day0-release

Day-0 Release

Drive a model from a pretrained checkpoint to a publish decision for a quantized checkpoint, in a fixed sequence with a gate after every stage. This skill is a conductor: it sequences the existing domain skills and enforces the gates — it does not re-implement quantization, serving, evaluation, or comparison.

Goal (the default day-0 criterion): a quantized checkpoint smaller than the source, with accuracy drop within the threshold (default <1%) on the standard benchmark set versus the matching baseline, plus a publish recommendation.

When to use

Use only for the full goal-driven release. For a single stage, route to the domain skill directly: quantize → ptq, serve → deployment, evaluate → evaluation, compare two existing runs → compare-results.

Inputs

Resolve these before starting (ask the user for anything missing):

• Model — HF handle or checkpoint path.
• Recipe / qformat — e.g. nvfp4, fp8, or a recipe path. One candidate for v1.
• Cluster / launcher — from clusters.yaml (see the common skill's environment-setup.md).
• Eval set — defaults to the evaluation skill's AA suite (recipes/tasks/aa/).
• Threshold — max accuracy drop; default 0.01 (1%).

The chain

setup ─▶ PTQ ─▶ canary ─▶ baseline-eval ─▶ quantized-eval ─▶ compare ─▶ verbosity ─▶ closeout
          │        │           │                │              │           │
       gate_ptq  /health    gate_run         gate_run    gate_compare  gate_verbosity
                 + 1 gen

The evaluation skill deploys the model it evaluates (it stands up its own endpoint per run), so there is no separate deploy stage — a serving failure during evaluation surfaces through the eval gate (DEPLOYMENT_HEALTH_FAILED) and triages to the deployment skill (see Step 4). The Step 2b canary is not that: it runs before any evaluation precisely so an unservable checkpoint is caught in ~15 min rather than after a multi-hour eval.

Accuracy (Step 5) and verbosity (Step 5b) are independent gates; closeout requires both.

Run each stage by invoking the domain skill, then run its gate before proceeding. Do not advance past a failed gate. Copy this checklist and track progress:

- [ ] Step 0: Resolve inputs; confirm threshold and eval set
- [ ] Step 1: Setup gate — creds present, cluster reachable
- [ ] Step 2: PTQ (ptq skill) → gate_ptq.py

Step 1 — Setup gate

Use the common skill's credentials.md and remote-execution.md to confirm credentials and cluster reachability. If either fails, stop with SYSTEMIC — do not start PTQ.

Step 2 — PTQ

Invoke the ptq skill to produce the quantized checkpoint. Then gate:

gate_ptq.py returns JSON {pass, failure_class, detail}. On pass: false, branch on failure_class (see Triage below). Do not evaluate an unvalidated checkpoint.

Step 2b — serving canary (MANDATORY before Step 3)

The canary itself is already specified by the ptq skill: see ptq/references/checkpoint-validation.md (required gate, canary query and the Serving canary row of its report table) and ptq/SKILL.md. Run it there rather than re-deriving it here — gate_ptq.py checks size, coverage and metadata, not whether the checkpoint loads, and skipping the canary has cost a full baseline eval against a checkpoint the serving stack could never load.

On failure use failure_class: CHECKPOINT_NOT_SERVABLE and drop to the deployment skill; do not proceed to Step 3.

What that spec does not cover: writing a canary that cannot lie. Both of these produced a wrong verdict on a large MoE, in opposite directions:

• Poll ceiling > load time, with headroom. A 50 min poll against a 51 min load reported failure for a checkpoint that serves fine. Large MoE loads are CPU-bound fp8 dequant (~50 min); 0% GPU during load is normal.
• Print an explicit RESULT: on every path and exit non-zero on failure. A poll loop that falls through to the generation test exits 0 and reads as PASS.
• Server log on shared storage, not node-local /tmp.
• Canary the as-exported artifact, not a copy you modified to make it work.

Step 3 — Baseline eval

The baseline is the source (pre-quantization) model on the same task set and sampling params. Always run a fresh baseline via the evaluation skill, which deploys the source model itself. Gate with gate_run.py.

Step 4 — Quantized eval

Invoke the evaluation skill on the quantized checkpoint, matching the baseline's task set and sampling params. The evaluation skill stands up the serving endpoint itself (it builds the deployment.command, e.g. a vllm serve …), so a serving failure surfaces here as a failed gate_run.py with DEPLOYMENT_HEALTH_FAILED. When that happens, drop to the deployment skill to reproduce and debug serving in isolation (serve the checkpoint standalone, confirm /health + one generation, iterate on flags / TP / image / env vars) rather than burning full eval cycles on a broken endpoint — then carry the working command back into NEL's deployment.command and resume the eval. If the checkpoint genuinely can't serve, POINT_INFEASIBLE.

Before submitting: assert baseline/candidate config parity. The candidate config must differ from the baseline's in nothing but checkpoint path and served model name. Diff mechanically — an eyeball pass misses this:

Any other difference biases the comparison and invalidates the gate: a mismatched parallelism between the two sides was worth ~2 pp, enough to invert the sign of the delta. If the model card splits sampling params per scenario, apply the same split to both sides. Never set an unbounded request_timeout (1e9) — it turns a transient stall into a job that holds its GPUs until the wall clock kills it.

Gate:

A pass: false here means the run is incomplete or invalid (judge/parse error, dropped samples) — do not compare scores from it.

Step 5 — Compare

Invoke the compare-results skill. It must perform the shared external baseline sanity check before the candidate-delta gate. A failed check is ANOMALOUS with failure class EXTERNAL_BASELINE_MISMATCH: investigate and rerun the baseline. If no credible comparable external score exists, record the baseline as externally unverified and continue using the validated measured baseline.

Statistical power — check BEFORE trusting any per-task verdict

A task whose measurement noise rivals the threshold cannot decide a gate. Confirm each task's repeat count gives a standard error below the threshold; otherwise mark it INDETERMINATE rather than reporting a pass/fail.

For example, on a 1 % gate on DeepSeek-V4-Pro, both tasks that originally failed passed once measured properly:

Add precision by submitting the benchmark more times, not by raising num_repeats within a run. SciCode's floor is 8 runs per side, pooled as a mean — recipes/tasks/aa/scicode.md.

Re-running does not guarantee fresh samples. With a warm NEL response cache a "re-run" can replay cached responses — two runs came back bit-identical to 16 digits. But an equal score is a signal to check provenance (invocation id, output dir, response artifacts), not proof of replay: discrete-scored tasks tie legitimately. Confirm replay before discarding a run as a duplicate.

After recording the external status, produce per-task deltas and run:

The threshold is a fraction of each task's score scale. Most AA tasks report 0-100, but some (e.g. tau2_bench_telecom Result) report 0-1; the gate infers each task's scale (0-1 if both scores are within [0, 1], else 0-100) and normalizes the drop accordingly, so --threshold 0.01 means "≤1 pt on a 0-100 task / ≤0.01 on a 0-1 task" uniformly. Pass --scales '{"task": max}' to override inference if a task's scores happen to fall in an ambiguous range.

gate_compare.py checks only the candidate delta; it cannot override a failed external baseline check. Combined decision:

• ACCEPT (accuracy only) — no external check failed and every task is within the candidate threshold → continue to Step 5b. This verdict covers accuracy alone; advance to Step 6 only once the verbosity gate also passes. A missing comparable external score is not a failure; report it as externally unverified.
• REGRESSION — one or more tasks exceed threshold. v1 stops here and reports which tasks regressed by how much. (Picking the next recipe and re-running is deferred — see Scope.)
• ANOMALOUS — external baseline sanity failed, or scores are otherwise implausible (e.g. baseline lower than candidate by a large margin, or a task score is outside its valid range) → correct the baseline or surface it.

Step 5b — Verbosity gate (MANDATORY; independent of accuracy)

gate_compare.py does not measure verbosity, so stopping after Step 5 leaves a hard gate unmeasured. Step 5's ACCEPT is accuracy-only — Step 6 requires both.

Exit codes follow the other gates: 0 pass, 1 the gate failed, 2 it could not read its input. A 2 means fix the invocation, not the checkpoint — check --baseline/--candidate, --glob (no_files_matched names the pattern that missed), and --exclude. If the detail names collapsed_keys, the artifact tree gave several tasks one name: point the gate at a tree whose task dirs are <harness>.<task> rather than <invocation_id>.<job_index>, or re-sync so each artifacts/ carries its own config.yml.

Read response_stats.avg_completion_tokens from each task's artifacts/eval_factory_metrics.json. Do not use the reasoning.* fields: reasoning.*_tokens are always 0 (only the reasoning/content split is missing) and the *_words siblings are a proxy that can disagree with the gate — one task read +6.10% FAIL in words and +1.32% PASS in tokens.

The gate is two-sided; a large drop in output length is also a change.

Two filters are mandatory: same reasoning effort (pass --exclude explicitly — it is empty by default, since which tier is canonical is run-specific) and complete runs only — runs within 1% of the largest successful_count that both sides can match (exact equality would make one dropped sample unmeasurable). A task with no such count on both sides is not_comparable. When the matched count is below a run one side has, the task carries truncated_comparison.

Tasks sharing no sample count are not_comparable, not a delta. Means under ~1000 tokens carry a short_output_warning — read absolute counts there.

Step 6 — Closeout

Report the decision with: source vs output size + ratio, per-task baseline / candidate / delta / within-threshold, the verbosity verdict from Step 5b, external source and sanity status, MLflow run IDs, and a publish recommendation (publish / do-not-publish). Archive artifacts to the workspace.

Publish exactly what was evaluated. Verify mechanically, not by path convention: inode-compare a shard in the evaluated directory against the one being published. A day-0 run leaves near-identical sibling exports that differ by one calibration suffix — prefix rejected ones REJECTED- so the artifact cannot be picked by autocomplete, and re-run the Step 2b canary against the final path after any move.

Triage (gate failure → decision)

Map a gate's failure_class to the next action:

gate_ptq.py also emits non-blocking notes (present on every result). Size growth is blocking by default, waived to a note only when the validation summary's source_precision is already at or below the recipe's target bits and the growth is within what that explains. Recording source_precision is part of the ptq skill's validation table, so the waiver is reachable from the normal pipeline. A BF16 source that failed to compress still fails, which is the case this check exists for.

SYSTEMIC (cluster down, dataset unavailable) aborts the whole run. POINT_INFEASIBLE means this (model, recipe) can't work as configured.

Output

Return a decision, not a raw artifact:

• ACCEPT + report + publish recommendation
• REGRESSION + which tasks failed the threshold and by how much
• ANOMALOUS / INFEASIBLE + reason and next automated action
• Always: workspace path + MLflow run IDs for traceability

Scope (v1)

In v1: the linear chain + gates + report. On REGRESSION, v1 reports and stops. Deferred to a follow-up: the evaluator-optimizer recipe loop (compare → pick the next recipe → re-run PTQ), which needs the bigpareto integration and a shared config/result schema.*