adaptive-quality-gate

Files (2)

• SKILL.md
• evals.json · evals/

SKILL.md (first lines)

# Adaptive Quality Gate

Set quality gates from the task's outcome and risk. Do not reuse a domain
checklist merely because it worked for a previous scraping, SEO, coding, or
content task.

## Core Contract

A gate is a decision rule backed by evidence. It is not a reminder to "check
quality."

For every material criterion, define:

- **Criterion**: the property that must be true.
- **Threshold**: the observable boundary between pass and fail.
- **Method**: how the property will be tested.
- **Evidence**: the artifact or result that proves the test ran.
- **Severity**: hard gate or advisory gate.
- **Failure action**: hold, repair and retest, request a decision, or accept a
  documented exception.

Apply gates before irreversible actions and before making a completion,
readiness, safety, or quality claim. Never convert missing evidence into a
pass.

## 1. Decide The Gate Level

Assess four factors:

- **Impact**: what happens if the result is wrong or incomplete?
- **Reversibility**: how easily can the action or artifact be corrected?
- **Uncertainty**: how much is unknown about inputs, requirements, or tools?
- **Exposure**: is the result private and provisional, or external,
  production-facing, regulated, or decision-driving?

Choose the smallest gate level that controls the material risk:

| Level | Use when | Typical gate size |
| --- | --- | --- |
| None | Trivial, easily reversible work with an obvious result | No formal table; perform a basic sanity check |
| Light | Low-impact deliverable with limited uncertainty | 1-3 focused checks |
| Standard | Material deliverable, multi-step workflow, or consequential claim | 3-6 hard or advisory gates |
| Strict | High-stakes, destructive, production, regulated, security-sensitive, or hard-to-reverse work | 5-8 gates plus independent or stronger verification where feasible |

For a `Standard` or `Strict` gate, state one concise sentence connecting the
chosen level to the material impact, reversibility, uncertainty, or exposure.
Do not expose a synthetic numeric risk score unless the task already has an
approved scoring model.

Use a formal gate when at least one of these is present:

- An external or decision-driving deliverable
- A factual, quantitative, completeness, freshness, or provenance claim
- A code, configuration, migration, release, or production change
- A destructive or difficult-to-reverse operation
- A multi-source data collection or transformation workflow
- Legal, medical, financial, privacy, security, or policy consequences
- An explicit request for QA, validation, acceptance criteria, audit, or a
  quality gate

Do not create gate theater for a simple explanation, brainstorming pass,
clearly provisional draft, or tiny reversible formatting change. State that a
formal gate is unnecessary when that decision is useful; otherwise just do the
sanity check.

## 2. Build The Task Contract

Before selecting checks, identify:

- Intended outcome and artifact
- Consumer or downstream action
- Explicit constraints and user-supplied acceptance criteria
- Material failure modes
- Unknowns and unavailable evidence
- The exact claim that would justify saying "done"

Use explicit requirements, specifications, policies, contracts, and user
constraints as the primary source of thresholds. Ask a concise question only
when the answer would materially change a hard threshold or authorize a risky
action. Otherwise, use a provisional gate and label the assumption.

## 3. Derive Task-Specific Gates

Select only dimensions that can change the release decision:

- **Correctness**: values, logic, calculations, or behavior match the source of
  truth.
- **Completeness**: required scope is covered against a defined denominator or
  inventory.
- **Evidence and provenance**: important claims trace to current, relevant
  sources or execution artifacts.
- **Functional behavior**: the artifact works in the target environment and
  important paths pass.
- **Safety and compliance**: destructive, security, privacy, legal, or policy
  constraints are respected.
- **Operational readiness**: backup, rollback, observability, ownership, and
  failure handling are adequate.
- **Usability and format**: the consumer can read, open, navigate, or act on the
  result as intended.
- **Freshness**: time-sensitive inputs and conclusions are current enough for
  the decision.

Do not automatically use every dimension. A gate that cannot affect acceptance
adds noise and weakens attention to the real risks.

### Threshold Rules

Prefer thresholds in this order:

1. Explicit user requirement, contract, policy, or specification
2. Deterministic test, schema, reference artifact, or authoritative source
3. Established project baseline, service level, or historical control limit
4. A clearly labeled provisional qualitative criterion with human review

Do not invent a percentage, sample size, zero-defect promise, or coverage
denominator. If the population is unknown, gate the inventory method and source
coverage instead of claiming an unsupported completeness percentage.

## 4. Place Gates At Decision Boundaries

From EthanYoQ/Skill-hub, MIT.

More from EthanYoQ/Skill-hub