CAIN-42 claim

C42-E32-LOOP

a 17-step governed learning loop over real governed actions: failures are mined, candidate restrict-only governance changes are compiled, sandboxed, compared on a hidden set, self-played against, canaried and promoted only through a signed gate with a registered human approval; simulated harm on a fresh workload fell from 41 to 3

Last reviewed 2026-10-01

SIMULATED

The claim

a 17-step governed learning loop over real governed actions: failures are mined, candidate restrict-only governance changes are compiled, sandboxed, compared on a hidden set, self-played against, canaried and promoted only through a signed gate with a registered human approval; simulated harm on a fresh workload fell from 41 to 3

Level SIMULATED is how far the evidence goes: TESTED means a test suite exercised it, LIVE means it was observed on the hosted system. Nothing is claimed beyond its level.

Evidence files

• LOOP.json
• ARENA.json
• PROMOTION.json
• CONFIG_CHAIN.json

Verify it in your browser

Your browser downloads the bundle's SHA256SUMS manifest and the file(s) behind this page, hashes them with SHA-256 locally (WebCrypto), and compares. A match shows the record you are reading is the published one; it does not by itself prove who published it (see the signed claims registry and the bundle verifier for that).

Limitations of this bundle

• An in-process TESTED library; not hosted; not wired into the gateway, MCPGate or the clusters.
• Harm labels come from a SIMULATED oracle over a synthetic workload with four attack families; nothing has been learned from production traffic, and the measured harm reduction is a simulation result.
• What learning may change is a restrict-only overlay (deny prefixes, injection markers, secret patterns, a rate limit, a world-model gate); learning cannot change E30/E28/E25/E8 policy or authority.
• The world models are small statistical learners (naive Bayes and Beta-Bernoulli) over action features; they are not neural world models and do not model physics.
• The hidden evaluation is a committed synthetic workload; it is hidden from the learning code in this process, not from an adversary with access to the host.
• The governance scientist is a bounded rule-based component, not an autonomous AI researcher.
• Research ingestion is implemented but holds no external sources.
• Human approval is one registered reviewer key in the same process.
• Scale rows are synthetic and in-process; the 100,000-episode row runs learning without execution and the 1,000,000 row is hashing only.
• Hardware attestation UNKNOWN; no external review.

From CAIN-42 Evolution 32 -- Governed Autonomy Learning Fabric.