Four tiers of evidence

20 July 2026

Most engineering claims arrive dressed as facts. The build passes. The scanner is clean. The feature works on Android. Ask what actually stands behind each one and the answers turn out to belong to different species, so we made the species explicit. Every claim in our engineering process carries one of four grades.

The last tier is the one that earns its keep. Almost every assurance failure we have studied, in our own work and in postmortems elsewhere, involves an unassessed claim promoted to clean by wishful accounting: the scanner that failed open, the platform nobody re-tested after a change, the reviewer who assumed the other reviewer looked. Naming the tier makes that promotion visible, and therefore refusable.

The grades apply to people and machines alike, and most usefully to automation that sounds confident. When an automated review asserts that something is absent, we require the search that would have found it: which pattern, which files. An unfalsifiable negative gets rejected regardless of how fluent it sounds, which turns out to be a good rule for colleagues of every kind.

None of this is novel; it is what good auditors have always done. Writing it down just means the discipline survives busy weeks, new tools, and the strong human preference for hearing that everything is fine.