How we work / Engineering standard

285 rules.
168 refuse the commit.

Most « engineering standards » are a document nobody reopens. Ours is executable : on every commit a mechanical pass evaluates the code against a registry of rules and refuses what does not hold. Not a reminder. A refusal.

285

rules in the registry

registry.json

168

blocking — they refuse, they do not warn

blocking = true

132

backed by an executable script

96 of them blocking

148

evaluated before the commit even exists

trigger = pre-commit

01 · The defect this fixes

A standard you can ignore is not a standard.

Three things break a written standard. We met all three, so we stopped writing it and started executing it.

The document nobody reopens

A standard in a PDF starts ageing the day it is signed. A registry runs : a rule that no longer matches the code fails, and gets fixed or retired.

The review that depends on who reads it

Two reviewers, two verdicts. A mechanical rule returns the same verdict at 3 a.m. as it does in a meeting — the only way a standard survives fatigue.

The « just this once » exception

It becomes the rule within three weeks. Here the exception exists, but it is named, written into the file, and visible in review — not muttered.

02 · When it bites

Five moments, earliest to latest.

The earlier a rule bites, the less it costs. Half the registry runs before the commit exists ; what needs the whole repository waits for the night.

pre-commitBefore the commit exists — the fault never enters the history148
nightlyWhat needs the whole repository — drift, coverage, debt46
pr-gateAt review — what can only be judged on a complete slice38
pre-pushBefore it leaves the machine — types, build, coherence24
release-tagAt ship time — provenance, signature, inventory12
03 · The last verdict, as it came

A standard that always says yes proves nothing.

This is the real output of the last pass on this very site. We publish it whole — including the column saying a third of the rules did not apply.

Evaluated

147

The rules whose trigger matched what was being staged.

Passed

106

72%. The rest were corrected before the commit existed — which is the point.

Not applicable

41

They did not bite on this slice. Counting a « not applicable » as a pass would be theatre.

What that number means, and what it does not.

106 of 147 is not a quality score : it is a count of applicable rules satisfied at commit time. A rule that does not apply says nothing good or bad. And a pass that never refuses anything is a signal to go read the registry, not to celebrate.

04 · The loop

A new rule warns before it refuses.

Making a rule blocking overnight breaks a team and gets routed around. Every rule goes through the same loop — that is what makes it survivable.

A defect repeats

The same incident lands twice. We write the rule that would have caught it, in warning mode, wired to the right trigger.

It warns, and we count

Calibration window : the rule runs without blocking and we measure its false positives. A rule that cries wolf is fixed or retired — it does not graduate.

It refuses, and we prove it

Once blocking, it is mutation-tested : we reintroduce the defect on purpose and require the rule to fail. A guard that never fails proves nothing.

05 · What it changes for you

You are not buying our discipline. You are buying its trace.

The deliverable carries its proof

Every delivery arrives with the verdict of the pass that let it out — evaluated, passed, not applicable. Not an assertion of quality : a reading.

Quality does not depend on who was there

The same rule applies to a newcomer's first day and a senior's tenth commit. That is what lets a team change without the bar moving.

What we did not verify is written down

Coverage gaps are declared, not hidden. A report listing no limits is a report worth re-reading.

Ask for the verdict before you sign.

We will show you the pass on a real repository — rules evaluated, refused, corrected. Not a demo : an output.

Where every figure on this page comes from

285 / 168 / 101 / 132 / 96 and the trigger split are counted in the engineering registry. The 147 / 106 / 41 verdict is the output of the pre-commit pass run on this repository on 2026-07-28. No figure is rounded or arranged.

What this page does not claim : no certification, no third-party audit, no external norm. It is an internal standard, mechanically applied. That is the entire claim.