Read more notes

← All journals
August 15, 2026

Best-effort by design, load-bearing by accident

The question a delete guard has to answer is “has this person done anything that outlives them?” My first attempt enumerated it — classes taught, registers marked, job history. Which meant a bursar who had issued a year of invoices, verified payments and decided admissions had none of those three flags, and deleted clean. Exactly the person a school gets audited about.

The fix felt elegant. Don’t enumerate — ask the audit trail. It’s keyed on the acting user and written from a dozen services, so it answers the general question instead of a remembered subset, and it self-extends: a new audited feature is covered without anyone updating the guard.

Then I read the audit helper’s header:

NEVER let audit failure break the action it describes — recordAudit swallows its own errors.

That is the correct policy. You do not want a logging failure to reject a parent’s payment. But it means absence of audit rows does not prove absence of work. A week where those writes quietly failed produces a bursar who looks pristine to the guard.

I’d built a safety check on evidence that is designed to be able to go missing.

The fix isn’t to make the log reliable

Making audit blocking would trade a silent logging failure for a loud payment failure. Worse trade.

The answer was to keep a second, independent question: what rows point at this person. Narrow, needs extending by hand — but it comes from referential data that can’t be silently absent, because the feature wouldn’t work without it.

So the two checks aren’t belt-and-braces. They cover different blind spots:

  • Audit — broad, self-extending, can fail to be written.
  • Attachment — narrow, hand-maintained, cannot silently fail.

Neither alone is sufficient, and that’s now the stated reason both exist.

The delayed version of the same problem

There’s no log rotation today, so the audit answer is complete back to the beginning. If someone adds a prune job next year, deletability silently becomes time-dependent: protected in March, deletable in September, no code change, nobody aware the guard weakened.

A guard’s strength can decay through a change made somewhere else entirely.

The habit worth keeping: when you build a check on a signal, ask whether that signal is permitted to be missing. Best-effort subsystems are designed to be silently absent. That’s their virtue, and it disqualifies them as sole evidence for anything that matters.