/eval-harness-first

name: eval-harness-first
description: Build the evaluation harness that gates every fine-tuning run — golden sets, per-failure-mode graders, judge calibration, and base-model baselines. Use when starting a fine-tuning effort, when converting traces into an eval set, or when calibrating a judge against human labels.

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:

1. **Collect traces** — production/agent spans, or synthetic tasks if none exist yet. 2. **Error analysis** — open coding on ≥100 traces, axial coding into 4–8 failure buckets. 3. **One grader per bucket** — deterministic first; calibrated LLM-judge only for genuinely subjective criteria. 4. **Prioritize** by frequency × severity × value. 5. **The labeled traces feed dataset curation, minus an explicit holdout.** Every `eval/goldens.jsonl` ID stays excluded from training data by ID. 6. **Train.** 7. **Re-run the same harness** on the checkpoint — not a different, looser one. 8. **Drift detection feeds back to step 2** — new production failure modes re-open error analysis.

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

• **From traces, when they exist:** run error

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.

• **Synthetic, when traces don't exist yet:**

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.

• **Goldens are versioned like code** — commit

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

Graders

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

• **Deterministic first.** Regex, schema validation,

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

• **LLM-judge only for genuinely subjective

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

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

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

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

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`.

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.

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

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

• Report **TPR and TNR**, not one blended accuracy

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

• **Pin the judge to a fixed model snapshot** and

recalibrate on judge-model change, quarterly regardless.

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

than the model under test.**

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

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

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-<model>.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:**