gz content advise-rendition (advisor-QC)

Overview

You are the advisor panel in the ADR-0.0.37 corpus pipeline (corpus → compress → advisor-QC → operator attest → committed rendition → playback). After gz content compose stages a candidate rendition, you read the candidate against its source corpus and judge how much information is retained per byte of the compressed result. You then record that judgment as a verdict via gz content advise-rendition.

This is an LLM-as-judge surface governed by ADR-0.0.39 doctrine:

• Advisory, never gating. A low retention score is evidence for the operator, not a fail-closed gate. The tool exits 0 for any score. You never block compose or commit on the verdict value.
• Explanation before verdict. Your reasoning is recorded before the score — state why first, then the number. The tool fails closed only on a structurally malformed verdict (an empty explanation), never on the value.
• The judgment is yours; the tool is deterministic. The gz content advise-rendition tool makes no LLM or network call. It validates receipt shape, writes the ARB receipt, and emits a ledger event. The information-retention read is your work, not the tool's.

Workflow

1. Identify the surface (e.g. AGENTS.md) and the consumer (e.g. root) whose candidate rendition you are judging. The candidate is staged at .gzkit/renditions/<surface>/<consumer>.candidate.md; the source corpus is at .gzkit/corpus/<surface>.jsonl.

2. Read both. For each compressible-tier corpus entry, ask: did the candidate retain the information (possibly compressed) or lose it? Invariant-tier entries must appear verbatim — their loss is not a low score, it is a compose-stage defect to surface.

3. Form a retention judgment as a ratio in [0.0, 1.0] — information retained per byte of the compressed rendition. High = dense retention; low = measurable loss per byte saved.

4. Write a concise explanation naming what was retained, what was combined, and what (if anything) was lost. This is the reasoning the operator reads before the corpus attestation.

5. Record the verdict:

   gz content advise-rendition <surface> [--consumer <vendor>] \
     --score <0.0-1.0> --explanation "<your reasoning>"
   

6. The verdict is written as an ARB receipt (arb-step-judge-<hash>) and witnessed by a rendition_advisor_verdict ledger event. Cite the receipt id in the operator's corpus attestation. It is not Gate 5, which names completed OBPI or ADR work only (.gzkit/rules/agents-md-map-doctrine.md § Attestation granularity).

Validation

• An ARB receipt is written under the configured receipts root (artifacts/receipts/ by default) named arb-step-judge-<32hex>.json, with exit_status: 0 and the explanation serialized before the verdict block.
• A rendition_advisor_verdict ledger event is emitted carrying surface, consumer, receipt_id, and score.
• A verdict with an empty/whitespace explanation fails closed: non-zero exit, no receipt, no ledger event.

Example

# After composing a candidate for AGENTS.md → root, judge and record the verdict.
gz content advise-rendition AGENTS.md --consumer root --score 0.94 \
  --explanation "All Mechanical bullets retained verbatim; two Promotable bullets combined without information loss; no Judgment content dropped."

# A low score is still recorded (advisory) — it is evidence for the operator, not a block.
gz content advise-rendition CLAUDE.md --consumer claude --score 0.55 \
  --explanation "One Judgment-tier rationale paragraph compressed past the point of recoverability; flagging measurable loss for operator review."