[ notes ]

Notes

Field notes from BARGO. Receipts from what we ship, the mechanisms we've deleted from it under evidence pressure, and the research that grounds both.

Field notes, ordered with anchor pieces first and the rest by date descending.
date	receipt	read
2026-08-12	What we deleted while shipping Three receipts do more explanatory work than the seven primitives combined.	6 min
2026-08-13	What we scored and chose not to build Four candidate shapes scored across six constraint axes. Every one failed.	3 min
2026-08-08	The active-inference fetcher our own benchmark refused 91-SKU A/B refused it: −76% cost, −40% grounded attributes. Published with the fixture that refused it.	4 min
2026-08-01	Per-field fusion with learned source reputation 756 lines across six files. TruthFinder / Accu-style per-field resolution weighted by Beta-Bernoulli per-host trust, clamped so no source is fully believed or fully written off.	3 min
2026-07-25	Calibration monitor with an ECE-drift alarm Fitted confidence with an ECE drift alarm. 94-SKU non-circular benchmark: ECE 0.018 after the fit.	3 min
2026-07-18	Self-verify — does the cited value actually appear at the cited URL? About one in six shipped citations was cited-but-wrong. This is what closed that class.	3 min
2026-07-11	Determinism alarm — regression + silent-seam liveness 162 lines. Built after a fleet-averaged dashboard hid a 0.88 → 0.69 → 0.62 per-organisation slide across three runs.	3 min
2026-07-04	A deterministic prompt-injection envelope 69 lines, deterministic, no LLM classifier. A fenced envelope for crawled content so untrusted page text cannot rewrite the instruction it was meant to inform.	3 min
2026-06-27	ETIM 10.0 — the classification model 5,640 classes, 76,625 feature mappings. Licensed. Modelled internally for enrichment.	2 min

← bargo.lv