· 3 min
how i think
these are not values. they are four mistakes i made expensively enough to remember.
every one of these has a specific system attached to it. i have tried to write the example rather than the slogan, because the slogan is the part that is easy to agree with and useless.
1 · put the invariant where it cannot be forgotten
a rule enforced by convention holds exactly as long as nobody writes the one line that forgets it.
- the usual multi-tenant answer is a tenant filter applied in application code on every read and write. it works until one query is written without it, and that query looks completely normal in review.
- so the boundary went into the database instead, forced, and the services check it against the live catalogue at boot and refuse to start if a table is unprotected. the invariant became something a machine asserts rather than something a reviewer remembers.
if a rule matters, make it impossible to violate rather than embarrassing to violate.
2 · disagreement is a signal, not noise
a system that produces one confident answer has thrown away the information you most needed.
- clinical pdfs defeat a single extractor in ways that fail silently — downstream, nothing can tell a sentence that was missed from one that was never written.
- running six engines and reconciling them by vote costs six times the compute per page. what it buys is that where they diverge, the pipeline knows the page is hard instead of quietly emitting one engine's guess.
the same shape shows up in agents arguing before they answer. cheap certainty is the expensive kind.
3 · a number without a source is decoration
if you cannot say how a figure was measured, it is not evidence, however true it happens to be.
- every metric on this site carries a source line and an as-of date, and the ones that come from my own role descriptions say so in those words rather than dressing themselves up as a production measurement.
- it is the same rule the products enforce: an answer that cannot name the clause it came from is an answer a lawyer cannot use, and one that cannot name its source is one nobody should act on.
holding myself to it is uncomfortable in exactly the way it should be.
4 · say what is still wrong
a write-up that claims nothing is broken is read as marketing, and it deserves to be.
- every case study here ends with what is unresolved. permission-aware retrieval is only ever as correct as the access state it last read. fallback parsing reports success for a run that was really rescued. hiring accuracy is not knowable inside a three-month engagement.
- none of that makes the work weaker. it is the part that makes the rest of it believable, and it is the first thing i look for when i read somebody else's.
the limitation you name yourself is the one nobody gets to discover for you.
show your work — cite specific text, admit uncertainty.
— evidane's house style, written into the product
where this shows up in practice is building; where it gets argued in public is writing.