A 200 carrying ok:false
Before rewriting our research, we probed every read tool by content, not by HTTP status. Thirty-odd tools worked. Four defects surfaced, one of which was not a defect: we had reported a number without reading the design behind it.
1 Nov 2026 · 2 min readOur earlier research write-ups were drafted while half the tools were called from memory rather than from their contract. Before rewriting them, every read tool was probed by content: a 200 OK carrying {"ok": false} counts as a failure.
Infrastructure: clean
hive check 6 of 6
vector collections 17 of 17 green
knowledge graph 1,046 entities · 10,060 relationships
research corpus 179 files · 5,370 chunks, with provenance tiers
rules 41 active · 142 triggers
About thirty read tools worked with correct arguments: memory, code, docs, sources, graph, rules, plans, files, git. One tool refused without a token, which is the guard working.
Three real defects, fixed the same day
- One search tool was missing from the tool contract. Its result shape could only be learned by calling it, and it was the one tool whose shape we got wrong.
- Duplicate points in the knowledge index. One query returned the same record twice at an identical score. Retrieval quality was being measured on a corpus that double-counts.
- A graph-neighbour tool required a numeric id, with no tool to look an id up from a name. The neighbours of a concept could not be reached in one call.
The fourth was ours
We reported “995 of 1,025 entities in cluster 0” as a defect. It was the documented behaviour: the clusters were connected components, not themes, and the code’s own header said so. A community algorithm would subdivide it but is unstable across runs, so it was deliberately not used. Reading the code before reporting would have caught it.
The real defect underneath: the graph metrics were never recomputed. Code ingest refreshed its metrics 30 seconds after changes; the knowledge graph had no equivalent, so tiers and clusters only refreshed when someone remembered a flag. Michal named the rule that came out of it: live database, never static.
The method
- Probe by content, not status. A tool that answers politely with nothing is the most dangerous kind.
- Read the design before reporting a number as a defect. Otherwise the audit adds noise to the system it audits.
- Audit the tools before writing about results. Anything built on an unaudited tool inherits its blind spots.
The measurements in this article come from these records in the running system.
~/memory/MEASURED-toolchain-audit-2026-09-17.md— toolchain audit, step 1 of the research rewrite
For three and a half hours our agent layer recorded none of five local-model runs. A network sensor that knows nothing about agents recorded all five, to the minute. Self-report was not wrong. It was absent, and absence reads as 'nothing happened'.
Every figure on this site comes from a running system. The standard, the time it failed us by a factor of eighteen, and where the numbers come from.
Una gave the correct physics and cited the encyclopedia instead of the primary paper sitting next to it. Four hypotheses, each tested by changing one variable. She does not pick the most authoritative source. She cites the one whose wording she reused.