Jevify

Jev is an optional classifier, not evidence of correctness or permission to edit. An explicit Jev request still needs a suitability decision; explain a better method when deterministic tools or direct reading answer the actual question more reliably.

Choose the method

• Prefer a compiler, search, AST rule or short script when a rule decides the answer.
• Prefer direct reading for small jobs or tightly cross-dependent reasoning.
• Consider Jev for repeated, independently judgeable semantic decisions with the

necessary shared context. Repetition or list length alone is not a reason to use it.

• Loading this skill or its helper does not require a judge call. If choosing Jev,

pilot before relying on results, including for read-only answers. A pilot may also show that a direct method is the better choice.

Optional Python helper, loaded in an eval cell of its own (combining %load with other code in one cell fails):

%load skill://jevify/jevify.py

Read [the on-demand reference](skill://jevify/reference.md) only for needed helper APIs, extraction details or existing recipes; it is not a second mandatory workflow.

Source-backed evidence preparation

For review comments, documentation promises, migrations, or issue/log intake, use jv.prepare_review, jv.prepare_documentation, jv.prepare_migration, or jv.prepare_issue with explicitly selected sources and bounded reference resolution. Structured fields stay deterministic. For prose, jv.extraction_request(plan) returns arguments for the jevify_extract tool, or None when extraction is unnecessary. Clairic defaults on for these selected inputs; --no-jevify-extract opts out. Selected private source leaves the machine for OpenRouter; exclude secrets. Credentials come only from the host resolver. Transport currently requires Linux; deterministic Python preparation remains available without a key or provider. Failure is unresolved evidence, never an invented candidate. When a result path exists, adopt it with jv.adopt_evidence(plan, result_path); otherwise keep the deterministic plan and its unresolved fields. An exact occurrence is not proof of a finding.

Use jv.workflow_review(evidence, Q) for ready-only per-question batches. Judge each batch with its own Q and extraction=batch["manifest"], then open the existing inbox with that manifest, the bound rows, and evidence=evidence. Missing base versions, contracts, wrappers, or related findings block dependent questions, not the scope ledger. Open the supporting sources before confirming. An unresolved sibling question prevents whole-record approval. Changed source/dependencies require explicit refresh; imported snapshots cannot be deleted. Explicit external imports preserve attribution, not authority. See the reference for the full workflow and interchange contract.

Live preparation experiments require separate approval of the frozen manifest digest and spending cap; default-on extraction is not experiment approval. The controlled runner compares Clairic plus deterministic preparation against the same deterministic preparation without Clairic. Synthetic/injected demonstrations never establish gains.

With the host notice

OMP 18.12.5's jevify-notice fires on the word "jevify" and directs bulk work to the built-in judge_batch() for about 20 or more items; it never mentions jv.*. Its item count is not a reason to use Jev: choose the method above first. If judge_batch() already ran, bind its output with jv.adopt(items, states, Q, batch=b) (drained items keep their served model; batch= adds batch ID, cost estimate and host counts), then continue with jv.flag, jv.pilot_check, review and jv.export.

If choosing Jev

1. 1Declare the question, source scope and unit boundaries. Define mutually exclusive labels, including uncertainty, against a concrete contract before judging data. Set escalation/action thresholds before the pilot; do not tune them to obtain desired outcomes. Scores and threshold estimates are not correctness guarantees. Default escalation: top choice probability below 0.7, unknown labels, and bool probabilities in 0.3–0.7; name additional flag labels in the rubric up front.
2. 2Check extraction coverage and context. Map independent source hits to units; matching totals alone do not prove coverage. Account for filters, dropped/empty states, truncation and oversize inputs. Include relevant helpers, siblings and consumers; sanitize logs and never submit credentials or keys. jv.plan checks full-request budgets; jv.pack groups without silent truncation. Bind source hashes and exclusions with jv.extraction before relying on coverage. Settle script-decidable fields with jv.facts (requires= warns when a unit lacks what a question asks about). Refuse elided reads (jv.refuse_elided, jv.units_from_table); pass the jv.states report to jv.flag(rows, report=...) so truncated or oversize units are flagged.
3. 3Pilot the chosen representation. Read flagged sources and near misses, compare verdicts with evidence, and repair misread wording or missing context. Rejudge the pilot after changes; abandon Jev if the judgments remain unsuitable. Run jv.pilot_check(rows, Q, rule=...) with the frozen rule: use_direct (flag rate, uncertainty band or catch-all share above about 40%) means stop and read directly or script it. jv.remaining reuses unchanged pilot verdicts in the sweep.
4. 4Freeze the rubric/questions and preserve their hash for full and after-action runs. A post-freeze wording change invalidates prior verdicts: re-freeze and rejudge the affected declared sweep, not just the flags.
5. 5When doing a sweep, judge every unit in the declared scope and reconcile every missing result, including originals dropped before grouping. A pilot is not complete coverage. Check the run's bound receipt as well as row flags; an empty flag list cannot prove completeness. Retain errors, fallback routes and unknowns as unresolved, never as safe labels; rerun or inspect their sources.
6. 6Persist the frozen questions/hash, source-to-unit mapping, rendered states and group slots, batch IDs, returned rows and receipt before discarding evidence. Use bundle_path for checkpointed jv.run recovery and jv.resume after a kernel loss. Recovery reconciles saved items and a surviving host batch; stale or missing batches remain unresolved unless resubmission is explicitly requested. This is not exactly-once execution; see the reference for crash and host-session limits.
7. 7A jv.splice of re-judged rows into an earlier run is never a full sweep; report it as spliced. Compare sweeps with jv.gate, which separates regressions from judge noise.

Confirm before action

• Read every flagged source in full, not just a truncated render or score; confirm

or overturn the claim from reachable behavior and the governing contract. Escalate uncertainty, near-threshold ambiguity, errors, explicit fallback rows and unknown provenance. Stock OMP (audited 18.8.7–18.9.1 and 18.12.5) exposes only a served model per item and model/cost in status(), no selected/primary, fallback or cache fields, so provenance_unknown stays true on every row. Cached repeats report the alias ~typesafe/jev-latest, fresh answers a versioned model, so such rows are not comparable; a familiar model label is not a substitute for provenance.

• Record source-bound confirm, overturn or unresolved decisions in the review inbox

(jv.review_queue, jv.review_questions, jv.adjudicate, jv.review_status), using a durable review path. approved_units lists units whose items are all confirmed and not stale; deletion additionally requires a settled review. The status always reports grants_authority: false. Stale approvals are reported under stale; decide them again rather than carrying them forward.

• Calibrate before bulk edits or destructive action; read-only calibration is

optional. Model agreement is an estimate, not source confirmation or authority. Poor agreement calls for revisiting the pilot/context, not relaxing thresholds.

• Grouping is a separate instrument, not an interchangeable cheaper flat run.

Pilot grouped judgments separately; calibrate them separately before bulk or destructive action. Do not mix grouped and flat rows in one decision.

• Act only on confirmed items within granted authority. Review authority is not

deletion authority. For deletion, call jv.delete_spans with the manifest, the review path and Q, dry-run first, and inspect the printed spans. Preserve dirty/unrelated work, and use LSP references for exported-symbol refactors. Delegate exact confirmed IDs for edits rather than asking editors to rediscover candidates.

• After edits, re-extract and remap coverage, then rejudge with the same frozen

questions. Unresolved flags/errors/unknowns or uncovered hits prevent a clean closure claim. Receipt completeness is independent of row flags and source coverage; a zero-unit extraction is not a model verdict.

Report honestly

Report the chosen method and declared scope; extracted/covered, dropped/truncated, piloted/judged and manually reviewed counts; errors, models/fallbacks, flags, calibration if performed, receipt completeness, review status (pending, unresolved, stale, settled) if an inbox was used, and changed/remaining items. Separate judge estimates from source-confirmed conclusions. Say when the result is pilot-tuned or sampled, not a full read; do not claim complete coverage or successful action without evidence.