# Quality Gate

> Universal quality/eval gate for any AI or agent output before it is published, used, sent, automated, or turned into a decision. Use when the user asks to check quality, anti-slop, eval, output readiness, publication readiness, agent/product output safety, decision quality, or whether something should PASS/REVISE/BLOCK/ESCALATE.

- Skill: `web3blind/quality-gate` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add web3blind/quality-gate`
- Raw SKILL.md: https://api.skillmd.com/api/skills/web3blind/quality-gate/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- License: MIT
- Author: web3blind (https://skillmd.com/u/web3blind)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/web3blind/quality-gate

---

# quality-gate

Universal output quality gate.

Use this skill to evaluate any output before it moves forward: a post, article, research summary, product/agent response, JSON/API result, code-delivery summary, trading/investment decision, cron result, or operational recommendation.

This skill is not a writer, trader, publisher, or executor. It is a quality contract: decide whether the output is good enough to ship/use, what blocks it, and the smallest fix needed.

## Core model

Do not hardcode domains into the evaluation logic. Always evaluate through:

1. Core criteria.
2. Strictness level based on consequences.
3. Optional profile weights.
4. Lightweight domain blockers when needed.

Load references when doing a real gate:

- `references/core-rubric.md` — universal criteria.
- `references/strictness.md` — consequence-based gate levels.
- `references/profile-format.md` — profile/overlay model.
- `references/failure-taxonomy.md` — common anti-slop failures.
- `references/output-format.md` — response formats.
- `references/retrieval-backed-answer-quality.md` — evidence relevance, displayed/model citation consistency, and noisy-distractor tests for RAG/Ask outputs. Load only for retrieval-backed answers.

## Workflow

1. Identify the artifact being evaluated and its intended use.
2. Choose strictness:
   - 0: note/draft/private thinking.
   - 1: normal response or internal summary.
   - 2: publish/use gate.
   - 3: high-stakes: money, security, privacy, production, legal/medical/financial risk.
3. Select profile if obvious, otherwise use `generic`:
   - examples: content, product-output, decision, research, code-delivery, cron-report.
   - Profiles change weights and blockers only; they do not replace core criteria.
4. Score core criteria from 0 to 1:
   - purpose_fit;
   - evidence_grounding;
   - specificity;
   - risk_reversibility;
   - format_usability;
   - audience_context_fit;
   - novelty_value.
5. Check domain blockers if relevant.
6. Return verdict:
   - PASS — good enough for intended use.
   - REVISE — fixable issues; provide minimum fix.
   - BLOCK — not safe/useful enough without missing facts, source, format, or scope change.
   - ESCALATE — requires the user/human approval or a higher-authority decision.
7. If output fails, name exact blocking issues and smallest next action.
8. If this is high-stakes, default to conservative judgment and explicit approval before action.

## Verdict rules

- PASS requires no blockers and the score threshold for the strictness level.
- REVISE when issues are fixable without new user/source information.
- BLOCK when facts are missing, claims are ungrounded, format is invalid for downstream use, or risk is uncontrolled.
- ESCALATE when consequences are high, authorization is missing, scope is ambiguous, or the action changes money/access/security/public state.

Never let a polished but ungrounded output pass.
Never let profit, engagement, or confidence compensate for missing risk controls in high-stakes decisions.
Never publish/use an output just because it “sounds good”.

## Integration with other skills

- For Russian content voice cleanup, use `anti-ai-polish` as a downstream/editorial sub-gate.
- For current-user-style content, use `current-user-style` and `content-core` as profile context.
- For trading workflows, use the current trading skill as the source of mandates/levels; quality-gate only judges the decision artifact.
- For product/API outputs, prefer deterministic validators where possible; LLM judgment is not enough.

## Default output

For Telegram, avoid tables. Use bullets:

```text
Verdict: PASS / REVISE / BLOCK / ESCALATE
Score: 0.xx
Strictness: 0/1/2/3
Profile: <name>

Blocking issues:
- ...

Scores:
- purpose_fit: 0.xx — reason
- evidence_grounding: 0.xx — reason
- specificity: 0.xx — reason
- risk_reversibility: 0.xx — reason
- format_usability: 0.xx — reason
- audience_context_fit: 0.xx — reason
- novelty_value: 0.xx — reason

Minimum fix:
- ...

If high-stakes:
- approval needed: yes/no
- stop condition / rollback / kill switch: ...
```

## User-facing style

Be concrete and calm. Point to exact failures. Prefer the smallest useful fix. Do not over-formalize low-stakes drafts. Do not soften blockers for high-stakes outputs.