Quality Assurance Process

How do you prevent defects before generated work compounds them?

Quality assurance moves the critical decision and its verifier upstream. Quality control then checks the finished output against the same gates used for human work.

Method

1. Declare the contract. Name the authority, setpoint, evidence boundary, and owner before work begins.
2. Place the pause point. Put a short READ-DO check before the first expensive or irreversible action.
3. Automate the invariant. Use a schema, test, or guard for facts a machine can compare.
4. Run the work. Preserve professional judgment for local interpretation and exceptions.
5. Verify and evolve. Run the normal output gates, record variance, and correct the earliest durable owner.

Principles

Quality control is a reactive part of quality management focused on detecting problems after they have occurred. It involves activities such as testing, measuring, and inspecting to ensure specified quality standards are met.

Quality assurance is a proactive process that focuses on preventing defects from occurring in the first place.

Deming's Checklist

14 Point checklist to incorporate quality control principles into the manufacturing process. With a focus on creating systems that got manufacturing right the first time.

1. Create constancy of purpose for improving products and services.
2. Adopt the new philosophy.
3. Cease dependence on inspection to achieve quality.
4. End the practice of awarding business on price alone; instead, minimize total cost by working with a single supplier.
5. Improve constantly and forever every process for planning, production and service.
6. Institute training on the job.
7. Adopt and institute leadership.
8. Drive out fear.
9. Break down barriers between staff areas.
10. Eliminate slogans, exhortations and targets for the workforce.
11. Eliminate numerical quotas for the workforce and numerical goals for management.
12. Remove barriers that rob people of pride of workmanship, and eliminate the annual rating or merit system.
13. Institute a vigorous program of education and self-improvement for everyone.
14. Put everybody in the company to work accomplishing the transformation.

Generator QA

Deming's Point 3 — "cease dependence on inspection" — applies to automated output too. A generator that produces code, content, or artifacts must pass the same gates as hand-crafted work. If a human commit would fail the pre-commit hook, a generated commit must fail the same hook.

Principle	What it means
Same gates	Generated output runs through identical linting, validation, and hook checks as manual output
Fail fast	If the generator can't pass the gates, the template driving the generator is broken — fix the template, not the gate
No inspection bypass	"It works at runtime" is not "it ships." The gap between runtime success and commit success is where defects hide
Instrument calibration	The tools that measure quality must themselves be measured. A broken issue tracker means the feedback loop is open

The test: run lint-staged simulation on generated files. If any file fails, the generator template is at Process Maturity level D (draft), not A (approved).

Contract Preflight

Output checks catch defects after the generator has already spent time and changed files. Add a READ-DO pause point before generation when the domain has a canonical registry, schema, taxonomy, or artifact index.

1. Find the authority. Read the canonical registry or schema before naming the output.
2. Classify the artifact. Decide whether the job is an existing canonical instance, a local specialization, or a researched proposal for a new standard.
3. Record the contract. Capture the canonical name, purpose, route, local job, and evidence needed to proceed.
4. Validate before build. Run a deterministic preflight that fails on missing or redefined canon.
5. Generate and recheck. Only then create the output and run the normal formatting, type, link, test, and commit gates.

The preflight should reject three states: a local artifact presented as canon, a canonical name with a different purpose, and a new standard proposed without evidence or approval. If the same classification still needs repeated human interpretation, promote the contract to typed data or a guard rather than adding more reminder prose.

Failure Modes

• The agent reads the authority only after creating files.
• A mutable scorecard replaces or deletes the durable checklist that defined its setpoint.
• A checklist asks for judgment that a deterministic guard could enforce.
• A local specialization silently changes a canonical name or purpose.
• The output passes runtime checks but skips commit or publication gates.
• A failed run adds reminder prose without changing the earliest pause point, template, or guard.

Proof Of Done

The process rejects a deliberately wrong contract before artifact creation, accepts the valid contract, and then passes the same formatting, type, link, test, and publication gates applied to human work. The run records which Pattern, Process, or Platform owner changed and the next-run assertion that will prove baseline lift.

Changes my mind: repeated runs show that output-only gates catch the same defects at lower total cost without wasted generation, cleanup, or false standard promotion.

Retrieval

Pull this method when an agent, generator, template, or repeated workflow creates artifacts inside a governed taxonomy, schema, registry, or publication system.

Version delta: adds domain-contract preflight so agentic work validates classification before it generates files, while retaining same-gate output verification.

Context

• Process Optimisation — The improvement loop these principles feed
• Standard Templates — Template readiness gates enforce these principles upstream
• Control System — PID mechanics: the gauge must work before the controller can correct
• Checklists — Design the pre-generation pause point as a short READ-DO control
• Tight Five Loops — Keep the QA method to five executable control moves

Questions

Next question: which domain contract can your agent validate before it creates the first file?

• What percentage of your automated output has never been tested against the same hooks that gate manual work?
• If the tool that logs defects is itself defective, how do you close the feedback loop?
• Which of Deming's 14 points is hardest to apply to AI-generated output — and what does that tell you about your process?
• When a generator passes runtime tests but fails at commit time, which earlier control was missing?