Systems that tell the truth about themselves.

The most common defect I keep finding isn't a missing check. It's a check whose name promises more than its query delivers.

Most AI evaluation suites drift toward checking whether the system behaved correctly instead of whether the work it produced was any good — build the coverage matrix before you trust a single metric in one.

TypeSafe's Jev returns a typed judgment and a calibrated number instead of a paragraph you then have to parse — worth testing against the heuristic it would replace, not worth trusting on the vendor's word.

The hardest part of document accessibility isn't fixing the tags — it's diagnosing whether a file can be fixed at all, and an honest diagnosis sometimes costs you the deal.

A model that reasons in tokens will hallucinate on physical law — the honest answer for deep-tech R&D is surrogate models grounded in physics plus a knowledge graph of prior art, and knowing where not to reach for an LLM matters as much as knowing where to.

Closure now requires a check that does not route back through the person whose work is being checked. A suspiciously clean report is a reason to look, not a reason to relax.

I shed the build and kept the seam: measurement, costing, and the connection between a technical system and a business outcome.

I reprioritise by irreversibility, not by who is asking loudest — the risks without a do-over come first. And 'nobody is forcing this' no longer means 'this can wait'.

Every comparison I present now states which fields the tooling could not read. An unlisted field is a stated gap, not a silent assumption.