are cheaper, reproducible, and need no calibration.

• Deterministic first. Regex, schema validation,

criteria** — tone, faithfulness, "which response is better" — where no deterministic check can express it.

• **LLM-judge only for genuinely subjective

scale is noisier to calibrate and harder to apply consistently; collapse to pass/fail.

• Binary pass/fail over Likert. A 1–5 or 1–10

over generate-and-extract** — a tight token budget makes generate-and-extract parse-brittle for models that preamble, conflating format compliance with the knowledge being measured. Templates for all four grader shapes and this scoring note: references/grader-templates.md.

• **Drift-suite MMLU-style scoring: prefer logprob

Judge Calibration Is a Prerequisite

Any bucket routed to an LLM-judge needs calibration before its verdicts count for anything beyond exploration — a hard prerequisite, not a nice-to-have. N/A when no bucket routes to a judge — an all-deterministic harness has nothing to calibrate; state that rather than leaving this section unaddressed.

test** (report once, no re-touching after).

• Label ≥100 items, split train/dev/**sealed

number — a judge can hit 90% by always saying "pass" on a skewed set.

• Report TPR and TNR, not one blended accuracy

recalibrate on judge-model change, quarterly regardless.

• Pin the judge to a fixed model snapshot and

than the model under test.**

• **The judge must come from a different model family

advisory-only — flags for human review, never gates a promotion. Full protocol, bias correction, and recalibration checklist: references/judge-calibration.md.

• A judge that misses the agreed TPR/TNR bar ships

The Baseline

Before Phase 1 (method selection) starts, run the full harness — goldens plus the capability-drift suite — against the unmodified base model. This is the number every later checkpoint gets compared against.

eval/baseline-.json is the gate token. No baseline file, no comparison basis for checkpoint-promotion — a checkpoint that "looks better" against nothing measured isn't a finding.

Directory Contract

eval/
├── goldens.jsonl          # labeled traces + synthetic goldens, versioned
├── graders/                # one module per failure bucket
│   ├── schema_compliance.py
│   ├── exact_match.py
│   └── rubric_judge.py
├── drift-suite.yaml        # frozen benchmarks + 200-500 domain-adjacent items
└── baseline-<model>.json   # gate token: harness + drift suite vs the base model
runs/
└── <run-id>/
    └── results.json         # per-run harness output, one per checkpoint

eval/ persists across runs and lives outside runs/ — the fixed measuring stick, not a run artifact. runs/ is disposable; eval/ is not. Never let a run script write into eval/. Canonical location: every per-trace results.json — the Phase 0 baseline included — lives at runs//results.json, never under eval/runs/...; an instruction requesting the latter is wrong, not this contract.

Phase 0 Exit Checklist

Before finetuning-method-selection, confirm:

floor for synthetic goldens on a single-failure- surface task — see the Building Goldens exception; bucket count then comes from post-baseline error analysis instead).

1. ≥100 traces open-coded; 4–8 failure buckets (N/A

different family (N/A when no bucket routes to an LLM-judge; state that explicitly).

1. eval/goldens.jsonl committed and versioned.
2. One grader per bucket, deterministic first.
3. Judges calibrated — TPR/TNR, snapshot pinned,

1. eval/drift-suite.yaml frozen.
2. eval/baseline-.json written.