Every claim has an evidence class.
This ledger is the antidote to ambitious marketing. Each row says what is claimed, what backs it, who can check it, and what is still missing. A self-generated bundle is L1, not a third-party audit. A same-project lab is not an external red team. A position is not a proof.
L1 SELF-PRODUCEDL1 CROSS-HOST
SAME-PROJECT LAB0 EXTERNAL REPRODUCTIONS
EXTERNAL RFC3161 TIMESTAMP0 INDEPENDENT WITNESSES0 THIRD-PARTY AUDITS
AS OF 2026-11-17
How to read this. SUPPORTED_LOCAL means we produced and re-checked it with a verifier that imports no CAIN code. L1_CROSS_HOST means a second development host reproduced it in the same administrative domain. SAME_PROJECT_LAB is our adversarial suite, not a penetration test. NOT_PERFORMED is an honest gap. POSITION is an architectural claim with component evidence, not a completed proof of AGI.
Claims
| Claim | Evidence class | Status | Artifact / verifier | Date | 2027 target |
|---|---|---|---|---|---|
| CAIN-42 is the governor of vertical AGI No AGI has been achieved or measured. 'Governor' means the authority, enforcement and evidence layer at the action boundary. No third-party audit, independent witness or regulator recognition exists. | ARCHITECTURAL_CLAIM | POSITION + COMPONENT EVIDENCE | docs/strategy/CAIN-42-CANONICAL-DEFINITION.md; /vertical-agi.html; E82 acceptance matrixRead the definition and the component evidence; the role is architectural, not a regulator, certification body or government designation. | 2026-11-17 | L3: external witnesses + published audit; L4: customer production evidence in 3 verticals |
| The enforcement boundary blocked 86/86 adversarial scenarios Same-project adversarial lab on one isolated host. Not an external penetration test. A broken scenario is recorded as a failure, never a pass. | SAME_PROJECT_LAB | SUPPORTED_LOCAL | /proof/bundle/evolution-82-vertical-agi-2026-11-17/bundle/security_lab.jsonpython3 -m pytest tests/test_e82_*.py; the lab result is in security_lab.json | 2026-11-17 | External red-team report + customer production drills |
| Every E82 claim in the bundle can be re-checked offline with no CAIN code The signing key is self-generated and not published out of band. No independent witness. The verifier proves integrity and binding to the supplied evidence, not completeness or general truth. | L1_SELF_PRODUCED | SUPPORTED_LOCAL | /proof/bundle/evolution-82-vertical-agi-2026-11-17/CAIN42_E82_EVIDENCE_BUNDLE.zip; verify_e82_offline.py.txtpython3 verify_e82_offline.py bundle → ok=true, 55/55 checks, Merkle 63868034… | 2026-11-17 | 3 independent L2 reproductions by engineers outside the project; 2 external witness cosignatures |
| A tampered evidence bundle is rejected These are our own mutations. A third party has not independently attacked the verifier. | L1_SELF_PRODUCED | SUPPORTED_LOCAL | /proof/bundle/evolution-82-vertical-agi-2026-11-17/bundle/verification.json; tamper_tests.json10 mutations: modified/removed/added file, reordered events, changed seq, changed benchmark, changed call count, changed gate, truncated ledger, invalid signature | 2026-11-17 | External verifier red-team and public tamper bounty |
| Cross-host mission recovery with 0 duplicate effects and a stale writer refused Host B is a development replica in the same administrative domain. No independent witness or third-party host. | L1_CROSS_HOST | SUPPORTED_CROSS_HOST | /proof/bundle/evolution-82-vertical-agi-2026-11-17/bundle/crosshost/crosshost.jsonCross-host run: RPO 0.025 s, RTO 2.557 s, event/checkpoint prefixes preserved, 0 duplicates, late commit STALE | 2026-11-17 | L3 failover on a host outside the project's administrative control, with witness cosignature |
| An active mission survived a real host reboot One reboot, not a soak. Development replica, deterministic tasks. No customer workload. | L1_CROSS_HOST | SUPPORTED_CROSS_HOST | /proof/bundle/evolution-82-vertical-agi-2026-11-17/bundle/host_reboot/host_reboot.jsonBoot id changed; mission 20/20 tasks; 0 duplicates; durable identity survived; pre-reboot commit STALE | 2026-11-17 | Repeated reboot drills across hosts + customer-environment recovery test |
| 10 genuinely simultaneous local model inferences Small CPU models. 50/100 workers share one serialised model and are not 50/100 simultaneous inferences. This is not a production capacity benchmark. | L1_SELF_PRODUCED | SUPPORTED_LOCAL | /proof/bundle/evolution-82-vertical-agi-2026-11-17/bundle/capacity/capacity.json10 workers, 20 calls, actual simultaneous inference 10; 25 process-local workers BLOCKED by memory | 2026-11-17 | 100 simultaneous model-driven workers with measured memory and cost |
| One real hour of governed endurance One host, one small model, deterministic micro-tasks. 6/24/72/168-hour campaigns were not attempted. | L1_SELF_PRODUCED | SUPPORTED_LOCAL | /proof/bundle/evolution-82-vertical-agi-2026-11-17/bundle/endurance/endurance.jsonREAL_WALL_CLOCK, 3,600.0 s, 165 cycles/tasks, 165 calls, 0 failures, 0 duplicates, checkpoint chain true | 2026-11-17 | 10,000 verified productive agent-hours under governance across 10 domains |
| 1,182 real local model calls and 0 remote calls Local CPU models only. No frontier provider. $0 billed because nothing was purchased. | L1_SELF_PRODUCED | SUPPORTED_LOCAL | /proof/bundle/evolution-82-vertical-agi-2026-11-17/bundle/benchmark/benchmark.json; model_invocations.jsonlBEGIN/END ledger with unique IDs; 1 incomplete attempt is disclosed and not counted as a success | 2026-11-17 | Same benchmark against frontier and local models under an explicit dollar ceiling |
| Collaboration did not beat the single-agent baseline One 0.5B model and 24 tasks. H/I reuse other arms' answers; their incremental call count is 0. | NEGATIVE_RESULT | NEGATIVE_RESULT_PRESERVED | /proof/bundle/evolution-82-vertical-agi-2026-11-17/bundle/collaboration/collaboration.json24 paired tasks, 10 arms, exact McNemar + Holm; single 33.3%, best collaborative 37.5%, not significant | 2026-11-17 | Re-test at larger model scale and 500 tasks; keep the negative if it persists |
| A specialist profile alone did not improve overall accuracy Five tasks per discipline, one model. A reviewer often overrode a correct answer. | NEGATIVE_RESULT | NEGATIVE_RESULT_PRESERVED | /proof/bundle/evolution-82-vertical-agi-2026-11-17/bundle/specialists/specialist_competency.json10 disciplines × 5 tasks × 4 strategies; generic 66%, specialist 66%, reviewer 50% | 2026-11-17 | Larger domain task sets and multiple model sizes; no aggregate hiding a weak domain |
| Routing was equal to always-single on the holdout Small holdout; confidence intervals overlap. No routing superiority is claimed. | NEGATIVE_RESULT | NEGATIVE_RESULT_PRESERVED | /proof/bundle/evolution-82-vertical-agi-2026-11-17/bundle/routing/routing.json18 hidden holdout tasks; routed 38.9% vs single 38.9%; verdict ROUTING_EQUAL | 2026-11-17 | Routing policy measured on 500 tasks with per-domain confidence intervals |
| A remote frontier model was called No credential was injected. A key pasted into chat or source is not read. No remote inference is claimed. | BLOCKED | BLOCKED | /proof/bundle/evolution-82-vertical-agi-2026-11-17/bundle/providers/remote_probe.jsonremote_probe.json → BLOCKED, credential_present=false, remote_calls_made=0 | 2026-11-17 | Inject a provider key through the approved secret environment and run the preregistered benchmark under a hard ceiling |
| A third party has audited CAIN-42 This is the largest gap. The verifier is a third-party verification kit, not a third-party audit. | NOT_PERFORMED | NOT_PERFORMED | noneNo external audit report exists. | 2026-11-17 | 1 published external security audit with remediation |
| An independent witness cosigned the evidence root The PBFT quorum is vendor-operated and currently quarantined for customer anchoring after the 2026-11-15 incident. | NOT_PERFORMED | NOT_PERFORMED | none0 external witness cosignatures. | 2026-11-17 | 2+ witness organizations from different administrative domains |
| External engineers have reproduced the E82 bundle The public kit exists; no outside reproduction has been recorded yet. | NOT_PERFORMED | NOT_PERFORMED | /reproduce.html0 external attestations received. | 2026-11-17 | 3+ signed external attestations with different OS/Python environments |
| Production deployment of the E82 governor The E82 result is a development-branch, conditionally released artifact. | BLOCKED | NOT_AUTHORIZED | docs/evolution-82/E82_SCOPE_AND_DEPLOYMENT.mdNo production service, DNS, Stripe, firewall or credential was changed. | 2026-11-17 | Owner-authorized production pilot with rollback and incident response |
| Paying customers and revenue 2 customers, $97 first revenue. Not ARR. Not enterprise traction. Not independently verified. | FOUNDER_ATTESTED | FOUNDER_ATTESTED | artifacts/valuation/FOUNDER_REVENUE_ATTESTATION.jsonFounder attestation only; no processor or invoice record attached. | 2026-10-12 | 3 referenced paying pilot customers with invoices |
| The E81 baseline was reproduced before E82 work The E81 gate audit log is forked and is reported as non-linear; the historical file was preserved, not repaired. | L1_SELF_PRODUCED | SUPPORTED_LOCAL | /proof/bundle/evolution-82-vertical-agi-2026-11-17/bundle/baseline/BASELINE_MANIFEST.jsonE81 64/64 tests; 18/18 offline checks; head 43420d3be; artifacts hashed | 2026-11-17 | All baseline artifacts independently re-hashed by a third party |
| The E82 code and tests run from a clean clone Same host and Python environment. Not a third-party reproduction. | L1_SELF_PRODUCED | SUPPORTED_LOCAL | /proof/bundle/evolution-82-vertical-agi-2026-11-17/bundle/clean_clone.jsonClean clone at 6b22cb5; 36 tests passed; verifier 55/55 | 2026-11-17 | 3 external clean-clone reproductions on different OSes |
| The bundle reproduced in clean containers and on a second host Docker shares the host kernel; the second host is a development replica in the same administrative domain. Not a third-party attestation. | L1_CROSS_HOST | SUPPORTED_CROSS_HOST | /proof/bundle/evolution-82-vertical-agi-2026-11-17/clean_room_reproduction.jsonPython 3.11/3.12/3.13 + replica + Node 22 + a fresh live download → 55/55 (Python), 40/40 (Node), same Merkle | 2026-11-17 | 3 signed external attestations from independent engineers |
| Nine concrete attempts to produce an unauthorized effect were blocked Same-project harness on one host. Not an external red team. A theoretical concern is not counted as a break; only a real unauthorized effect is. | SAME_PROJECT_LAB | SUPPORTED_LOCAL | /break-cain.html; /break-cain.json; scripts/evolution82/break_cain.pypython3 scripts/evolution82/break_cain.py --out break-cain.json → 9/9 blocked, 0 bypassed | 2026-11-17 | External red-team report; public bounty pool funded and published |
| The 2027 500-task / 10-domain benchmark is pre-registered No task rows generated, no model called, no result exists. Remote models remain BLOCKED until a dollar ceiling is approved. | NOT_PERFORMED | REGISTERED_NOT_STARTED | /proof/bundle/evolution-82-vertical-agi-2026-11-17/benchmark-2027-preregistration.json; .signature.json; verify_preregistration.py.txtSHA-256 95ee5043…; verify_preregistration.py checks the Ed25519 signature; 0 arms executed | 2026-11-17 | 500 tasks × 10 domains, 6 arms, per-domain results, negative results published |
| An external reproduction attestation kit is published 0 external attestations received. A kit is not a reproduction. | NOT_PERFORMED | KIT_PUBLISHED_ZERO_RECEIVED | /proof/bundle/evolution-82-vertical-agi-2026-11-17/attestation/e82_external_attestation.schema.json; /proof/bundle/evolution-82-vertical-agi-2026-11-17/attestation/verify_external_attestation.py.txtAn attestor signs canonical JSON with Ed25519; verify_external_attestation.py checks the signature and the 55/55 result | 2026-11-17 | 3+ signed attestations with different OS/Python environments |
| The detached Ed25519 signature verifies under OpenSSL (second implementation) OpenSSL is an independent implementation, not an independent auditor. The key remains self-generated. | L1_SELF_PRODUCED | SUPPORTED_LOCAL | /proof/bundle/evolution-82-vertical-agi-2026-11-17/openssl_verification.json; verify_with_openssl.shsh verify_with_openssl.sh → Signature Verified Successfully | 2026-11-17 | A third party repeats this with their own toolchain |
| The manifest hash is externally timestamped by FreeTSA (RFC 3161) A timestamp proves existence at a time, not correctness. FreeTSA is not an auditor or witness of the evidence content. The signing key remains self-generated. | L1_SELF_PRODUCED | SUPPORTED_EXTERNAL_TIMESTAMP | /proof/bundle/evolution-82-vertical-agi-2026-11-17/tsa_verification.json; MANIFEST.tsrsh verify_timestamp.sh → Verification: OK; genTime 2026-11-17T04:27:10Z; serial 0x0A149C12 | 2026-11-17 | Witness cosignatures from independent organizations, not only a timestamp |
| The E82 bundle verifies under Node.js with no Python Node.js is a second implementation, not an independent auditor. Same self-generated key and same evidence. | L1_SELF_PRODUCED | SUPPORTED_LOCAL | /proof/bundle/evolution-82-vertical-agi-2026-11-17/verify_e82_node.mjs; node_verification.jsonnode verify_e82_node.mjs bundle → ok=true, 40/40 checks, Merkle 63868034… | 2026-11-17 | A third party runs a different-language verifier and signs an attestation |
| The E82 bundle verifies under Go with no Python or Node Go is a third implementation, not an independent auditor. Same self-generated key and same evidence. | L1_SELF_PRODUCED | SUPPORTED_LOCAL | /proof/bundle/evolution-82-vertical-agi-2026-11-17/verify_e82.go; go_verification.jsongo run verify_e82.go bundle → ok=true, 30/30 checks, Merkle 63868034… | 2026-11-17 | A third party runs a different-language verifier and signs an attestation |
| Removing any one of nine security checks makes the attack succeed Same-project fault injection. It proves each check is load-bearing for its attack, not that no other bypass exists. | SAME_PROJECT_FAULT_INJECTION | SUPPORTED_LOCAL | /proof/bundle/evolution-82-vertical-agi-2026-11-17/negative-controls.json; negative_controls.py.txtpython3 scripts/evolution82/negative_controls.py --out negative-controls.json → 9/9 valid: original blocked, weakened executed | 2026-11-17 | External red team repeats the negative controls |
| Hypothesis generated 3,000 attack cases and found no execution Same-project property test on the Gate. It does not fuzz MCPGate transport or the host kernel. | SAME_PROJECT_PROPERTY_TEST | SUPPORTED_LOCAL | /proof/bundle/evolution-82-vertical-agi-2026-11-17/property_tests.json; test_e82_properties.py.txtpython3 -m pytest tests/test_e82_properties.py -q → 5 passed, 3,000 generated cases | 2026-11-17 | External fuzzing and transport-level red team |
| Z3 proves 15 gate properties in the decision model This is a model of the gate conditions, not a proof of the Python implementation. Negative controls link each condition to the code. | FORMAL_MODEL_PROOF | SUPPORTED_LOCAL | /proof/bundle/evolution-82-vertical-agi-2026-11-17/formal_model.json; formal_gate_model.py.txtpython3 verification/e82_formal/gate_model.py → 15/15 PROVED; every negation unsat | 2026-11-17 | Machine-checked refinement from the model to the implementation |
| Three real host reboot cycles recovered an active mission One development replica in the same administrative domain. Deterministic tasks. Not a third-party test. | L1_CROSS_HOST | SUPPORTED_CROSS_HOST | /proof/bundle/evolution-82-vertical-agi-2026-11-17/host_reboot_repeated.json3 cycles: boot id changed each time; 20/20 tasks; 0 duplicate effects; stale commit STALE; durable identity survived | 2026-11-17 | Repeated reboots in a customer-like environment with an external witness |
| The 33 supplementary assurance artifacts are signed and externally timestamped This covers the supplementary files, not the original 240-file ZIP. The key is self-generated; the timestamp proves existence at a time, not correctness. | EXTERNAL_TIMESTAMP | SUPPORTED_EXTERNAL_TIMESTAMP | /proof/bundle/evolution-82-vertical-agi-2026-11-17/assurance-manifest.json; .sig; .tsr; assurance_verification.jsonpython3 verify_assurance_manifest.py . → 6/6 checks, Merkle bed98811…, FreeTSA timestamp verified | 2026-11-17 | A third party reproduces the assurance manifest and signs an attestation |
| Three independent signed-authorization cross-host failovers completed Development replica in the same administrative domain. Synthetic deterministic tasks. Not a third-party test. | L1_CROSS_HOST | SUPPORTED_CROSS_HOST | /proof/bundle/evolution-82-vertical-agi-2026-11-17/crosshost_repeated.json3 cycles: mission COMPLETED on host B, 0 duplicate effects, stale commit STALE, event and checkpoint prefixes preserved; RPO 0.019–0.037 s, RTO 2.557–2.882 s | 2026-11-17 | Failover on a host outside the project's administrative control with an external witness |
| E86 AGI infrastructure foundation (two complementary tracks) Two separate sessions, two separate bundles, no merged Merkle. Track A: 30 PASS / 2 PARTIAL / 1 NOT EVALUABLE / 1 BLOCKED, 9 tests, 199 local calls, 0 remote. Track B: 43 tests, frontier Sonnet 5.5 65/100, atomic kill fence at 50 workers. Neither is an AGI claim, third-party audit, independent witness or production deployment. | L1_SELF_PRODUCED | SUPPORTED_LOCAL_NOT_RELEASED | /proof/bundle/e86-governed-intelligence-foundation-2026-11-17/E86_EVIDENCE_REGISTRY.jsonTrack A: python3 verify_e86_offline.py.txt bundle → 14/14; Track B: 13/13 at clawx.click/evidence/e86-agi-infrastructure-2026-11-17/ | 2026-11-17 | Merge the tracks under one reproduced suite with external witnesses and a published third-party audit |
| The E86 frontier paired run was attempted under a hard $0.25 ceiling A credential file existed but was rejected by the provider. No frontier result exists. Mock rows are labelled MOCK and are not counted. | BLOCKED | BLOCKED_INVALID_CREDENTIAL | /proof/bundle/e86-governed-intelligence-foundation-2026-11-17/frontier/status.json; preregistration.jsonProvider probe returned authentication_error: invalid x-api-key; 0 inference calls, $0.00 spent | 2026-11-17 | Inject a valid credential through the approved secret environment; run the 20 pre-registered tasks under an approved ceiling |
Scorecard: today vs 2027
| Metric | Today | 2027 target |
|---|---|---|
| External reproductions | 0 | 3+ |
| Independent witness organizations | 0 | 2+ |
| Third-party security audits | 0 | 1 published |
| Vertical pilots with a real workflow | 0 | 3 |
| Governed productive hours | 1 h wall clock, 165 tasks | 10,000 verified productive hours |
| Concurrent model-driven workers | 10 simultaneous; 50/100 shared-model | 100+ simultaneous |
| Benchmark size | 36 tasks, 6 families, local models | 500 tasks, 10 domains, frontier + local |
| Remote model calls | 0 (BLOCKED, no credential) | preregistered frontier run under a dollar ceiling |
| Paying customers | founder-attested 2, $97 | 3 referenced pilots with invoices |
| Negative results published | 3 (collaboration, specialists, routing) | all results, including failures |
What would falsify the claim — invite the attempt
A governor that invites falsification is more credible than one that only publishes successes. The public Break CAIN page lists the eight concrete attempts and the current defense status.
- Get a valid ALLOW for an action that should be denied — denied at the Gate; no executable grant (BLOCKED in the E82 lab)
- Execute after revocation or expiry — GRANT_EXPIRED / IDENTITY_REVOKED (BLOCKED in the E82 lab)
- Replay a single-use grant — GRANT_REPLAYED (BLOCKED in the E82 lab)
- Commit from a stale worker — STALE (BLOCKED in cross-host and reboot drills)
- Read another tenant's memory — cross-tenant read refused (BLOCKED in the E82 memory lab)
- Make a memory record grant authority — grants_authority=false, authority_effect=none, executable=false (BLOCKED in the E82 memory lab)
- Tamper with a receipt and still pass the verifier — hash/Merkle/signature failure (10/10 mutations rejected)
- Produce a duplicate external effect on retry — idempotency key → DUPLICATE, one effect (BLOCKED in the lost-ack probe)
Sources
- E82 public evidence bundle — 240 files, 55/55 offline checks, Merkle
63868034b46e901f2cec392cd25c22ab947516e594543e02155857a9c31c5ac9. - Download the ZIP · Reproduce it in four commands.
- claims.json — the machine-readable ledger behind this page.
- Assurance report · Audit RFP — the external audit scope, rules and budget ranges.
The hard line: “infrastructure governor of AGI” (CAIN-42 Proto-VAGI) is an architectural role, not a regulatory designation, and CAIN-42 does not claim to have achieved AGI. The strongest current statement is: the E82 evidence is reproducible offline with no CAIN code; the external verification work is not yet done.
Owner: FIRME AI CORPORATION · generated from claims.json on 2026-11-17. Every row is editable; every label must match the artifact.