CAIN-42 evidence library

AGENTIC 75 verification run: 2026-11-13

Evidence bundle: 0 claims, 0 of 0 invariants held.

Last reviewed 2026-11-13

Browse the raw bundle

The bundle's own README

AGENTIC 75 verification run: 2026-11-13#

One run on one host. Observed numbers, including the negative ones.

This bundle records what a verification pass on this host actually produced. It does not repeat the product bundle's claims; it tests them. Where a requested claim is not supported by the run, it is listed under Not established rather than softened.

What was verified#

1. Selftest. Every one of the 75 DELIVERABLE SKUs ran on its fixed sample and its SHA-256 manifest re-verified: 75/75. Selftest output line: 75/75 DELIVERABLE SKUs produced a verified manifest. 2. Tests. 2687 passed in 216.85s (0:03:36). The selection is the 75 tests/test_a75_*.py files plus the catalog, CLI and live-try-route tests. 3. Signed samples verify offline. 75/75 signed samples verify with their pinned public key and a verifier that imports no CAIN code. Outer verdict strings: VERDICT: VERIFIED x 74, VERIFIED_RUN x 1. Inner per-engine verdicts: SOUND x 53, SUPPORTED x 4, COMPLIANT x 4, CONTAINED x 2, ACTIVE x 2, ALLOW x 1, VERIFIED x 1, CONSISTENT x 1, SAFE x 1, PASS x 1, ACCEPT x 1, CLOSED x 1, CERTIFIED x 1, PRESERVED x 1, CONFORMANT x 1, ALLOWED x 1. Per-SKU rows, payload name and verdict are in RUN.json under offline.results. 4. FALSEWORK per engine. 75 packages graded separately (the whole directory exceeds the detector's 512,000-byte scan cap, so a single call returns UNKNOWN with nothing analysed). Grades: {"PARTIAL": 9, "REAL": 66}. Findings by severity: {"HIGH": 7, "LOW": 184, "MEDIUM": 11}. 5. Quarantine latency, measured not quoted. 300 iterations, quarantine_agent p50 1.972 ms / p95 2.19 ms / p99 2.557 ms / max 3.19 ms; is_quarantined p50 0.334 ms / p99 0.525 ms. Scope: single process, local SQLite WAL file, in-process cache; NOT cluster-wide, NOT measured across the 75 engines.

Not established by this run#

These are the claims that were requested for this run and that the evidence does not support. They are recorded here so the bundle cannot be read as endorsing them.

1. "A cohesive super-agent control fabric." The 75 packages are independent engines. 73 of 75 import only the shared signing kit _kit.py; 0 import another engine package; no package composes them into one runtime. What is shared is a record format and a per-engine signing key, not authority, state or a control loop. The two engines that do not use _kit (CAIN-A75-001 longrun, CAIN-A75-002 labpreflight) have their own signers and verifiers. 2. "Fully linked via hash-chained SQLite WAL journals." Engines that use SQLite: none. Records are hash-chained (prev_record_hash over canonical JSON) and Ed25519-signed, but the chain is per record sequence, not a SQLite WAL journal shared across the suite. SQLite-WAL durability exists elsewhere in the repository (cain/quarantine.py), not in cain_a75_engines/. 3. "Ed25519 capability token leases" between the products. Engines mentioning a capability token: ['realtimegate'] (as a field of supplied sample data, not a lease the engine issues or enforces). Ed25519 is used to sign each engine's record. ActionCapabilityToken lives in the CAIN kernel, and this run found no wiring from it into the 75 engines. 4. "Zero un-fail logic or hard-coded stubs detected by FALSEWORK." FALSEWORK reports 202 findings, including 7 HIGH. The HIGH findings are all in samples.py fixture data (typewritten success counts such as passed: 412), which the detector is right to flag: they are simulated inputs, not computed results. STUB count is 0; PARTIAL is 9. 5. "Sub-50 ms fail-closed revocation benchmark." No such benchmark existed in the repository for the 75 products. The number in this bundle is a new single-process measurement of cain/quarantine.py (see above), not a suite-level property and not a cluster-wide result. It is one host, one process, one local SQLite file. 6. "Published via a public ledger ingest API, with immutable public ledger indexing." /evidence/ingest is operator-only (X-Gateway-Admin-Key) and writes to the gateway's local evidence database; it returns {"status": "success"} and appends to the transparency log best-effort. It is not a public endpoint and it does not return a ledger index or receipt. This bundle is published as static files on clawx.click, which is a file publish, not a ledger registration. 7. "Establishes the 2027 paradigm." A paradigm claim is not a testable proposition. Nothing here can establish one, and this bundle does not assert it.

Limits#

• Samples are fixed inputs shipped with each engine, not customer systems.
• deliver writes findings.json, report.md and an unsigned manifest. The signed

samples in this repository come from python3 -m cain_a75_engines.<module>.cli sample, not from deliver.

• FALSEWORK is heuristic. REAL means the detector found no fabrication pattern, not that

the engine is correct.

• No customer, revenue, partnership, certification, third-party audit or patent claim.
• The team is a three-person team.

Generated 2026-11-13T05:42:49Z on vultr (3.14.4).