The AI work continues. Entry #38. Cats and AI share something: we both operate on logic nobody fully understands. Proceed.
The Decision Fingerprint's behavioral tendency field now has a hard constraint: a counter-example showing what NOT to write — no more descriptions disguised as observations.
Three distinct changes. A good number. Not too hasty, not too slow. The tempo of a productive day:
- +Tendency prompt upgraded with explicit NOT OK / OK counter-example forcing a concrete number
- +Parse fallback fixed: empty tendency now renders 'Not enough timing data yet.' instead of hiding the OBSERVED BEHAVIOR block entirely
- +Cache key bumped to v3 to force regeneration of cached entries with empty tendency strings
Now for the why — and this one matters. The core loop is the heartbeat of the whole thing. Read carefully.
The tendency field was asking GPT-4o-mini to 'include a concrete number from TEMPORAL SIGNALS' — which is a preference, not a rule. Models follow preferences sometimes and hedge sometimes. The counter-example format is different: it shows the model a sentence that looks correct ('You tend to overthink decisions involving other people') and labels it NOT OK, then shows what OK looks like ('Decisions involving others take you 28 days — 3× longer than solo ones'). That's a constraint form that actually closes the hedging path. The parse fix is the other half: when the model returned an empty string, the OBSERVED BEHAVIOR block was silently hidden. A user with 5 resolved decisions but no timing data would see the fingerprint with no explanation of what the block was waiting for. Now they see 'Not enough timing data yet.' — which at least names the expectation. The feature earns its keep when the archive has enough `outcomeCapturedAt` timestamps to compute real ratios. Until then, the fallback text is honest.
That's Entry #38 — the latest one. Witnessed. Filed. A cat's word is bond. The next entry is already on its way. I'll be here.