sl-complete

name	sl-complete
description	Complete a subjective-labeling corpus after sealed evaluation by freezing a validated production policy, running preflight and idempotent labeling, routing risky items to humans or unresolved status, reconciling one terminal disposition per item, and performing a probability-based final audit with provenance. Use for /sl-complete, production labeling, or corpus completion.

Complete the corpus

Apply only a production route that passed the sealed evaluation for the intended claims. Finish with one accountable disposition per corpus item and a human-audited quality statement.

Read first

Read:

• ../../ref/ref-contract.md
• ../../ref/ref-schema.md
• ../../ref/ref-stages.md
• ../../ref/ref-assets.md
• ../../ref/ref-architecture.md
• ../../ref/ref-cascade.md
• ../../ref/ref-output-style.md

Preconditions

Require valid G*, D_cal*, and T*; closed scorecards; a candidate that passes all required quality floors; and a declared remaining in-scope corpus. Otherwise stop with HOLD or select human-only completion explicitly.

Protocol

1. Freeze production manifest. Record corpus scope, policy and executor checksums, full routing rule, thresholds, abstention, risk queue, human capacity, budget, shard plan, retries, provenance tiers, and final-audit design.
2. Run preflight. Use a frozen probability sample to estimate route shares, risk queue size, quality linked to T*, cost, latency, and capacity. Do not change thresholds without creating a new manifest and preserving the old result.
3. Execute idempotently. Write every attempt append-only with item, run, policy, executor, wrapper, input checksum, output, reason codes, confidence, and timestamps. Resume failed shards without overwriting prior attempts.
4. Apply risk routing. Escalate declared disagreement, boundary, novelty, coverage, policy-mismatch, protected-stratum, duplicate-conflict, and drift conditions. Keep over-capacity or unresolvable items unresolved; never relabel them as NONE.
5. Record human review. Human judgments override machine outputs semantically and retain links to both attempts. Production review does not silently mutate G*; a material policy issue opens a new calibration/versioning decision.
6. Reconcile terminals. Create exactly one terminal disposition for every in-scope item: human-confirmed, audited-machine, machine-accepted, accepted-unresolved, excluded, or invalid. Report missing and duplicate terminal ids as hard failures.
7. Run final probability audit. Freeze sampling strata, seed, probabilities, blind human protocol, thresholds, and claims. Estimate weighted errors and intervals, inspect protected strata and route-specific failures, and document repairs.
8. Close or reopen. Create completed D* only when completeness, provenance, quality, and risk gates pass. Otherwise repair, re-audit, narrow claims, or reopen calibration with explicit versions.

Result

Return total scope; terminal count; labels and dispositions by provenance; unresolved, excluded, and invalid counts; route shares; human-review load; cost and latency; final audit estimate and intervals; repairs; protected claims; limitations; and artifact paths.

Embeddings, classifier confidence, and LM consensus are permitted only inside the exact validated route. If the production runner, reconciler, audit sampler, or provenance writer is absent, emit HOLD; do not fall back to legacy k-NN inheritance or panel majority.