Installs into .claude/skills of the current project.
Are you the author of Gz Advisor Qc?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/tvproductions-gz-advisor-qc)
---
name: gz-advisor-qc
description: Use between gz content compose and corpus attestation to judge a candidate rendition and record its advisory QC receipt.
category: agent-operations
lifecycle_state: active
owner: gzkit-governance
last_reviewed: 2026-10-17
metadata:
skill-version: "0.2.2"
model: sonnet
gz_command: gz content advise-rendition
---
# 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:
```bash
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
```bash
# 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."
```