Ideas · Builder Notes

Kognita 0.4 in a day: a use-case register, policy packs and 0.4.1

Everything that went into Kognita alpha 0.4 on 18 October: the use-case register, a policy language, ingestion, adapters, origination and RM review, a flight recorder, circuit breakers, packs, a cross-review and two releases.

Context

0.4 was the release that had to answer the MAS mapping: a named owner for every use of AI, measured outcomes, and a full answer to all ten reconstruction questions. I built it between 9 am and 6 pm and released it twice.

Accountability

  • A versioned, effective-dated use-case register replaced the free-text purpose list. A decision must name a registered use case, and an empty owner fails closed.
  • Origination records why a client or product was chosen as cited facts, and refuses a bare score. RM review stores a hash of what the RM saw, a diff of what they changed and a provenance label.
  • A review faster than the minimum time, or an approver at a 99.9 per cent approval rate, is flagged as a rubber stamp. The thresholds are a starting point and need judgement.
  • A client send is its own governed event. With that, all ten reconstruction questions could be answered from evidence for the first time.

Authoring and integration

  • A YAML policy language compiles to the existing policy rows, with scenario tests next to the rules. An uncited rule won't load.
  • Docling ingestion splits documents by section and classifies each one, so one confidential section doesn't decide the whole file.
  • Adapters for OpenAI Agents, LangGraph, Pydantic AI and Hermes, each under 200 lines so policy logic stays in the core.
  • Five starter packs and a private-banking pack, with local overlays that updates never overwrite.

Operations

A flight recorder reads runs and alerts straight from the evidence chain and shows a missing measurement as unmeasured. Circuit breakers trip on registered thresholds, revocation works by agent, model, provider, tool, client or use case, and kill-switch drills are recorded; a high-materiality use case with no drill denies. Outcome metrics and near misses come from the interaction record, and a confirmed problem adds a review flag instead of a prohibiting policy. Gateway reads were bounded to rows in force after I measured overhead growing with closed history.

Review and release

Before tagging I asked for an adversarial cross-review. It found real gaps: the principal could come from somewhere other than the bound client configuration, evidence rows could be updated, tail truncation went unnoticed, an empty register allowed, and held calls missed policies that came into force during the hold. Each got a fix and a named test. 0.4.0 shipped mid-afternoon, and 0.4.1 followed at 6 pm after a trigger failed on SQLite older than 3.47.

Open

The append-only triggers are SQLite-specific, so other databases rely on verification failing closed. Some adapter extras pin conflicting versions and cannot be installed together.

Test the artefact you ship. The packaging test caught a broken wheel before any user did.

All ideasEmail Ahmed