The owner of the answers below is code: skills/herdr-foreman/classify/classify-report.sh and skills/herdr-foreman/classify/report_verdict.py ask the questions and label a report, and skills/herdr-foreman/foreman/report_gates.py decides and records gates. This file says how the pieces fit, what a gate obliges, and how the bands change. It restates no threshold, answer set or composition rule; those are the owners' contracts.

Labelling a Report

Input: one delivered report, or a round's reports through skills/herdr-foreman/classify/classify-reports.sh.
Output: one label per report, the JSON object the skills/herdr-foreman/classify/report_verdict.py docstring documents (verdict, per-question answers, the report's sha256, the question hash and the model id). The batch adds unannotated, one entry and reason per report that got no label.
The questions live in skills/herdr-foreman/classify/report-questions.json. Changing one changes every label's question hash; update the file's changed date with it.
Step 12 runs the batch, skills/herdr-foreman/classify/classify-reports.sh. Its exit contract is in that script's header. A failed classification lands in unannotated and is never fatal. Exit 2 is a usage error, and its output is never used as labels.
The deterministic checks that refuse one report before any call are in the skills/herdr-foreman/classify/classify-report.sh header.

Adapters

Jev (TypeSafe System One) is the default. It needs TYPESAFE_API_KEY (.env.example). The client is skills/herdr-foreman/classify/typesafe_client.py, shared with the evidence assessor (#472).
An unavailable Jev, or one whose client refuses the request (the report carries the key), produces no label. The report lands in unannotated and takes the reasoning path, a full read. No other classifier is asked.
--agent codex|claude|grok runs an LLM adapter for measurement (skills/herdr-foreman/classify/evaluate.sh). An LLM label carries no probabilities and never gates.

Report Gates

foreman report-gate-record --labels <classify-reports output> records the gate each label earns and prints {"schema_version": 2, "recorded": [...], "replayed": [...], "no_gate": [...]}: recorded holds each new gate with its report, level and reason; replayed a gate already on record for the same report bytes and classification; no_gate each report its label leaves ungated, with the reason. Exit 1 records nothing and names the cause on stderr. The owner computes the level from the label's probabilities, the pinned model and the bands alone; the label's own verdict and gate are never inputs. A report whose bytes changed since classification is refused: reclassify it.
block — the report cannot be accepted until foreman report-gate-clear --report <path> (--evidence <report> --reason <why> | --decision <attention id>) records why it does not block. close-member refuses an accepted ledger decision and record-report refuses an approved verdict while it is open. A needs_work or blocked decision is never refused by a block.
reread — the report cannot be gated at all until a full re-read is recorded with foreman report-gate-reread --report <path> --evidence <re-read report> --note <what it verified>. Both close-member and record-report refuse while it is open.
No gate — probabilities below the bands, an LLM label, an unpinned model, an unannotated report. The report is read and gated as it would be without a classifier.

Who Resolves a Gate

The owner reads who resolved a gate from records it already holds, never from the caller. The gated report's task and role come from the applied dispatch whose supervision enrollment binds it; a report no enrollment binds cannot be resolved until the enrollment is restored.

Worker — --evidence names a report whose current bytes supervision observed (report_observed) or recover-report recovered after the gate was recorded, for another applied dispatch on the same task in the gated report's role. It clears a block or records a re-read.
Judge — the same delivery evidence, for an applied dispatch of the pinned judge (config.json judge) in adjudication mode on the same task. It clears a block; it never records a re-read.
Operator — --decision names an attention decision on the same task, resolved with the operator's user_answer after the gate was recorded. Its answer summary is the recorded reason; --reason is refused.
report-gate-clear resolves only block gates and report-gate-reread only reread gates. A command with no open gate of its level refuses and names the other command; a report carrying both stays unaccepted until both are resolved.
Anything else records nothing and names what is missing. The foreman records a resolution and never decides one.
A gate never approves, accepts or skips a check. A recorded clear or re-read undoes it.
foreman report-gate-status [--report <path>] lists open and resolved gates.

Verdict Gates

The same store holds a second gate source (#646): a report whose owner-parsed VERDICT: line is blocking. No label is involved and no label ever touches one.

assess-specialist (a replay included) and record-report record it, source: verdict, level: block, with the dispatch the verdict was recorded against. One gate per (report, bytes): a replay returns the existing gate whatever its status; new bytes record a new gate.
It never refuses acceptance: close-member and record-report ignore it.
apply refuses a fresh release dispatch on its task while it is open, dry runs included. A completed release replay is exempt.
report-gate-clear clears it with --decision as above, or with --evidence naming the gated responsibility's approved re-check. A judge's report is refused: a ruling decides, and the re-check that cites it clears. The re-check predicate is _rechecked's docstring.
A clear on a report carrying both sources resolves every pending gate by its own source's rule, or none.

Sidecar schema 2

skills/herdr-foreman/foreman/report_gates.py owns <canonical selected state>.report-gates.json: {"schema_version": 2, "state_path", "gates": [...]}. Each gate carries schema_version, source (classifier|verdict), report (resolved, canonical path), sha256, level, reason, dispatch, at, status (open|cleared|reread) and resolution (null, or schema_version, at, action, by (worker|judge|operator; worker|operator for a verdict gate), reason, evidence (attention for an operator, or path, sha256 and dispatch)). A classifier gate also carries probabilities, model, question and bands, and dispatch is null. A verdict gate carries none of the four, is always block, and names its dispatch. Every record is validated whole on every read. Writes take the sidecar's own lock. close-member and record-report hold that lock from their gate check through their commit, so a gate recorded meanwhile is refused rather than slipped in between; record-report writes its verdict gate under that same lock. A missing file is first use; an unreadable or unsupported one, or a symlink at its path, refuses every reader, never reading as no gates. A classifier replay is the same report bytes under the same model, question, bands and level; any other classification of those bytes records a new gate. close-member reads it and never writes it.

Schema 1 held classifier gates alone, without source or dispatch. Every read goes through the owner, so the first read of a schema-1 document migrates it: each gate becomes source: classifier, dispatch: null, its resolution's schema_version becomes 2, and the document is rewritten under the sidecar lock before it is returned. A read while another process holds that lock is refused and retried, never served unmigrated. A reader at schema 1 meeting a schema-2 document refuses it as unsupported rather than reading it as no gates. The sidecar takes rules/stateful-artifacts.md Migration Policy's gate-store exception: an open gate refuses an action, so "no usable prior state" would read as no gate.

Changing the Bands

The live bands are the named constants at the top of skills/herdr-foreman/foreman/report_gates.py, with BANDS_VERSION naming their calibration. They ship uncalibrated and conservative. Calibration proposes; a reviewed pull request installs:

Export TYPESAFE_API_KEY in the shell that runs the evaluation.
Run bash skills/herdr-foreman/classify/evaluate.sh --agent jev --results <labels.json>.
Run python3 skills/herdr-foreman/classify/scoring.py calibrate <labels.json>. It prints proposed bands and the counts it used, and writes nothing. It refuses to propose from too few held-out labels; its held-out test and selection rules are in the skills/herdr-foreman/classify/scoring.py docstring and top-of-file constants.
Open a pull request that changes the constants and BANDS_VERSION, with the proposal's counts and the resulting false-gate counts in its CHANGELOG entry.

Recalibrate after every JEV_MODEL bump and every question change.