Skip to content

Code Quality

A handful of things here get confused for each other constantly. Syntaxis keeps them apart on purpose — each is answering a genuinely different question, and none of them can stand in for another.

Proof The question it actually answers
Verify Do the project’s own checks pass?
Check Does this work satisfy its spec, in the app as it’s actually running?
Audit Is the code sound, secure, and consistent with our standards?
Independent audit Does someone with zero history on this build agree?
Check guide Can a human follow the steps and judge it themselves?
Test run Do the designed cases still hold across areas and platforms?

None of these substitute for the others. A green pipeline says nothing about behavior. A clean review doesn’t mean the app runs. A passing suite is silent about everything it never touched.

The project’s own repeatable checks — typecheck, tests, build — defined once through /ci and written down in AGENTS.md. Local work and GitHub both run that exact same recipe.

A review across quality, security, performance, and test coverage, either scoped to a branch or run against the whole project, narrowable to a specific lens when that’s all you need.

The rule that matters more than any other here: the ledger never decides what gets reviewed. The code gets looked at fresh every time, and only afterward do findings get recorded. Treating open findings as a checklist — verifying only those — is exactly the failure this exists to prevent, because a repair can introduce a brand-new defect that no existing entry was ever pointing at.

Either a genuinely fresh reviewer session, or a configured isolated reviewer child, examines one exact approved checkpoint and leaves behind a receipt: target and base SHAs, the spec’s hash, who was asked to review it versus who actually did, the check result, commands run, evidence, findings, and whatever risk remains.

A later change that actually matters makes that receipt go stale, and /complete won’t proceed on a stale, pending, or changes-requested review.

The declared adapter, model, and fresh-context identity are recorded metadata — not cryptographic proof of anything. Syntaxis writes down what ran; it doesn’t claim more certainty than that.

Reviews are held to the same bar builds are:

  • New abstractions, dependencies, services, configuration surfaces, or security machinery all need a current requirement behind them.
  • When scale is unknown, the smaller and more reversible design wins by default.
  • Simplicity is never an excuse to skip real trust-boundary validation, data-loss prevention, accessibility, an explicit security requirement, a configured test, or a project rule.

A pure proportionality complaint sits at P2 or P3. It only climbs to P0 or P1 once an actual code path, a broken contract, or a real security boundary is demonstrated — not just suspected.

P0 and P1 block completion, so they’re reserved for failures that have actually been shown to be real. A hunch gets recorded as unverified — tracked as a lead, without holding up a merge.

See The Findings Ledger for the full mechanics.