Skip to content
Back to skills

Gz Advisor Qc

ASecurity

Use between gz content compose and corpus attestation to judge a candidate rendition and record its advisory QC receipt.

  • 9 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added October 19, 2026
ai-agentsgobash

Security analysis

A100/100

Scanned October 19, 2026

npx -y skills add tvproductions/gzkit --skill gz-advisor-qc --agent claude-code

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.

Security grade badge for Gz Advisor Qc
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/tvproductions-gz-advisor-qc/badge)](https://www.skillsdirectory.com/skills/tvproductions-gz-advisor-qc)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

SKILL.md
---
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."
```

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.