CAIN-42 Public Evidence Index

Source commit 4b16f6b0cd73000afb7088ac0b7d17436cffa1c3 · release 4.0.0 · published 2026-10-09T16:30:00Z · publisher: CAIN-42 operator (FIRME AI; pre-incorporation).

What this is. A read-only index of what CAIN-42 actually implements, what is currently deployed, which claims have reproducible evidence, and which do not. It is designed so an outside reviewer can reach an independent verdict of PASS, FAIL, or NOT_ESTABLISHED without trusting CAIN's marketing. No login, cookie, JavaScript or API key is required.

Independence, stated plainly. Every artefact here is operator-generated. Independent reproduction: 0. Independent audit: 0. A self-published key proves the holder of the key signed the bytes; it does not establish who the holder is.

Headline claims

Claim	Status	Key evidence
Every hosted decision is ordered by a live 4-replica PBFT quorum.	OPERATOR_MEASURED	live status, latest hourly proof. At 2026-10-09T16:10Z: 4/4 reachable but agreement false, node-2 quarantined, verdict NOT_OPERATIONAL.
3-of-4 Ed25519 quorum certificates.	REPRODUCIBLE	sample bundle + offline verifier: 29/29 checks, 160 certificates, 7 negative controls.
Approximately 7.3 million model-checked states.	REPRODUCIBLE	models & commands, PBFT TLC, MCPGate TLC. 7,305,092 distinct states summed across 5 bounded runs. Not a proof of the code.
1,963 passing tests out of 1,978.	OPERATOR_MEASURED	test manifest, JUnit XML. 1,963 + 7 skipped + 8 xfail = 1,978. Re-running needs source (not public).

Statuses: DOCUMENTED · OPERATOR_MEASURED · REPRODUCIBLE · INDEPENDENTLY_REPRODUCED · INDEPENDENTLY_AUDITED · NOT_ESTABLISHED · FAILED · SUPERSEDED. Full registry: claims.json.

External Review Core Bundle (v2)

One complete, small bundle: README · manifest · publisher key · publisher attestation (_publisher/review-core-2026-10-09.json) · quorum certificate · certificate verifier · membership · replica keys · canonicalization · schema · negative fixtures · test reconciliation · formal summary · enforcement boundary · soak verdict · E42 verifier v2 · E42 verifier-assurance results · separated category totals · known limitations.

Documents

claims.json — full claims registry with exact text, status, evidence, falsifiers.
deployment.json — clusters, replicas, regions, failure domains, protocol assumptions.
status.json — current status (dynamic; server-generated UTC timestamp).
manifest.json — curated artefacts with bytes, SHA-256, content type, URL.
inventory.json — complete hashed inventory of every evidence artefact (7,400+ files, ~249 MB), discoverable here rather than by guessing paths.
retrieval-log.json — anonymous public-access audit: HTTP status and served-byte SHA-256 for every inventoried URL.
changelog.json — material claim changes, each linked to evidence.
known-limitations.md — what this package does not establish.
HOW_TO_REPRODUCE.md — step-by-step checks, labelled offline / online / source-required.
verification-pack/ — runnable offline pack with pinned dependencies.

Reproduce it yourself

curl -sO https://cainstudio.online/evidence/verification-pack/verify_all.py
curl -sO https://cainstudio.online/evidence/verification-pack/requirements.txt
python3 -m pip install -r requirements.txt
python3 verify_all.py            # offline: quorum certificates, formal models, chains
python3 verify_all.py --online   # adds read-only public endpoint checks

IP and disclosure policy

CAIN-42 publishes verification artefacts — evidence bundles, hashes, public verification keys, zero-CAIN-import standalone verifiers, abstract TLA+ models and raw model-checker output. It withholds the proprietary implementation source (platform-gateway/, cain/, the test suite, private keys, host IPs, overlay addresses and secrets) as a trade secret; there is no public git remote. Consequence, stated honestly: you can reproduce the verification of CAIN's evidence without CAIN's code, but you cannot rebuild CAIN from this package. That withholding is recorded as a reproducibility limit.

Known limitations (summary)

All evidence is operator-generated; nothing is independently reproduced or audited.
cain-mr-01 is 4 replicas on 3 hosts / 3 regions (two replicas share one host); both clusters use one provider (Vultr), one operator, one control plane.
At the last check cain-mr-01 was not in agreement and the hourly proof verdict was NOT_OPERATIONAL.
The 72-hour soak has not passed (RUNNING, 34.1h of 72h).
The 7.3M figure is a sum across 5 bounded models; no proved model-to-code refinement.
The 1,978-test run is not publicly reproducible (source withheld).
Bypass-resistance demonstrations are NOT_ESTABLISHED; they require operator approval and an isolated environment.
Performance figures use different measurement boundaries and lack public benchmark code.
No verified customers or revenue; no patent filed; no certification; no incorporation.

Full list: known-limitations.md.

This page is server-rendered HTML with no client-side JavaScript. The same content is available as JSON at manifest.json.

Verification package (offline QC verifier, negative fixtures, formal reconciliation, IP & licensing): /evidence/review/