CAIN-42 · TWO LABS: AGI + RSIAGI: NOT ESTABLISHEDTOP EVIDENCE UPDATE #10 · REFRESH · 2026-11-21T18:19:32Z · next 2026-11-21T18:31:32Z120 updates/day · every 12 min · verified evidence only

MOST SIGNIFICANT AGI + RSI EVIDENCE

AGI LAB · GENERAL INTELLIGENCE
Continual learning beats no-memoryPASS
0.6667 vs 0.5556 · +11.1 pts · p=0.00049
Retention 0.667 vs 0.500; poison 1/1 refused. Ties retrieval/episodic — bounded claim.
evidence · continual_results.json
100-identity research labPASS
100 registered · 5 peak real workers · 60 tasks · 7 missions
177.87 s measured; 0 failed, 0 dead, 144 hash-chained events. Registration is not activity.
evidence · E127_RUN_REPORT.json
Verified collective intelligencePASS
0.666 vs 0.480 best single · 0/500 Byzantine flips
+18.6 points, p=2.7e-9; 22 agents banned; rule-based agents, not LLMs.
evidence · collective_results.json
RSI LAB · RECURSIVE SELF-IMPROVEMENT
E132 governed RSI loopPARTIAL
main 30/48 · matched fixed 28/48
Library 21→117; depth 1→8; tier-6 1.000 vs fixed 0.500; tiers 8/12/16 = 0.000; verifier 1,740/1,740.
evidence · artifacts/e132/README.md
E129 self-written libraryPASS
8→28→35→35 sealed solves
POSITIVE_BOUNDED; library 21→96, depth 1→8; 11/11 gates; stops growing at generation 2.
evidence · artifacts/e129/README.md
RSI ceilingPARTIAL
Tiers 8, 12, 16 = 0/6
Bounded constructed integer domain; no model trained; no external reviewer. Not unrestricted RSI.
evidence · artifacts/e132/README.md
48-HOUR SOAK: FAILED / NOT ESTABLISHED. The latest full-duration attempt (2026-11-18) ran 48.008 h but failed verification: 71 consecutive unreadable cluster checks and replicas ended at different heights [345052, 345053, 345054]. The earlier 2026-11-15 attempt failed at checkpoint 0 on a PBFT safety incident. There is no 48-hour pass to display. The supported long run is the 24.003-hour soak (2026-11-02).
No breakthrough is scheduled. A new top result appears only after a preregistered experiment, independent verification and a committed evidence record. E127 report · 20 gates · offline verifier · AGI+RSI evidence index · two-lab evidence
E128 · PUBLIC RESEARCH OBSERVATORYAGI NOT ESTABLISHEDDATA FRESHNESS: HISTORICAL SNAPSHOT · 2026-11-21T17:30:09ZMERKLE 7bb44c0c03c594… · FIRME AI CORPORATION

CAIN-42 E128 — THE EIGHT-WORKER GENERAL INTELLIGENCE RESEARCH CAMPAIGN

8 specialized workers executed 2000 preregistered trials (64 genuine local Ollama LLM trials, 9387 tokens, $0.00 paid). Explore every trial, failure, and Merkle proof.
SPECIALIZED WORKERS
8/8
VALID TRIALS
2000/2000
LOCAL LLM TRIALS
64
UNIQUE FAMILIES
87
INTERNAL REPRO
7
EXTERNAL REPRO
0
PAID SPEND
$0.00
KILL-SWITCH
VERIFIED
Open Interactive Experiment Explorer →Final Report26 Release GatesOffline VerifierEvidence RegistryHonest limits: constructed environments + local 0.5B–1.5B Ollama models; 0 external reproductions; external benchmarks BLOCKED; single-host SQLite queue.
E127 · LIVING RESEARCH LAB · PUBLIC OBSERVATORYAGI: NOT ESTABLISHEDSTATIC SNAPSHOT · 2026-11-21T16:21:27Z · RUN COMPLETED

CAIN-42 — THE OPEN GENERAL INTELLIGENCE RESEARCH LABORATORY

144 hash-chained events; 7 completed missions; evidence registry v39; failures remain visible.
REGISTERED IDENTITIES
100
CONFIGURED
8
PEAK ONLINE
5
PEAK ACTIVE
5
EXPERIMENTS DONE
60
MISSIONS DONE
7
EVENTS
144
UPTIME
177.9s
Honest state: 100 identities are registered, not 100 running agents. Peak 5 real workers; online is 0 after the run. Uptime is 177.9s, not 24/7. No experiment was fabricated. report · 20 gates · verifier
E126.b · EVIDENCE UPDATEAGI: NOT ESTABLISHEDSTATIC SNAPSHOT · 2026-11-21 · 30 canonical records
Registry now has 30 records: PASS 15 / PARTIAL 7 / FAIL 7 / INTERIM 1. Failures remain visible.
Continual learning with verified semantic memory · full 0.6667 vs no-memory 0.5556 PASSPre-task success calibration · Brier 0.0926 / ECE 0.0082 PASSExpected-information-gain intervention selection · active MAE 0.1101 vs random 0.1336 FAILNovel-environment confirmation, 200 instances · full 0.8950 vs retrieval 0.8600 vs E125 0.8200 PARTIAL
Persistent learning passed against no-memory but ties retrieval; active causal selection did not win its paired test; novel-environment H1 missed p<0.05. report · 25 gates · verifier
CLAWX · PUBLIC EVIDENCEAGI STATUS: NOT ACHIEVEDSTATIC VERIFIED SNAPSHOT · 2026-11-21

CAIN-42 RESEARCH EVIDENCE REGISTRY

Every research claim has a source. Every experiment has a result. Every result has limitations. Every published bundle can be inspected.

LATEST
E126
RECORDS
21
PASS
8
PARTIAL
6
FAIL
6
INTEGRITY
19/20
PRIORITY RECORDS · scientific status and evidence integrity are separate
IDEvoClaimScienceIntegrity
E125-BCE125Preregistered compositional generalization (B+C)PASSPASS
E126-ENVE126Novel-environment adaptationPARTIALPASS
E126-TRANSFERE126Safe transfer with zero false acceptsPASSPASS
E126-COLLECTIVEE126Collective advantage with verified evidencePASSPASS
E126-GOVE126Three-process authorization and kill switchPASSPASS
E126-DISCOVERYE126Eight scientific discovery cyclesPASSPASS
E126-CONTE126Persistent learning across 120 episodesFAILPASS
E126-CAUSAL-ACTIVEE126Active intervention selectionFAILPASS
E126-ADAPT-RETENTIONE126Prior-capability retention after adaptationFAILPASS
Latest researchEvolution historyGeneralizationContinual learningCausalTransferNeural adaptationCollectiveSelf-improvementLong missionGovernanceFailed experimentsBlocked experimentsReproductionExternal review

Do not confuse a verified hash with a scientific finding. E126 persistent learning and active causal selection failed; novel-environment H1 missed p<0.05; external benchmarks and external human review are BLOCKED. canonical registry · verifier · report

PENNY-FOR-PENNY EVIDENCE spent $168.17tasks generated 497resolved 257$ / resolved $0.65$ / task $0.34rank #75 / 136board →
measured SWE-bench Verified partial run · unattempted tasks count as unresolved · 0 resolved task ids in the feed · generated 2026-11-21T18:17:12Z · AGI NOT ACHIEVED
SWE-BENCH VERIFIED · CAIN-42 RANK #136 of 136 — 0.0% resolved (0/500) · 50% horizon — · official 500-task run Rank table → Evidence
CAIN-42 · VERIFIED WINS
CAIN-42
PROTO-VAGI
6-HOUR MISSION PASSED · 24-HOUR SOAK PASSED · 48-HOUR SOAK FAILED · AGI NOT ACHIEVED
6-HOUR 21,600 s · 39,449 tasks · 39,449-record chain, 0 breaks · 33,813 solved  |  24-HOUR SOAK 24.003 h · 0 anomalies · 0 divergences · 67 faults · 10,994 ALLOW / 177 DENY
VAGI E118 0.264→1.000 · E119 0.333→1.000 · E122 0.146→0.917 · E123-OW 0.188→0.906  |  8/8 BUNDLES VERIFY · 192/192 CHECKS
6h evidence → 24h evidence → Live VAGI test board → AGI evidence index →
LIVE 24/7 · — PASS · — FAIL · — RUNNING · — test runs · last win: — · next: —
48-hour soak (2026-11-18) failed its signed verdict and is disclosed. PASS = duration + journal integrity, not task success. AGI not achieved. No third-party audit.
E123 · OPEN-WORLD INTELLIGENCE BREAKTHROUGH · 2026-11-2016 GATES PASSVERIFIER PASS · 30/30 checks passedOVERALL PARTIAL (16 PASS / 1 PARTIAL / 3 BLOCKED)AGI: NOT ACHIEVED
E123 Open-World Bundle →20-Gate ScorecardOffline VerifierStudio MirrorE123 AGI Research Lab →Live Win Board →AI Investor Portal →All AGI Evidence (E111–E124) →
Capability Acquisition (G02–G03)PASS
32 unfamiliar task families (160 sealed instances, 12-stage cycle, 13-field Capability): Full CAIN E123 0.906 vs compute-matched 0.188 & no-causal 0.469 (Holm p=0.00e+00). Strategic Test (Software→Cybersecurity, n=36): acquired capability 1.000 vs base Qwen2.5-0.5B 0.389 & compute-matched 0.556.
Causal World Model & Transfer (G04–G07)PASS
48 confounded/shifting/noisy worlds: active EIG+HMM 0.917 vs 2x passive 0.000 (unpredictable abstain 1.00). 6 verticals: 10/12 positive transfers (5 structurally distinct), 12/12 harmful transfers rejected; 9-role discovery (4 cycles, 2 rejected, reusable skill proven).
Local LLMs, Metacognition & Memory (G09–G10)PASS
Real Qwen2.5-0.5B (0 weight updates, 0 authority, n=24): base 0.000 → complete CAIN 0.292. Fresh AMA-Bench (n=12) F1 0.2560→0.2444; fresh ARC-AGI-2: 0/6 (honest). 5-state metacognition Brier 0.1250 (vs 0.3121); 0/25 memory poison accepted; SIGTERM crash-recovery PASS.
Collective, Self-Improve & MCPGate (G12–G16)PASS
7-arch collective: dynamic team 1.000 vs best single 0.458; Byzantine memory contamination 0/30. Self-improvement: 24 candidates across 11 stages, 4/4 governance attacks blocked, 0 self-authorized. Multiprocess TCP MCPGate + WAL kill-switch: 14/14 PASS, 0 effects after KILL.

Honest limits & BLOCKED gates (FIRME AI CORPORATION · branch evolution-123-open-world · commit 1c7fcb6d6): G11 (12-hour wall-clock mission) BLOCKED; G18 (independent external human review) BLOCKED; G19 (separate administrative failure domain) BLOCKED; ARC-AGI-2 unsolved (0/6 on SmolLM2-135M CPU), ARC-AGI-3 unavailable, SWE-bench deferred to E124, hosted frontier APIs unavailable; 0 neural weight updates; enforcement tier is loopback_multiprocess_execution. AGI: NOT ACHIEVED.

CAIN-42 · PROTOTYPE VAGI & AGI INFRASTRUCTURE · FIRME AI CORPORATION AGI STATUS: NOT ACHIEVED

CAIN-42 (by FIRME AI CORPORATION) is prototype Vertical AGI (Proto-VAGI) and AGI infrastructure — combining:

  1. AGI Infrastructure (Governed Execution Runtime): the 7-moat Byzantine trust fabric, CAIN Studio, MCPGate default-deny enforcement boundary, and ClawX independently verifiable evidence registry (NO AUTHORIZATION → NO EXECUTION).
  2. Prototype VAGI Research Organism: experimental neural + symbolic abstraction synthesis, causal world modeling, provenance memory, cross-domain transfer, autonomous research loops, and governed long-horizon missions (E111–E124 published at cainstudio.online/agi-evidence-index.html), while keeping AGI Status: NOT ACHIEVED.
E122 · NEURAL GENERALIZATION & AUTONOMOUS DISCOVERY · 2026-11-20 OVERALL PARTIAL (15 PASS / 3 PARTIAL) Verifier: 37/37 checks passed AGI: NOT ACHIEVED
Real LLM Neural Abstraction (G03–G05)
Local PyTorch LLMs (Qwen2.5-0.5B/1.5B, SmolLM2-135M, 0 authority): sealed unfamiliar Python tasks (n=48) E120 symbolic 0.000, neural alone 0.146, one-shot 0.542 → CAIN neural organism 0.917 (Holm p=3.00e-11; overfit rejected).
External Benchmarks & Repairs (G06–G07)
AMA-Bench (MIT, n=20) Token F1: recency 0.3158 → CAIN provenance 0.3538 (+0.0380, 14W/5L). ARC-AGI-2 (Apache-2.0, n=8): 0/8 (unmerged, honest raw score). Real-source Python repairs: 0.00 → 1.00.
Causal, 10-Role Research & Transfer (G08–G16)
Neural+causal planning 0.250 → 0.750; 10-role 3-cycle research (override blocked); 5/5 positive + 3/3 negative-rejected transfers; continual learning (0 weight updates, 0 forgetting); 7-arm collective; 18 self-improve candidates.
Governed Neural MCPGate & 12h Mission (G17–G18)
Real Qwen2.5-0.5B malicious/benign tool calls enforced across separate OS processes via cain/mcp_proxy.py + SQLite WAL kill-switch: 9/9 pass, 0 effects after KILL; 12h neural mission active (1658.91s / 43,200s, 95 records, 0 dupes).

Honest limits & PARTIAL gates (FIRME AI CORPORATION · commit e65a75971): Frontier hosted APIs BLOCKED (no API key); 0 weight updates performed; ARC-AGI-2 unsolved (0/8 on 0.5B CPU); G02 E119 6h PASSED; E120 12h and G17 E122 12h neural missions remain PARTIAL while running real wall-clock time; G18 multi-org 2-of-3 quorum NOT met (external human review NOT YET PERFORMED). AGI: NOT ACHIEVED.

CAIN-42 · LEADERBOARD — #1 VERIFIABLE GOVERNANCE (only fully evidenced entry) · general autonomy 2.0 h on a CAIN-native suite, below the METR frontier (6.4–17.4 h) and not comparable · not a universal #1 Board E124 evidence E123 METR All evidence
E124 · CAIN LONG-HORIZON AGENT — 70B open model, **13/18** on a 6-task CAIN-native suite · 50% horizon **≈2.0 h** · still **below** the METR frontier (6.4–17.4 h) and **not METR-comparable** · governance/evidence axis: CAIN's measured lead Result →
CAIN-42 vs FRONTIER AGENTS · METR 50% TIME HORIZON
1 claude mythos preview early 17.4h · 2 claude opus 4 6 12.0h · 3 gemini 3 1 pro 6.4h · 4 gpt 5 2 5.9h … · CAIN-42: UNRANKED (no METR suite; 6h governed workload RUNNING) · Governance/evidence axis: CAIN's measured differentiator Full board →
E121 · INDEPENDENT VERIFICATION LAYER — E118 26/26 · E119 16/16 from clean copies · 6h PASS · external reviewer BLOCKED · independent admin BLOCKED · AGI NOT ACHIEVED Verification →
AGI EVIDENCE INDEX · E111–E121 — every claim with its verifier and limits · AGI: NOT ACHIEVED · Open the index →
E119 · LEARNED GRAMMAR + REAL TASKS — sealed 0.333→1.000 (gain +0.667, p≈1e-7; replicated +0.583) · 8 macros learned · 8 real tasks · 6h PASS (21,600 s; 39,449 tasks; journal intact) · external reviewer BLOCKED · AGI NOT ACHIEVED Evidence ↓
STRONGEST VAGI EVIDENCE SO FAR · E118 — unseen primitives + depth-3 compositions: OOD 0.264 → 1.000 (gain +0.736, p≈0; replicated +0.778) · 3/3 capabilities · memory 1.00 vs 0.17 · self-improvement 1.00 vs 0.17 · real data 0.627→0.972 · AGI: NOT ACHIEVED Evidence ↓
E120 · GENERALIZING SELF-IMPROVING VAGI ORGANISM · 2026-11-20 OVERALL PARTIAL (17 PASS / 3 PARTIAL) Verifier: 47/47 checks passed AGI: NOT ACHIEVED
Unified Generalization (G03–G06)
Zero task-ID shortcuts; 4 sealed conditions vs no-learning (budget 1200): A +0.625, B +0.438, C +1.000, D +0.500 (all Holm p<0.001); retention 1.000; RF ref 0.000.
Causal World Model & Research (G07–G08)
Confounded SCM planning 0.275 → 0.775 (p=0.000325); 8-role loop: AdversarialReviewer rejected director shortcut H1; accepted H2 (0.333→0.667).
Self-Improve, 5 Verticals & Transfer (G09–G14)
14 candidates classified (3 beneficial promoted, 3 unsafe governance blocks, rollback verified); 5/5 real-source verticals improved (Z3/SymPy/sklearn/AST); 4/4 positive transfers across 4 domains; 2/2 negative transfers rejected.
Collective, Memory & Real MCPGate (G15–G18)
Matched-compute collective 1.000 vs strongest single 0.483 (homo gain 0.0); 8-layer memory 1.000 vs 0.250 (0/20 poison); real cain/mcp_proxy + WAL kill-switch 9/9, 0 effects after KILL; 9/9 faults recovered.

Honest limits & PARTIAL gates (FIRME AI CORPORATION · commit 8c6ef11a3): G02 E119 6h mission PASSED (21,600 s / 21,600 s; 39,449-record journal intact); G17 E120 12h mission is RUNNING (2856.55 s / 43,200 s, SIGTERM resume verified 0 dupes); G19 clean-room verifier passes as nobody, external human review NOT YET PERFORMED. Constructed symbolic/grid suites + public datasets; not achieved AGI.

CAIN-42 PUBLIC AGI EVIDENCE (E111 → E116 · 2026-11-20) — E116 Organism (13 PASS / 7 PARTIAL / 2 BLOCKED · 54/54) · E115 Memory INTERIM (7/7 · paid BLOCKED) · E114 Lab 3.0 (6 PASS / 14 PARTIAL / 1 NOT TESTED / 1 FAIL · 24/24) · E113 Collective INTERIM/RUNNING (12/12) · E112 Lab 2.0 (7 PASS / 2 PARTIAL / 2 FAIL · 33/33) · E111 Maturation (7 PASS / 4 PARTIAL / 1 FAIL / 1 BLOCKED) · AGI Benchmarks · Auditor Index →
LATEST PUBLIC AGI EVIDENCE · E112 → E116 Verified Public Evidence Through E116 — Every Bundle Offline-Verifiable (0 CAIN Imports)

Honesty rule: CAIN-42 Proto-VAGI is governed AGI infrastructure built by FIRME AI CORPORATION. It is NOT yet AGI. Every card below reports measured outcomes from raw bundle files alongside negative, interim, and blocked results.

E116 · 2026-11-20PARTIAL Verifier 54/54 · 13 PASS / 7 PARTIAL / 2 BLOCKED
Self-Evolving VAGI Organism (Generalization, Memory, World Model, Transfer, Governance)

Sealed generalization: fixed baseline 0.306 → per-task 0.939 → world memory 0.983 → memory+prior 0.989. Acquired R_world_memory_policy: sealed 0.817 → 0.950 (p=0.02148), replicated 0.850 → 0.967 (p=0.03906). 5/5 skills promoted; memory reload 0.989, poison rejected; hybrid world model 0.876 vs single-table 0.522 (shift detected); routed specialists 1.000 vs best single 0.167 (redundant voting 0 gain); 5/5 transfer pathways; tabular adaptation 0.627 → 0.993; KILL 0 effects after order.

Limit: Disclosed finite grammar and constructed worlds; H2 rejected, H3 blocked; LLM fine-tuning BLOCKED; 6h interrupted & 24h BLOCKED; single host; no external reviewer; NOT AGI.

E115 · 2026-11-19INTERIM Verifier 7/7 · Paid BLOCKED
Verified Memory Organism (AMA-Bench Preregistration, Attack Audits, Rule Discovery)

Preregistered on external AMA-Bench (ICML 2026, official token F1; prereg SHA-256 31ae27decaa4…). Model-free retrieval diagnostic (dev, 4000-tok clean, novel-token recall): M0_none 0.000 / M1_naive 0.393 / M2_evidence 0.409 / M3_causal 0.391 / M4_validated 0.391 / R_recent 0.351. 5 hostile-write attack audits recorded at 5% attack rate. Sealed rule discovery: shift worlds 33/60 (open enumerator) vs 18/60 (fixed); frontier worlds (inexpressible in both grammars) 0/60.

Limit: Primary model-backed AMA-Bench scores do NOT exist yet: paid Claude arms BLOCKED (no API credit), local model arms PENDING; diagnostic recall is not an official AMA-Bench score; NOT AGI.

E114 · 2026-11-19PARTIAL Verifier 24/24 · 6 PASS / 14 PARTIAL / 1 NOT TESTED / 1 FAIL
AGI Research Lab 3.0 (Governed Research Loop, Active Causal Discovery, Cycle-2)

Information-gain experiment selection identified hidden causal rules in 10.28 vs 18.48 experiments for random (120 sealed worlds, 110 wins / 4 losses). Memory-prior transfer to new 6-variable worlds p = 0.009; poisoned prior rejected; routed experts beat naive continual learning 10/10 sealed seeds; 44/44 governance attacks held; 0 effects after KILL across 16 processes; SIGKILL resume 0 duplicates.

Limit: Constructed simulations; 0/5 self-improvement candidates promoted; both cycle-2 follow-ups negative (B2 gated skills, B1 grammar growth); no model inference; single host; NOT AGI.

E113 · 2026-11-19INTERIM · RUNNING Verifier 12/12 · Phase 3 Running
Model-Backed Collective Intelligence Study (Preregistered Matched-Compute Pool)

Preregistered study testing whether a model-backed governed collective beats the strongest individual agent under matched CPU compute (Qwen2.5-0.5B, Qwen2.5-1.5B, SmolLM2-135M; sealed split hash committed before sealed outputs). Governance attack lab: 18/18 cases refused; 5/5 multi-process KILL races with 0 effects after the fence (delegated-permit expiry bug found and fixed); E111 (87/87) and E112 (33/33) baselines reproduced.

Limit: Primary collective-vs-individual endpoint is still running (no primary intelligence result yet); small local CPU models only; cloud BLOCKED; single host loopback; NOT AGI.

E112 · 2026-11-19PARTIAL Verifier 33/33 · 7 PASS / 2 PARTIAL / 2 FAIL
AGI Research Lab 2.0 (2,000 Governed VAGI Agents, 1,000 Processes, Specialization)

2,000 unique VAGI agent identities across 6 verticals (identity grants no authority), 12,000 governed executions, 1,000 real OS processes (peak PSS 1456 MB), 10,000 hash-chained receipts, and 0 effects after KILL. Complementary specialization: routed 6-specialist team 1.000 vs best single agent 0.167; redundant voting 1.000 = one exact agent (negative collaboration result preserved).

Limit: Deterministic rule engines (not LLMs); accuracy constant 0.736 at every scale; fan-out slower on tiny tasks (15.832 s vs 1.07 s); local file-lock KILL, not production MCPGate; NOT AGI.

CAIN-42 PROTO-VAGI · INFRASTRUCTURE GOVERNOR OF AGI CAIN-42 Proto-VAGI is the infrastructure governor of AGI and a VAGI research platform. It is not yet AGI; its intelligence results are early and bounded. E116 organism →E115 memory →E114 lab 3.0 →E113 collective →E112 lab 2.0 →E111 maturation →AGI benchmarks →Auditor index →AGI dossier →Proof factory →
CAIN-42 PROTO-VAGI CAIN-42 Proto-VAGI is the infrastructure governor of AGI. CAIN-42 Proto-VAGI: early, measured signs of intelligence, and every action it takes comes with a signed receipt. Auditor index: every VAGI bundle, verifier re-runs, manifest hashes →
Early, measured signs of intelligence
  • E95: on a frozen held-out set, composing procedures it had not seen went from 0/90 to 79/90 after one harness improvement (two candidates rejected). evidence (Constructed tasks; a configuration change, not a weight update.)
  • E93: open procedure synthesis solved 369/510 held-out tasks versus 200 for selecting from a supplied library. evidence (At equal compute the two were about equal (208 vs 200).)
  • E105: it learned from its own experience where hidden preconditions sit (33.0% of its learned lift on the column holding 10% of cells), with no code that knows the answer. evidence (Learned, not yet useful: steering with it did not raise success (24 vs 27 of 100).)
  • E75: on held-out unfamiliar code repairs it solved 8/8, where the previous version solved 0/8. evidence (Narrow single-function repair; frontier tasks 1/4; the local model alone 1/8.)
  • E79: model-backed agents answered 35/104 tasks vs 10/104 for the deterministic baseline (about 3.5x). evidence (One 0.5B model; collaboration gave no significant gain; host address redacted (disclosed).)
  • E110 verified transfer: knowledge learned in one constructed domain improved the other in 4 of 5 pathways (e.g. 69 vs 38 solved), and harmful transfers were rejected in 5 of 5 replicates while blind transfer lost 12 incidents. evidence (Near transfer between two constructed domains; not the six real verticals.)
Every action comes with a signed receipt
  • 12 CAIN-42 VAGI Agents live: 35/35 published signed records verify offline with a verifier that imports no CAIN code. evidence (Deterministic rule engines; nothing executed on real systems.)
  • Authorization boundary: 27/27 adversarial cases pass, 0 unauthorized effects; kill switch: 0 effects after the KILL order. evidence (Loopback on one host, not the production MCPGate.)
  • E110 governed organism: every action of the E108 organism in 120 sealed incidents went through the hardened boundary: 9,240 signed, hash-chained receipts; outcomes identical to the direct runs (120/120); 0 effects after an independent KILL. evidence (Loopback on one host, not the production MCPGate.)
  • E110 collective: when 7 of 8 workers lie, an unverified team 'discovers' a rule that does not exist (p = 0.001); independent recomputation rejected all 17 fabricated rows and kept the verdict correct. evidence (Speed gain only (2.1-3.8x); no quality gain over one agent.)
  • E110 (in progress): hardened execution boundary passes 22/22 attack cases (the E102 boundary: 19); independent emergency KILL/REVOKE path; 0 effects after KILL in a 40-step race. Six-hour governed mission running. evidence (Interim; loopback, one host; no gate claimed PASS yet.)

Proto means early. CAIN-42 is governed AGI infrastructure; it is not yet AGI. The intelligence results are on constructed tasks, measured by the same project and not independently reproduced. Also on record: negative results (E103, E104, E106) in the proof factory.

EVOLUTION 110 · IN PROGRESS

The public research record of CAIN-42

Every evolution, experiment, benchmark, negative result, verifier and current blocker, navigable by status. 25 VAGI bundles today: 8 CONTEXT, 2 CORRECTED, 2 MIXED, 2 NEGATIVE, 4 RECEIPTS, 7 SUPPORTS.

Open the auditor index (by status, with manifest hashes and fresh verifier re-runs) · E110 interim · all evidence

E110 gate matrix (interim, no gate claimed PASS): FAIL 2 · NOT TESTED 2 · PARTIAL 12 · RUNNING 2. GATE_MATRIX.json · auditor index

PARTIAL G1 unified architecture
NOT TESTED G2 held-out performance
PARTIAL G3 continual learning
PARTIAL G4 causal reasoning
FAIL G5 skill acquisition (>=5)
PARTIAL G6 cross-domain transfer (>=3 of 5)
PARTIAL G7 metacognition
RUNNING G8 six-hour governed mission
PARTIAL G9 collective intelligence
PARTIAL G10 recursive improvement
PARTIAL G11 resource-side enforcement
PARTIAL G12 kill switch
PARTIAL G13 identity
PARTIAL G14 evidence
NOT TESTED G15 memory integrity
PARTIAL G16 independent verification
RUNNING G17 endurance
FAIL G18 operational independence

CAIN-42 Proto-VAGI is the infrastructure governor of AGI. Proto means early. CAIN-42 is not yet AGI; its intelligence results are early, bounded and on constructed tasks. Built by a three-person team (FIRME AI CORPORATION). Loopback results run on one host.

NEW · CAIN-42 VAGI AGENTS Governed vertical agents for 12 industries: 12 of 12 live Open the agents →

Each CAIN-42 VAGI Agent proposes the work for your case, decides every action — ALLOW, HOLD FOR HUMAN or DENY — under published rules, and returns a signed, hash-chained record you can re-check offline. Other agents ask for a click; these agents give you proof.

Verify the claims: evidence bundle (signed samples, rule sets, tests) · offline verifier · signing key · live API · auditor index (all VAGI evidence)

CAIN-42 is governed AGI infrastructure. It is not yet AGI; its intelligence results are early and bounded. The agents are deterministic rule engines; nothing is executed on real systems; prices are proposals; no customers, audit or certification are claimed for them.

AGI INFRASTRUCTURE EVIDENCE CAIN decides, the resource enforces, every effect is recorded — measured, with limits. All infrastructure evidence →
Authorization boundary (CAIN decides, resource enforces)
27/27 adversarial cases pass; 0 unauthorized effects (forged, replayed, expired, mismatched permits; revocation; scope escalation; partition fails closed)
Limit: Loopback, three OS processes on one host; not the production MCPGate (cain/mcp_proxy).
Kill switch during an active mission
0 effects after the KILL order in the resource's own hash-chained ledger; mission refused with KILL_SWITCH_ENGAGED
Limit: Same host; the first run used a looser counter and was replaced by the ledger-order check (disclosed).
Byzantine worker, independent verifier
30 false success claims, all rejected; 0 honest claims rejected; only verified experience enters shared memory
Limit: The verifier replays deterministic worlds; one Byzantine worker of four.
Consensus cluster (PBFT, n = 3f + 1)
proof 2491 at 2026-11-19T16:53:09Z: cluster cain-mr-01, agreement True, 3 signing, health MINIMUM_QUORUM
Limit: Vendor-operated validators in one admin domain; safety incident 2026-11-15 (conflicting certificates on cain-mr-02, quarantined) and view incident 2026-11-18 on mr-03.
Public anchoring of daily evidence
daily evidence root recorded in Sigstore's public Rekor log; standalone verifier published
Limit: Anchors integrity of the record, not the truth of its contents; 0 independent witness cosignatures.
Persistent governed mission (6 h)
0.5545 h, 1343/1541 episodes, 1 process starts, finished False
Limit: One host; every action through the loopback boundary; forced kill and resume at the halfway mark.
Earlier infrastructure bundle (E86)
governed intelligence foundation bundle (2026-11-17)
Limit: Its source-code folder was withdrawn on 2026-11-17 (source stays private); see its WITHDRAWN notes.
ALL EVIDENCE 332 evidence bundles · 19,792 files · 287/287 signed-claim pins match Open the full registry → catalog generated 2026-11-19T16:50:21Z
Product & engine evidence (238)
Consensus & cluster (38)
Soak & endurance (28)
Registry, provenance & audit kits (17)
Proof loop (detect / fix / prove) (11)
VAGI PROOF FACTORY Every VAGI claim has a run, a verifier, and a published failure boundary. Open the full factory → CAIN-42 Proto-VAGI is the infrastructure governor of AGI. It is not yet AGI; its intelligence results are early and bounded.
VAGI ORGANISM · ALL EVIDENCECAIN-42 Proto-VAGI: early, measured signs of intelligence, and every action it takes comes with a signed receipt.From a single pulse to an organism that learns hidden rules from its own experiments — every step, including the negatives and corrections.
E98 · PulseGAIN
A learned world model improves closed-loop decisions.
35/50 model-based vs 4/50 random (+62.0 pp)
E99 · Rapid heartbeatCORRECTED
Recurrent advantage did not survive E102's re-score; the world model did.
GRU 93/84/88 of 100 vs MLP 68/63/70 vs random 20/24/23
E100–E101 · Integrated organismCORRECTED
Measured on already-solved environments; re-scored by E102.
Delayed 52/60 vs MLP 43/60; hysteresis 58/60 vs 56/60
E102 · Persistent governed organismGAIN
Learns the world while it acts; beats frozen and reactive baselines.
252/300 vs frozen 198/300 vs reactive 154/300 (McNemar p=8.1e-09); 40 poisoned memories: undefended -44.5 pp, causal -1.0 pp
E103 · Transfer and skillsMIXED
Cross-domain transfer positive; skills failed held-out.
service -> grid: faster on 56 vs 14 tasks (p=4.3e-07); grid -> service: 78 vs 52 (p=0.03)
E104 · Strategy switchNEGATIVE
Rejected by the preregistered gate, which prevented a loss.
sealed grid 256 vs 257 of 300; key-and-door 35 vs 35; service lock 18 vs 18; plain service 54 vs 65 (a loss the gate prevented)
E105 · Where preconditions areNEGATIVE
Learned the drain column, but it did not raise success.
edge-lock 24 vs 27 of 100 with vs without the prior; uniform-key worlds 25 vs 26 of 60 (no harm)
E106 · Subgoal + permutation gateSTOPPED
Stopped before a sealed run: the gate was inverted. Nothing claimed.
development only; no sealed run
E107 · Runs its own experimentsGAIN
Interventional attribution + valid gate: more incidents solved, none lost.
edge-lock 91 vs 64 of 200 (27 won, 0 lost, p = 1.5e-08); uniform-key worlds unchanged 57 vs 57 and 74 vs 74
E108 · Learns many kinds of hidden rulesGAIN
Selection-aware gate; four rules in three frames found and used; none lost.
edge 72 vs 41/150 · row0 77 vs 48/150 · start_row 98 vs 74/150 · barrier_adj 101 vs 85/150; uniform controls unchanged (63 vs 63, 57 vs 57)
E110 · Every action receiptedRECEIPTS
Organism acts only through the hardened boundary; a signed receipt per action; KILL fences it.
9240 receipts for 9120 actions + 120 opens; parity with direct runs 120/120; permit effects after KILL 0; pre-issued permit after KILL refused
E110 · Transfers what it learnedGAIN
Laws carried to a new domain, checked by 3 experiments: 4 of 5 directions gain; wrong laws rejected.
P1 99 vs 76 · P2 88 vs 71 · P3 69 vs 38 · P4 73 vs 38 · P5 63 vs 60 (of 100); N1 verified 51 = none 51, blind 39

Every step links to its evidence bundle and offline verifier (hover for limits). TEST results on constructed tasks; no external reviewer. Receipts: the governed boundary issues an Ed25519-signed receipt for every action it admits (E102, loopback); the E107/E108 sealed runs called the world directly and carry no per-action receipts. CAIN-42 Proto-VAGI is the infrastructure governor of AGI; it is not yet AGI. Source code is not published. Auditor index →

E110 verified transfer between domainsTEST
P1 99 vs 76 · P2 88 vs 71 · P3 69 vs 38 · P4 73 vs 38 · P5 63 vs 60 (of 100); N1 verified 51 = none 51, blind 39
A law learned in one domain, carried as one hypothesis and checked against 3 experiments in a new domain, raised solved incidents in 4 of 5 directions with no losses; where the law did not hold the check rejected it every time (blind transfer lost 12 incidents there).
Limit: Near transfer between two constructed domains (grid, service lock); reverse direction P5 not significant; G6 reported PARTIAL; no external reviewer. Not AGI.
E110 every organism action carries a signed receiptTEST
9240 receipts for 9120 actions + 120 opens; parity with direct runs 120/120; permit effects after KILL 0; pre-issued permit after KILL refused
The E108 organism ran 120 sealed incidents with every action routed through the hardened resource boundary: signed permit, resource-side check, Ed25519-signed hash-chained receipt. A KILL sent to the resource on an independent key stopped every later effect while the decision service still answered ALLOW.
Limit: Loopback, OS processes on one host (not separate failure domains, not the production MCPGate); a subset of E108 sealed incidents; no external reviewer. Not AGI.
E108 the organism learns many kinds of hidden rulesTEST
edge 72 vs 41/150 · row0 77 vs 48/150 · start_row 98 vs 74/150 · barrier_adj 101 vs 85/150; uniform controls unchanged (63 vs 63, 57 vs 57)
From its own experiments the organism found four different hidden laws (a column, a row, its start's row, next to the barrier) in three frames, stayed silent on pattern-free worlds with a selection-aware null, and used each law: 100 incidents won, 0 lost.
Limit: Coordinate laws in designed frames, one constructed domain family; uniform gates off at p 0.318 / 0.21; no per-action receipts in these runs; no external reviewer. Not AGI or VAGI.
E107 the organism runs its own experimentsTEST
edge-lock 91 vs 64 of 200 (27 won, 0 lost, p = 1.5e-08); uniform-key worlds unchanged 57 vs 57 and 74 vs 74
In practice the organism tests one suspect cell at a time to find what opens a blocked passage, learns where such causes tend to be (found the drain column on its own; a permutation gate stays off on worlds with no pattern), and then uses it: 91 vs 64 of 200 sealed incidents solved.
Limit: One regularity type in two constructed domains; the no-pattern gate was off at p = 0.018 (thin margin, likely selection); 1 of 84 edge experiments mis-attributed (disclosed); no external reviewer. Not AGI or VAGI.
E105 learning where hidden preconditions areTEST
edge-lock 24 vs 27 of 100 with vs without the prior; uniform-key worlds 25 vs 26 of 60 (no harm)
From its own experience the organism located the hidden drain configuration (33.0% of its learned lift on the drain column, which is 10% of cells), but using that knowledge did not raise success.
Limit: Knowledge learned, not yet useful; the no-false-pattern guardrail gave a false positive and needs a shuffled-null test; no external reviewer.
E104 hidden-precondition discovery — rejected by the gateTEST
sealed grid 256 vs 257 of 300; key-and-door 35 vs 35; service lock 18 vs 18; plain service 54 vs 65 (a loss the gate prevented)
A strategy switch that looked strong on 16 development tasks was rejected by the preregistered paired gate; the sealed test confirms it does not help and would have hurt one domain.
Limit: Negative result: the hidden-key bottleneck remains open; the bandit's unpaired estimate was misleading; no external reviewer.
E103 cross-domain transfer and skillsTEST
service -> grid: faster on 56 vs 14 tasks (p=4.3e-07); grid -> service: 78 vs 52 (p=0.03)
Experience moves between two verticals: service-configuration memory makes grid tasks faster, and grid memory speeds service incidents on more tasks than it slows. Learned skills did not survive the sealed test.
Limit: Gates 2 PASS / 3 PARTIAL / 3 FAIL; near transfer between analogous domains; no success gain; skills and skill composition FAILED on held-out tasks; no external reviewer.
E102 persistent organism — learns the world while it actsTEST
252/300 vs frozen 198/300 vs reactive 154/300 (McNemar p=8.1e-09); 40 poisoned memories: undefended -44.5 pp, causal -1.0 pp
An online world model (hypothesis posterior, change points, causal memory) beats a frozen model on 300 sealed unfamiliar worlds; causal memory resists poisoning.
Limit: Gates 5 PASS / 10 PARTIAL / 5 FAIL; info-gain probing did not beat random probing; no cross-domain transfer or skills yet; loopback governance, not production MCPGate; no external reviewer.
Integrated organism — held-out family (corrected by E102)TEST
Delayed 52/60 vs MLP 43/60; hysteresis 58/60 vs 56/60
CORRECTED 2026-11-19 by E102 Phase 0: these numbers came from environments that were already solved before the agent acted. Re-scored correctly, GRU 53/60, MLP 57/60, random 4/60, few-shot 56/60 — the recurrent advantage and the adaptation failure do not hold. Original bundle kept unchanged.
Limit: Skill 45/50; failure F1 0.6667; cross-domain and 6-hour BLOCKED; no external reviewer.
Latent generalist — rapid heartbeat (corrected by E102)TEST
GRU 93/84/88 of 100 vs MLP 68/63/70 vs random 20/24/23
CORRECTED 2026-11-19 by E102 Phase 0: these numbers came from already-solved environments that every arm reused in turn. Re-scored correctly, the committed GRU solves 88/86/86 vs random 8/16/12 — the world model is real — but a retrained MLP scores 87/86/86, so the recurrent advantage does not hold. Original bundle kept unchanged.
Limit: Six gain families; skill 32/50; natural-failure F1 0.5455; no external reviewer.
Learned intelligence pulseTEST
35/50 model-based vs 4/50 random (+62.0 pp)
A trained residual neural world model improves closed-loop decisions on a sealed control set.
Limit: Three gain families; skill synthesis 7/24; natural-failure F1 0.0; no external reviewer.
Governor holds at the actionTEST
0 effects after the kill switch; 0 duplicate worker seeds
Every real E98 action was Ed25519-authorized and 3-of-4 BFT-notarized; the kill switch stopped execution.
Limit: In-process CAIN/MCPGate bridge, not a live network MCPGate request.
Evaluation integrityTEST
E96 solver 0/18 on unseen schemas; E97 seed-collision fix 0; skill selection labelled
We audited our own benchmark and downgraded it instead of defending it.
Limit: Same-team audit; no external reviewer. E96 90/90 is template-matched, not general.
VAGI Behavior SuiteTEST
library 500/500 in-distribution; shift 5/100
A 10-domain, 500-task constructed suite with a published failure boundary.
Limit: Hand-written procedure library; constructed tasks; the 5/100 shift is a coverage gap.
Recursive harness improvementTEST
E95: 0/90 → 79/90 held-out composition; familiar 60/60
A measured harness change improved held-out composition, with two candidates rejected.
Limit: Constructed numeric substrate; in-distribution composition; no external reviewer.

Labels: LIVE = production activity; TEST = executed test; SIMULATED = synthetic; BLOCKED = unavailable. No lane is LIVE. No AGI capability claim. CAIN-42 Proto-VAGI is the infrastructure governor of AGI; it is not yet AGI. Every lane is TEST unless marked LIVE; simulated activity is never shown as live.

E98 · VAGI INTELLIGENCE PULSE · MEASURED A learned neural world model moved sealed closed-loop success from 4/50 to 35/50. Evidence + verifier → Mirror →
35/50
model-based, sealed
4/50
random baseline
+62.0 pp
closed-loop gain
0.017954
one-step state MAE
0.022124
calibration ECE

What the pulse is: a trained residual neural ensemble (6902 transitions), closed-loop planning on 50 sealed environments, and a +62.0 pp gain over random. The governor held: every real action was Ed25519-authorized, 3-of-4 BFT-notarized, and the kill switch produced 0 effects after the fence. 10 real OS workers, 0 duplicate seeds.

What it is not: general VAGI intelligence. Skill synthesis is 7/24; natural-failure F1 is 0.0; the environment has three gain families; no external reviewer. These failures are on the front, not hidden.

E95 · RECURSIVE HARNESS IMPROVEMENT A measured weakness, five candidates, two rejections, one frozen held-out gain. Evidence + verifier → Mirror →
79/90
held-out composition, selected harness
0/90
held-out composition, atom baseline
60/60
familiar non-regression (baseline 60/60)
9→9
macros after backward revision

What actually changed: the cognitive harness gained a conditional composition operator over validated actions and an adaptive atom-first fallback. The engine observed 0/90 on development composition, generated five candidates, rejected depth-2 (no gain, higher cost) and unguarded composition (familiar regression), and promoted the adaptive harness. Two large nested-composition families remain unsolved.

Limits, on the front: constructed numeric substrate; harness configuration, not model weights; in-distribution composition over the same development atom pool; one seed set; no external reviewer, no witness, no third-party audit. This is not AGI, not VAGI, and not unrestricted recursive self-improvement.

VAGI PROGRAM · EVIDENCE OF RECORD 10 domains. 500 held-out tasks. The generalization gap is published, not hidden. Raw rows + standalone verifier → Per-domain table →
500/500
procedure library, in-distribution
5/100
unfamiliar variants (distribution shift)
72/500
no-procedure floor
112/500
one fixed procedure
481/500
raw Haiku 5.5, same policy card
74/100
raw Haiku on the shift probe (library 5/100)

Frontier baseline, same policy card, no tools: Haiku 5.5 481/500 (96.2%) and Fable 5.2 96/100. That is why the 500/500 library score is not a frontier-model discriminator; the 5/100 shift is the honest gap. Model baseline + verifier →

Shift probe, raw model: the library scores 5/100; Haiku 5.5 scores 74/100 and Sonnet 5.5 81/100. So the 5/100 is largely a library-coverage gap, not a general-intelligence gap. The probe is adversarial and partly underspecified: the software hidden target is not inferable, and all models scored 0/10 on tool orchestration. Shift baseline + verifier →

Check it here: the button recomputes SHA-256 of /vagi-suite.json in your browser and re-checks every headline count. No CAIN code, no server trust required for the arithmetic.

Not checked yet. The check is deterministic; a failure is shown in red, never hidden. pinned sha256 56876ef17f245a9d…

Limits, on the front: 500/500 is a hand-written procedure ceiling on a constructed suite, not a generalization result. The shift probe is 5/100 for the library, but raw models score far higher, so that number is a library-coverage map, not an AGI test. Gate G3 = PARTIAL: no external verifier, no independent witness. This is not evidence of AGI or VAGI. Self-reported by FIRME AI CORPORATION.

Per-domain counts (in-distribution / distribution shift)
domainin-distributionshift
software repair50/505/10
security triage50/500/10
data analysis50/500/10
business decision50/500/10
science calculation50/500/10
planning50/500/10
memory retrieval50/500/10
tool orchestration50/500/10
recovery50/500/10
governance50/500/10

Suite seed 9402 · generated 2026-11-18T13:59:49.980105+00:00 · 500 held-out tasks · governed arm 500/500. verify_offline.py · daily VAGI evidence ledger

E93 · DAY 1–2 · ALL EVIDENCE · NOT VAGI All the VAGI evidence so far, honestly measured. Raw rows + verifier → Daily VAGI evidence →

Shown: open search with no templates found new procedures that scored 369/510 held-out tasks vs 200 for picking from the supplied library and 87 with no procedures; replicated on new families (367 vs 223 / 114). More verified experience gave steadily better results (87 → 224 → 250 → 311 → 311 → 369). Software bug-fixing: learned from experience 16/30 vs fixed strategy 3/30 (p = 0.002). Kill switch: 20/20 trials, 0 effects after the fence. Claim register · 16-dimension scorecard · 104/104 checks recomputed from raw rows.

Limits: at equal compute synthesis only matches the library (208 vs 200); synthesis tasks are in-distribution; software is at 53%, below the 70% bar; 0 positive transfers; a local model's 34 proposals were all invalid; same team, no external review. Real-domain runs and collaboration are not done yet; a 6-hour mission is in progress. Current designation: Governed agentic intelligence infrastructure (outcome 3 of 3 in the E93 rule). No AGI, VAGI, production or certification claim.

E92 · CORRECTION · NOT DEMONSTRATED We withdrew our E92 procedure-synthesis result. Read the correction → Daily VAGI evidence →

Earlier today this banner said CAIN-42 had generated 3 procedures with no designer-supplied solution (120/120 vs 70/120). A code review found that the "generated" procedures came from four hand-written templates, each the full solution to one task family, and that the three comparison arms all ran the same untuned default program. Procedure generation is not demonstrated. E92 is NOT RELEASED, and gates 3, 4, 5 and 16 are FAIL.

The governance lab (15/15 in-process), the typed DSL and promotion pipeline, and the recomputation of the E91 baseline still stand. Real software defects, the 6-hour mission and independent verification remain BLOCKED. We make no AGI, VAGI, production or certification claim. A real synthesis experiment, with an open search and real baselines, is the first gate of E93.

FOR AI FOUNDERS AND INVESTORS · FIRME AI CORPORATION

CAIN-42 is building the governed infrastructure for vertical general intelligence. Target: 2027.

CAIN-42 gives AI agents memory and enforceable limits. On an independent benchmark it doubled a frontier model's score over no memory, at about one-seventh the cost of the brute-force approach, and every number is published with a verifier you can run yourself.

Autonomous AI agents will not be trusted with real work until something establishes who they are, what they may do, stops them at the point of action, and keeps a record anyone can check. CAIN-42 is that layer, and it is model-independent. Our 2027 goal is Governed Vertical General Intelligence (VAGI) infrastructure: persistent, learning, multi-agent intelligence that stays under enforceable authority. Here is where we stand, measured. Prepare for 2027: the governance layer is being built now, and the evidence is open for you to check.

Measured this week
CAIN memory raised Claude Opus 5.5 on Continual Learning Bench (an outside benchmark) from 0.220 to 0.439 per scan, better on 90/90 scans; reproduced in a fresh run (0.401).
CAIN used 7% of ICL's input tokens and $2.47 vs $17.90 per run: 3.8x more gain per dollar.
Plain in-context learning still scored higher (0.637); CAIN wins on cost, not yet on raw score.
New: CAIN hybrid memory on Claude Opus 5.5 reached 0.478/scan, 75% of plain ICL's 0.637, at $3.81 vs $16.93 per run (23% of the cost; average of 3 runs, which varied widely).
New hybrid memory on Claude Haiku 5.5 (recent context + CAIN long-run table): 0.430/scan, 94% of plain ICL's 0.459 at 29% of its cost, CAIN structured 0.420, basic retrieval 0.220, no memory 0.220.
Second task (CL-Bench poker, Claude Haiku 5.5, 1 run each; profit in big blinds per hand): CAIN hybrid +0.24 vs its no-memory +0.35; plain ICL +0.05 vs +0.21. Second poker run: CAIN hybrid +0.20 vs +0.31; plain ICL +0.02 vs +0.76.
6-hour faulted mission running now; results posted when complete.
Road to VAGI infrastructure (2027)
  • BUILTGoverned execution: identity, authority, resource-side fence, kill switch
  • PARTIALLearning from experience on an outside benchmark (beats no-memory; not yet plain ICL)
  • MEASUREDCost efficiency of learning (gain per dollar vs the best simple method)
  • IN PROGRESSLong-horizon autonomy: 6 h now; 24 h and 72 h next
  • NEXTCross-domain transfer and collective (multi-agent) gains
  • SEEKING PARTNERIndependent third-party evaluation
Check it yourself
Every number above comes from a signed bundle with an offline verifier (84/84 checks). The weak results are published next to the strong ones.
Self-run results; independent evaluation not yet obtained. VAGI is a 2027 goal, not a current capability. This page is not an offer to sell, or a solicitation of an offer to buy, any security; any offering is made only through definitive documents to eligible investors. Forward-looking statements are goals, not guarantees.
42 claims, each with its proof — open the investor evidence portal
  1. CAIN memory doubled a frontier model's score on an outside benchmark MEASURED
  2. On a frontier model, CAIN hybrid memory reached 75% of the best method's score at about 23% of its cost MEASURED
  3. 3.8x more learning gain per dollar than the brute-force method MEASURED
  4. The benchmark gain reproduced in a separate run MEASURED
  5. Hybrid memory: 94% of the best method's score at 29% of its cost MEASURED
  6. Kill switch: 0 effects after the fence at up to 250 worker processes REPRODUCED
  7. 11 of 11 prior headline results reproduced from raw data REPRODUCED
  8. Signed evidence bundle with an offline verifier that catches tampering VERIFIED
  9. MCPGate runs a tool call only with a consensus-committed authorization bound to that exact action VERIFIED
  10. A 4-node Byzantine-fault-tolerant cluster produced authentic 3-of-4 quorum certificates VERIFIED
  11. Live network-partition tests: an isolated or split cluster commits nothing VERIFIED
  12. Safety held under 10% packet loss, delay, duplication and reordering VERIFIED
  13. Two replicas lost their storage at once and were rebuilt from off-host backups VERIFIED
  14. Live rollback to the previous engine and forward again, one replica at a time VERIFIED
  15. The cluster image rebuilds bit-for-bit from its commit VERIFIED
  16. Every hosted decision record since 2026-09-27 is signed VERIFIED
  17. TLA+ models of consensus and the MCPGate gate, model-checked exhaustively VERIFIED
  18. The consensus fast path's view-change rule was model-checked VERIFIED
  19. DAG data is availability-certified and ordered identically on all 4 nodes VERIFIED
  20. A live 3-member etcd coordination cluster across three US regions VERIFIED
  21. An agent acts only inside a zone of authority the live cluster authorized VERIFIED
  22. Authority leases are invalidated under 9 tested conditions VERIFIED
  23. When authority lapses, the agent's next tool call is refused VERIFIED
  24. A policy change becomes active only after the live cluster commits it VERIFIED
  25. The system governor runs in the production gateway and refuses ungoverned calls VERIFIED
  26. An agent's model or tool change is gated in production; authority is never evolvable VERIFIED
  27. A caught attacker no longer gains autonomy on the production gateway VERIFIED
  28. A public proof fabric anyone can verify: signed manifest and per-file digests VERIFIED
  29. 28 executable safety invariants: no quorum, no authorization, no execution TESTED
  30. 22 kinds of agent operation run through one explicit governed pipeline TESTED
  31. Continuous trajectory governance: only a governance key, never the agent's, authorizes TESTED
  32. Signed, versioned agent identities with key lifecycle and scoped authority TESTED
  33. Changes to an agent's memory, skills, tools or routing must pass a governed gate TESTED
  34. System governance: emergency freeze and quorum-governed state TESTED
  35. Every allowed action yields a proof-of-governance record TESTED
  36. A 17-step governed learning loop over real governance outcomes TESTED
  37. Proof-carrying agency: each governed operation yields one signed proof TESTED
  38. Execution mesh: identity, capability and authority checked before every execution TESTED
  39. Governed agent factory: a mission compiles into a bounded agent specification TESTED
  40. Governed multi-agent action: agents act on the world only through governed decisions TESTED
  41. A recorded real run of the defined CAIN-42 test suites TESTED
  42. We attacked our own decision path, found holes in nine areas and fixed them TESTED
E91 · LEARNS FROM VERIFIED EXPERIENCE · 6 DOMAINS Learning from verified experience: governed agent 114/160 (71.2%) → 153/160 (95.6%) on 160 unseen tasks in 6 domains (exact McNemar p<0.001); all of the gain came from promoted, checker-verified procedures (no change on tasks without one: 44→44 of 50). Cross-domain transfer: a skill learned only in software fault isolation raised industrial fault diagnosis 5/10→10/10; experiment design learned in science raised business pricing decisions 7/10→10/10; on look-alike tasks where they do not apply the gate refused them (blind reuse would score 2/10 and 0/10). Causal reasoning 40/40 vs 30/40; 41/41 attacks blocked; kill switch 0.15 ms, 0 effects after; 59.4-min faulted mission, 7,200 verified, 0 duplicates. Multi-agent: no collaboration benefit demonstrated (single agent 10/20 beat every team under a shared tool budget). Claude Haiku 5.5, $0.31 total. Constructed tasks, self-run, PARTIAL EVIDENCE; no independent witness. Verify E91 bundle (91/91) → Limitations →
E89 · GOVERNED AUTONOMOUS INTELLIGENCE INFRASTRUCTURE External benchmark (Continual Learning Bench, not written by us): CAIN memory raised Claude Opus 5.5 from 0.220 to 0.439 per scan, better on 90/90 scans, 3 runs. Plain in-context learning with the same model scored higher: 0.637/scan (beat CAIN on 82 of 90 scans, 1 run). CAIN used 7% of ICL's input tokens and $2.47 vs $17.90 per run: 3.8x more gain per dollar. New: CAIN hybrid memory on Claude Opus 5.5 reached 0.478/scan, 75% of plain ICL's 0.637, at $3.81 vs $16.93 per run (23% of the cost; average of 3 runs, which varied widely). Replication by us (fresh process, 2 runs): 0.401/scan. New hybrid memory on Claude Haiku 5.5 (recent context + CAIN long-run table): 0.430/scan, 94% of plain ICL's 0.459 at 29% of its cost, CAIN structured 0.420, basic retrieval 0.220, no memory 0.220. Second task (CL-Bench poker, Claude Haiku 5.5, 1 run each; profit in big blinds per hand): CAIN hybrid +0.24 vs its no-memory +0.35; plain ICL +0.05 vs +0.21. Second poker run: CAIN hybrid +0.20 vs +0.31; plain ICL +0.02 vs +0.76. 6-hour faulted mission: in progress. E88 reproduced 11/11 from raw rows. Goal: VAGI infrastructure by 2027 — not yet demonstrated. Self-run, 0 independent witnesses, not released. Verify E89 bundle (84/84) → Benchmark numbers → Limitations →
E86 · AGI INFRASTRUCTURE FOUNDATION Two complementary tracks: (A) governed intelligence foundation, 14/14 offline; (B) frontier + atomic kill fence, 13/13. Both NOT RELEASED · 0 independent witnesses · no AGI claim · no production deployment. E86 evidence registry → Verify track A → Verify track B →
E85 EVIDENCE Evolution 85 measured: 100 held-out tasks, 9 arms, 2,122 real local model calls, 0 remote; ten-scope kill switch; 1 h endurance. Signed bundle, 75 files, 43/43 offline checks. Decision: NOT RELEASED; 0 independent witnesses. Verify bundle → Final report → Acceptance matrix →
AGI INFRASTRUCTURE · E86 An external agent registered, received scoped authority, executed through the governed boundary, and produced a receipt verified offline; the mission completed only with independent evidence. Fenced stop: 0 effects after the engage transaction at 50 workers. Frontier Sonnet 5.5: 65/100 with validated strategies. 43 tests. Not AGI achieved; not independently certified; NOT RELEASED. Verify E86 bundle → Final report →
GOVERNOR OF VERTICAL AGI LIVE EVIDENCE SURFACE E82 · 2026-11-17 FIRME AI CORPORATION loading clock…

CAIN-42 Proto-VAGI is the infrastructure governor of AGI.

Not another model. CAIN-42 is the authority, enforcement and evidence layer that decides whether a vertical AGI may act — then proves what actually happened. Every consequential action is bound to an identity, a delegated authority, an intent, a policy decision, a single-use capability and an observed effect. The record is signed, hash-chained and re-checkable by a verifier that imports no CAIN code.

ZERO-CAIN VERIFIER THIRD-PARTY VERIFICATION KIT 0 EXTERNAL REPRODUCTIONS · BE THE FIRST 240-FILE SIGNED BUNDLE 55/55 OFFLINE CHECKS 10/10 TAMPER MUTATIONS REJECTED 86/86 SAME-PROJECT LAB BLOCKED 0 REMOTE MODEL CALLS MERKLE 63868034…
60-SECOND VERIFICATION 1 · Drag the ZIP into your browser → SHA-256 2 · node verify_e82_node.mjs bundle → 40/40 3 · python3 verify_e82_offline.py bundle → 55/55 Open the verifier →

Live governor console

REAL ENDPOINTS · NO MOCK VERDICT

Loading the live scenario list from /fabric/try…

Your text selects the nearest registered test scenario. The page never executes your text.

Pick a scenario, then run the real decision pipeline.

The verdict is whatever the live /fabric/try endpoint returns. If it is down, the widget says so and shows no verdict. See the live stage list →

Live surface

FETCHED NOW
Decision pipelineCHECKING…
PBFT quorumCHECKING…
AETHER-33 invariantsCHECKING…
Public manifestCHECKING…
The governor's contract. No authorization → no execution. Only ALLOW executes. Unknown, error, expired, revoked, replayed or stale-fenced states never become permission. A model, memory, router, reputation or skill may propose; none may authorize.

“Governor” is an architectural role: it authorizes, enforces, records and verifies at the action boundary. It is not a regulator, a certification body or a government designation.

Download the evidence. Run the verifier. No CAIN code.

REPEATABLE · OFFLINE · SHA-256

1 · Signed evidence bundle

240 files: raw rows, model ledger, mission events, recovery records, gates, Merkle root and Ed25519 signature. 1.1 MB ZIP.

⬇ CAIN42_E82_EVIDENCE_BUNDLE.zip

2 · Standalone verifier

Python standard library + cryptography. Imports no CAIN module. Recomputes the headline numbers from the raw rows.

⬇ verify_e82_offline.py

3 · Manifest, signature, key

Check the bundle against the detached signature and the published Ed25519 public key. The key is self-generated; no independent witness.

MANIFEST.json .sig key

4 · Audit and acceptance

44 artifact-derived gates: 42 PASS / 2 BLOCKED. Conditional release, blockers and the full final audit.

Final audit Decision

5 · Integrity list

SHA-256 for the ZIP, verifier, manifest, signature and key. Compare before you unpack.

SHA256SUMS.txt

6 · Historical evidence

E81 governed workforce bundle, preserved unmodified, including its known forked gate log reported honestly.

E81 bundle

What “independent” means here: independent of CAIN code, not independent of the project. The verifier is a third-party-capable kit. A third-party audit and an independent witness are still NOT PERFORMED.

The technical case: why vertical AGI needs a governor, not another model

1. Vertical AGI is an actor, not a chat model.

A vertical AGI is a domain-specialised system that acts: it plans, calls tools, edits repositories, queries databases, moves money, changes infrastructure, and delegates to other agents. Its failure mode is not a bad paragraph. Its failure mode is an unauthorized effect — a destructive write, a credential exfiltration, a payment, a clinical or financial action that should never have crossed the boundary. Prompt filters inspect the conversation channel; the effect happens in the execution channel. That gap is the governance problem vertical AGI creates.

2. The governor is a runtime, not a guardrail.

CAIN-42 places a deterministic control plane between the proposal and the effect. The canonical path is WHO → AUTHORITY → INTENT → SECURITY CONTEXT → POLICY → RISK → TRUST → PREDICTION → DECISION → ENFORCEMENT → EXECUTION → EFFECT → EVIDENCE → CAUSALITY → COUNTERFACTUAL → OUTCOME → TRUST UPDATE. The model is interchangeable: local Qwen and SmolLM weights, a hosted frontier model, or a customer's own model. The agent framework is interchangeable: MCP, A2A, LangGraph, AutoGen, CrewAI or a bespoke runtime. The governor does not care which model proposed the action; it cares whether the exact action is authorized, whether enforcement actually halted it, and whether the observed effect matches the intent.

3. Authority is a lease, not a permission.

An agent never receives standing authority. It receives a signed, scope-bound, short-lived, single-use capability bound to the exact tool, arguments, tenant, mission and expiry. Delegation is attenuating: a child can never hold more authority than its parent. Every claim is re-checked at the boundary — signature, agent, tool, scope hash, expiry, fencing token, revocation, replay, budget. A stale worker cannot commit after losing its lease. A replayed grant cannot produce a second effect. An unknown or errored state fails closed. This is the difference between “the model promised not to” and “the system cannot”.

4. Evidence is the product.

Every governed action produces a receipt: the canonical action, the signed grant, the policy and risk decision, the enforcement event, the execution, the observed effect, and the hashes that bind them. Receipts are SHA-256 chained and rolled into an RFC 6962 Merkle root. The public bundle ships a standalone verifier that imports none of CAIN-42's code — only Python's standard library and cryptography. It recomputes the headline numbers from the raw rows instead of trusting a report. That is what makes the claim repeatable: an engineer can download the bundle, run the verifier, and get the same 55 checks on their own machine.

5. One spine, many verticals.

Vertical AGI does not need one model per domain; it needs one authority and evidence spine that every domain-specific agent plugs into. CAIN-42 supplies that spine: identity federation, delegated authority, capability tokens, deterministic policy, predictive blast radius, the enforcement boundary, the evidence fabric, the adversarial engine and the autonomous control loop. The verticals differ in their detectors, policies, tools and risk models — software engineering, cybersecurity, finance, medicine, legal information, infrastructure, research. They do not differ in the question that matters: was this exact action authorized, enforced, executed and proved?

6. Measured, not asserted.

E82 measured 1,182 real local model calls and 0 remote calls. Ten genuinely simultaneous local model inferences ran; 25 process-local workers were blocked by memory rather than overclaimed; 50 and 100 concurrent workers ran against a shared serialised model and the page reports simultaneous inference = 1. A one-hour real wall-clock campaign completed 165 tasks with 0 failures and 0 duplicate effects. Cross-host failover completed with RPO 0.025 s, RTO 2.557 s, a stale writer refused and 0 duplicate effects. A real host reboot interrupted an active mission; 20/20 tasks recovered, the boot id changed, and durable identity survived. The adversarial lab blocked 86/86 scenarios. The collaboration and specialist-prompt experiments produced negative results and those negatives are preserved.

7. What remains unproven — stated on the page.

No third-party audit, SOC 2, ISO 42001 certification or independent witness exists. The signing key is self-generated. No remote model was called because no credential was injected. No production service was deployed. 25 process-local workers are memory-blocked; 6/24/72/168-hour endurance was not attempted. The PBFT cluster is vendor-operated and remains quarantined for customer anchoring after the 2026-11-15 safety incident. “Governor” is an architectural role, not a regulatory designation, and CAIN-42 does not claim to have achieved AGI. The evidence is reproducible; the ambition is not allowed to overwrite it.

Third-party status, in plain words. The verifier is a third-party verification kit: it imports no CAIN code and can be run by anyone. That is not the same as a third-party audit. No external auditor has certified CAIN-42, no independent witness has cosigned the Merkle root, and the signing key is self-generated. We publish the distinction because the difference is the point. See the full list of what is not built or not authorized →

Public evidence surface: E82 · 2026-11-17 · FIRME AI CORPORATION. All live panels call same-origin public endpoints; if an endpoint is down, the panel says NOT LIVE and invents nothing. The page does not execute your text. No production deployment, DNS, Stripe, firewall or credential change was made for this page.

Skip to content
CAIN-42 is a Byzantine trust fabric and the governed execution paradigm for autonomous AI. This is its evidence registry. The Agentic 75 is its operating model for agentic AI in 2027: 75 governed products, each with a runnable engine, a signed sample and an offline verifier. Verify the 75 → Full catalog →Proof Standard →Integrations →
EVERYTHING FOR SALE 205 sellable surfaces: 99 products, PROOFLOOP and 105 services. 126 active payment links. The counts below are backed by the signed audit bundle. Browse everything → Buy a product → PROOFLOOP → Audit evidence →
CAIN-42 network:
CAIN-42 Proto-VAGI Evidence Registryclawx.click · public record
connecting…
LONG-HORIZON GOVERNED OPERATION · MEASURED 2026-11-21T18:17:13Z

A real mission has been running for 12.00 hours — 2142 research cycles, 2136 useful, 214 signed checkpoints

persistent mission
12.00 h of 12.00 h target
cycles executed
2142
useful research units
2136
failed units
0
signed checkpoints
214
restarts survived
0
journal verified
YES
memory facts retained
327
state
COMPLETED

External long-horizon rank: UNRANKED. No external METR-style time-horizon evaluation has been run for CAIN, so a top-10 position cannot be claimed. The measured local run is reported instead.

Why this is evidence and not a badge: every cycle does a capability induction on a never-seen task family, records the raw outcome, and appends a SHA-256 hash-chained journal record; a corrupted chain or memory digest makes the mission refuse to resume (exercised). A running daemon is not a completed mission — completion requires 43,200 real seconds, and the panel reports the wall clock it actually has. AGI status: NOT ACHIEVED.

Open the live lab → · mission JSON · SWE-bench board · evidence bundle

CAIN-42

The CAIN-42 agentic evidence portal · live evidence for autonomous AI agents

Everything needed to audit CAIN-42 is on this one site, checked by your browser.

Auditors only need clawx.click. The live production clusters, real quorum certificates, live decisions, every signed claim with its real status, every evidence file, the signing keys, the public Sigstore anchors and the offline verifiers are all reachable from this page. Nothing requires an account, and nothing asks you to trust us.

CAIN-42 is a Byzantine trust fabric. This site is where its evidence is kept: the Ed25519-signed claims register, every evidence bundle, the hash-chained proofs written by the two live validator clusters, and the monitors that re-check them. Statuses are copied from the signed files exactly as published, whatever they say. When this page loads, it recomputes the hashes and verifies the signatures with your browser's own WebCrypto. You don't have to take our word for anything.

Why execution-layer governance

Evidence-root key (Ed25519)
loading…
Signed evidence index digest
loading…
Catalog generated
loading…

The architectural case

More capable agents. Explicit authority. Verifiable evidence.

Authorization claims should be checkable.

Long-running agents and coordinated swarms need more than instructions. They need clearly bounded authority, enforceable tool access, operational continuity, and records that can be checked. These 13 talks and demonstrations explain complementary parts of that challenge. Every one can be played on this page.

An agent's ability to propose an action is not the same as its authority to execute it. CAIN-42 is designed to place an enforceable authorization boundary between AI agents and governed tools or real-world effects, with evidence of authorization decisions.

The goal is not to make agents less useful. It is to make broader autonomy practical by keeping permissions, limits, revocation, and evidence outside the model's discretion. Prompts guide behavior. Execution controls determine which governed actions may actually run. Security-critical authorization should not depend solely on the model following its prompt.

The industry is converging on the execution layer

The same requirement is now being described in public by some of the largest names in the field, each from its own position. Speakers and publishers in the 13 sources below include:

  • Anthropic
  • AWS
  • Microsoft (speaker)
  • NVIDIA
  • Visa
  • Okta
  • Kong
  • The Linux Foundation
  • LF Decentralized Trust
  • Moonshot AI (Kimi)
  • EnforceAuth

Independent sources, not CAIN endorsements. These sources discuss related architectures and engineering principles. None of them mentions CAIN-42. CAIN-specific capabilities and performance require CAIN-specific evidence, which is linked on every card.

Where the boundary sits

  1. ProposeAn agent or swarm proposes an action.
  2. CheckThe authorization boundary checks policy and delegated authority.
  3. Execute if authorizedThe governed tool runs only on ALLOW.
  4. RecordDecision and execution evidence are recorded as supported.
  5. Monitor and revokeMonitoring and revocation stay available within the documented scope.

A recorded authorization decision shows what was permitted. It does not prove the action succeeded, and not every external effect can be reversed.

What this means

  • Prompts guide the agent.
  • Policies define permitted actions.
  • Enforcement controls governed execution.
  • Evidence supports later verification.
  • Revocation limits future authorized activity within the system's actual coverage.

This site holds the records: who proposed an action, which decision was made, and the signed, hash-chained entry that says so, with verifiers that run offline. Four things are kept distinct: an approval, an attempted execution, a completed execution, and a successful outcome. A valid signature can establish the origin and integrity of a record. It does not, by itself, establish that the record is complete, that every assertion is true, or that the underlying action succeeded. The validator clusters that co-sign records are operated by the vendor across several regions; they are not independently administered.

Watch the 13 sources

13 relevant talks and demonstrations.

  1. SOURCE 01 OF 13 · OPEN-SOURCE PROJECT TALK

    Policy-Checked, Attributable, Revocable, Provable AI Agent Authorization

    Speaker names not confirmed in this review · published by LF Decentralized Trust · 2026-10-02 (streamed live)

    • Authorization and identity
    • Evidence and trust

    How it relates to CAIN-42: It makes the same separation CAIN-42 is built on: what an agent is able to do is a different question from what it holds authority to do, and the second must be answered outside the model.

    What it explains, how CAIN applies it, and what it does not prove
    What the source explains
    Introduces Caracal, an open-source AI authorization project in the LF Decentralized Trust mentorship program. The publisher's description says it covers why API-key and identity models fall short for agents, what authority means for autonomous systems, and an architecture for policy-based authorization, attribution, revocation and auditability.
    How CAIN applies this principle
    MCPGate refuses a governed tools/call that does not carry a signed, short-lived capability token bound to that call's parameters and usable once. A changed argument or a replayed token is refused. Implemented; public evidence limited
    What this does not prove
    Caracal is a separate project. Its talk does not mention, validate or endorse CAIN-42.
    CAIN evidence and documentation
    Try a blocked tool call in the MCPGate playground · Where MCPGate sits in the request path
  2. SOURCE 02 OF 13 · INDEPENDENT TECHNICAL PRESENTATION

    Trust and Guardrails Governing Autonomous AI Agents on Azure

    Nicholas Chang · published by Nicholas Chang (individual channel) · 2026-10-24

    • Authorization and identity

    How it relates to CAIN-42: That dedicated layer between proposal and execution is the position CAIN-42 occupies.

    What it explains, how CAIN applies it, and what it does not prove
    What the source explains
    Per the publisher's description: authorization changes once software interprets intent and acts for a user. Agents should have their own identities and narrowly scoped permissions, and policy such as spending limits and approval thresholds should live in a dedicated layer between the agent's proposed action and execution, not inside a prompt.
    How CAIN applies this principle
    CAIN decisions are made per proposed action. Only ALLOW is executable; REQUIRE_APPROVAL pauses for a person; unknown or error states are denied. Budget and spending envelopes exist in the AOS mission runtime, which is a reference implementation. Implemented; public evidence limited
    What this does not prove
    This is an individual's presentation about building on Azure. It is not presented here as official Microsoft policy.
    CAIN evidence and documentation
    Try a blocked tool call in the MCPGate playground · Agentic 75 governance map
  3. SOURCE 03 OF 13 · VENDOR PRESENTATION

    How to Build Audit-Ready Governance for Autonomous AI Agents | Okta

    Okta speakers (The AI Identity Summit) · published by Okta · 2026-07-16

    • Authorization and identity
    • Evidence and trust

    How it relates to CAIN-42: Attributable actions and permission management across an agent's whole lifecycle, including the ability to stop it, are core CAIN-42 requirements.

    What it explains, how CAIN applies it, and what it does not prove
    What the source explains
    Per the publisher's description: a masterclass on keeping a traceable, verifiable audit trail as autonomous systems act for an enterprise, moving from static security snapshots to continuously enforced identity controls, with clear ownership and accountability.
    How CAIN applies this principle
    Every MCPGate decision is tied to an agent identity and written to a hash-chained, signed record. Revocation is checked mid-flight at the boundary, and a fleet quarantine (kill switch) stops further authorized actions by an agent or swarm. Implemented; public evidence limited
    What this does not prove
    "Audit-ready" is the vendor's phrase, not an independent compliance finding, for Okta or for CAIN-42. A kill switch stops future governed actions; it does not undo completed ones.
    CAIN evidence and documentation
    CAIN-42 evidence registry · Offline verifiers
  4. SOURCE 04 OF 13 · VENDOR PRESENTATION AND DEMONSTRATION

    Agentic Identity: Securing MCP Servers with Kong AI Gateway

    Andrew, product manager at Kong (as named by the publisher) · published by Kong · 2026-11-05

    • Authorization and identity

    How it relates to CAIN-42: A gateway placed between agents and MCP tools, acting as both network and identity perimeter, is the architectural position of MCPGate.

    What it explains, how CAIN applies it, and what it does not prove
    What the source explains
    Per the publisher's description: agents should stop carrying every credential themselves and route actions through a governed gateway. Covers monitoring MCP traffic, inbound identity, outbound credential translation, token vaulting and fine-grained authorization policy at the gateway.
    How CAIN applies this principle
    MCPGate is a JSON-RPC proxy (stdio and HTTP) between an agent and its MCP servers. A tools/call that is unknown, malformed or unauthorized gets a JSON-RPC error and is not forwarded to the server. Implemented; public evidence limited
    What this does not prove
    A gateway governs only the calls routed through it. Tools reached with other credentials or by another path are outside its control, for Kong's gateway and for MCPGate alike.
    Publisher's chapter timestamps
    5:59 Gateway identity perimeter · 8:17 MCP governance ladder · 23:25 Gateway authorization models
    CAIN evidence and documentation
    Where MCPGate sits in the request path · Try a blocked tool call in the MCPGate playground
  5. SOURCE 05 OF 13 · CONFERENCE TALK

    Trust Is the Next Bottleneck: Why the Agentic Economy Needs Confidential Computing

    Pawan Khandavilli, Microsoft · published by The Linux Foundation · 2026-07-17

    • Evidence and trust

    How it relates to CAIN-42: CAIN-42 shares the goal of substantiating execution claims with evidence that a relying party can check, instead of only recording an identity.

    What it explains, how CAIN applies it, and what it does not prove
    What the source explains
    Per the publisher's abstract: agents can authenticate but cannot yet prove what actually ran. OAuth handles authorization, not runtime integrity; prompt-level governance shapes intent, not enforceable policy. The talk applies hardware attestation and measured execution to agent trust and names four gaps, including measured policy-as-code and portable attestation evidence.
    How this informs CAIN's roadmap
    CAIN-42 today produces software-signed, hash-chained records with offline verifiers. Hardware-backed attestation is on the roadmap: an appraisal engine exists in code, and no trusted-execution hardware is deployed. Planned
    What this does not prove
    Signed application records are not hardware attestation. CAIN-42 does not claim hardware-attested execution.
    CAIN evidence and documentation
    Offline verifiers
  6. SOURCE 06 OF 13 · VENDOR PRODUCT PRESENTATION

    The Authorization Gap: Securing Autonomous AI at Runtime

    EnforceAuth · published by EnforceAuth · 2026-05-20

    • Authorization and identity

    How it relates to CAIN-42: Per-action authorization during ongoing agent activity, rather than a one-time login check, is how CAIN-42 treats authority.

    What it explains, how CAIN applies it, and what it does not prove
    What the source explains
    Per the publisher's description: the central question for AI security moves from "who are you?" to "what are you allowed to do right now?", answered by runtime authorization enforcement with policy-as-code and auditability for agents, machine identities and MCP tools.
    How CAIN applies this principle
    Each governed call is decided at the moment it is proposed. The proxy fails closed: a gate error, an unknown trust state or an unavailable replay store refuses the call. Implemented; public evidence limited
    What this does not prove
    Latency, scale and "immutable" audit figures in this video are the vendor's own claims about its own product. They are not repeated here and say nothing about CAIN-42.
    CAIN evidence and documentation
    Try a blocked tool call in the MCPGate playground
  7. SOURCE 07 OF 13 · CONFERENCE TALK

    Claude for Long-Horizon Tasks

    Lance Martin, Anthropic · published by AI Engineer · 2026-07-22 (premiered)

    • Long-running agents

    How it relates to CAIN-42: Separating the reasoning model from the environment where actions run is where an authorization boundary naturally sits. The talk covers the harness; CAIN-42 covers who may act through it.

    What it explains, how CAIN applies it, and what it does not prove
    What the source explains
    Per the publisher's description: lessons on building agent harnesses for reliable and secure long-horizon work, including decoupling the brain and hands, self-verification, self-learning, and designing harnesses that evolve.
    How CAIN applies this principle
    The CAIN AOS mission runtime journals each step, writes an execution intent before an effect, and requires independent verification before a mission is accepted. It is a reference implementation and is not deployed. Demonstration only
    What this does not prove
    Session persistence is not authorization enforcement, and this talk is about Anthropic's platform, not CAIN-42.
    CAIN evidence and documentation
    Agentic 75 governance map
  8. SOURCE 08 OF 13 · WORKSHOP

    Anthropic Workshop: Build Agents That Run for Hours

    Ash Prabaker and Andrew Wilson · published by AI Engineer · 2026-05-18 (premiered)

    • Long-running agents

    How it relates to CAIN-42: If an agent runs for hours, its permissions and stop controls have to stay meaningful for hours. The workshop's point that a producer should not grade itself matches CAIN-42's rule that a producer never verifies its own work.

    What it explains, how CAIN applies it, and what it does not prove
    What the source explains
    Per the publisher's description: why self-evaluation is a trap and adversarial evaluator agents work better, why structured handoffs beat context compaction for coherence drift, decomposing work into testable contracts, and reading traces as the main debugging loop.
    How CAIN applies this principle
    In the CAIN workforce fabric a result is accepted only after a different agent re-checks it. Published endurance evidence is scoped: one continuous one-hour mission (252 of 252 tasks verified) and 1,018 aggregate worker-hours across concurrent workers. Implemented and publicly evidenced
    What this does not prove
    Hours of demonstrated operation in a workshop do not establish multi-day endurance, and CAIN-42 has not demonstrated a multi-day mission.
    Publisher's chapter timestamps
    2:29 Challenges: context, planning and judgment · 17:28 State-of-the-art harness patterns · 25:04 The generator-evaluator contract
    CAIN evidence and documentation
    Evolution 81 workforce evidence bundle
  9. SOURCE 09 OF 13 · CONFERENCE TALK

    Your Old Prompts Are Making Claude Worse

    Lance Martin, Anthropic · published by AI Engineer · 2026-11-06

    • Long-running agents

    How it relates to CAIN-42: Reliable autonomy is a system-engineering problem as well as a prompt-writing one. Prompts still matter; they guide the agent.

    What it explains, how CAIN applies it, and what it does not prove
    What the source explains
    Per the publisher's description: auditing prompts for outdated habits, using the prompt cache, designing evals a model can climb, and how a managed-agent harness that separates brain from hands keeps days-long agents reliable and secure.
    How CAIN applies this principle
    CAIN-42 keeps security-critical authorization out of the prompt: a model, a memory or a plan may propose an action, and none of them can authorize one. Implemented; public evidence limited
    What this does not prove
    The talk does not say prompts are unnecessary, and neither does CAIN-42. Cost and benchmark figures in the talk are Anthropic's own.
    Publisher's chapter timestamps
    1:32 Audit your prompts · 18:53 Harnesses for days-long agents · 20:21 Managed agents: separating brain from hands
    CAIN evidence and documentation
    Agentic 75 governance map
  10. SOURCE 10 OF 13 · COURSE LESSON AND DEMONSTRATION

    Multi-Agent Patterns: Agent Swarms

    AWS Developers (Strands Agents course, chapter 12) · published by AWS Developers · 2026-07-14

    • Swarms

    How it relates to CAIN-42: When leadership moves between agents, authority has to stay explicit at every handoff. That is the delegation problem CAIN-42 addresses.

    What it explains, how CAIN applies it, and what it does not prove
    What the source explains
    Per the publisher's description: in a swarm, agents share a workspace, each with its own tools, prompt and specialization. They hand off to each other dynamically and no single agent is permanently in charge.
    How CAIN applies this principle
    In the CAIN workforce fabric a worker acts only under a signed grant attenuated from its mission, plus a single-use ticket for one task. A child cannot exceed its parent's envelope, and registering a worker grants nothing. Implemented and publicly evidenced
    What this does not prove
    Swarm coordination does not by itself demonstrate security, correctness or general intelligence.
    Publisher's chapter timestamps
    1:30 Configuration and safety · 2:48 Execution and shared state
    CAIN evidence and documentation
    Evolution 81 workforce evidence bundle
  11. SOURCE 11 OF 13 · PRODUCT DEMONSTRATION

    Meet Kimi K2.6 Agent Swarm

    Kimi AI · published by Kimi AI (Moonshot AI) · 2026-04-22

    • Swarms

    How it relates to CAIN-42: Hundreds of sub-agents taking thousands of steps make bounded authority and per-agent attribution a practical necessity.

    What it explains, how CAIN applies it, and what it does not prove
    What the source explains
    A product demonstration of the Kimi K2.6 Agent Swarm. The publisher advertises up to 300 sub-agents and 4,000 coordinated steps; those are the publisher's figures.
    How CAIN applies this principle
    CAIN-42 has run 1,000 real worker processes on one host, each with its own generated Ed25519 identity and every result independently re-checked. Those workers ran deterministic tools, not model reasoning. Implemented and publicly evidenced
    What this does not prove
    Advertised scale is not evidence of reliability, and CAIN-42's worker count is not a claim about model intelligence.
    CAIN evidence and documentation
    Evolution 81 workforce evidence bundle
  12. SOURCE 12 OF 13 · VENDOR PRESENTATION

    Visa Intelligent Commerce: Building Trust for the Future of Agentic Commerce

    Visa · published by Visa · 2026-05-27

    • Commerce and real-world deployment

    How it relates to CAIN-42: An agent being able to start a payment is different from holding authority to make it. That is the capability-versus-authority distinction applied to money.

    What it explains, how CAIN applies it, and what it does not prove
    What the source explains
    Per the publisher's description: as agents search, compare and transact for consumers, trust has to come first. Consumers need confidence the agent stays within their instructions; merchants and issuers need visibility into intent, identity and authorization; consumer control is preserved through clear authorization and secure credential handling.
    How CAIN applies this principle
    CAIN-42 ships wallet-mandate and checkout-mandate review engines that check a supplied mandate record against published rules and sign the result. Implemented; public evidence limited
    What this does not prove
    CAIN-42 has no integration, approval or partnership with Visa, and its mandate engines review records; they do not process payments.
    CAIN evidence and documentation
    Agentic 75 product engines
  13. SOURCE 13 OF 13 · KEYNOTE HIGHLIGHTS

    Jensen Huang's Keynote Highlights at NVIDIA GTC Berlin 2026

    Jensen Huang, NVIDIA · published by NVIDIA · 2026-10-25

    • Commerce and real-world deployment
    • Swarms

    How it relates to CAIN-42: Execution governance is one component of a much larger agent stack that runs from chips to robots.

    What it explains, how CAIN applies it, and what it does not prove
    What the source explains
    Per the publisher's description: how AI factories are evolving to run agents and why physical AI is the next frontier, with announcements across AI infrastructure, safe agents, open models, enterprise AI, healthcare, mobility and robotics.
    How this informs CAIN's roadmap
    CAIN-42's physical-systems work is simulation only. No hardware is governed today, and physical deployment is on the roadmap. Planned
    What this does not prove
    A keynote vision does not establish CAIN-42's suitability for safety-critical physical deployment.
    CAIN evidence and documentation
    Agentic 75 governance map

What these sources do, and do not, establish

They do establish

  • That identity providers, gateway vendors, model developers, cloud platforms, payment networks and open-source foundations are publicly treating agent authorization, attribution, revocation and evidence as engineering requirements.
  • That separating the reasoning model from the place where actions execute is a mainstream design for long-running agents.
  • That swarms of specialized agents are a shipping pattern, which makes delegated authority a live problem.

They do not establish

  • That any of these organizations has reviewed, endorsed, partnered with or integrated CAIN-42.
  • That CAIN-42 was developed in response to these talks, or that CAIN-42 meets any vendor's stated performance.
  • That any system here has passed an independent audit or is safe for every deployment. Vendor presentations are not peer-reviewed research.

2027–2028 outlook: thousands of governed agents, with a kill switch

The goal FIRME AI CORPORATION is building CAIN-42 toward: thousands of specialized AI agents working for thousands of hours across many different fields, every action authorized before it runs, every result backed by evidence, and one kill switch that sits outside the model. Governed swarms of specialists, one vertical at a time, are our path to that.

This is a goal and an outlook, not a forecast presented as fact. If agents become more persistent, more interconnected and more economically consequential, their authority will need to remain explicit and enforceable throughout their operation. CAIN-42 is being developed around that requirement.

Where that path stands today, measured

  • 1,000 real worker processes on one host, each with its own signed identity, every result independently re-checked.
  • 1,018 aggregate worker-hours across concurrent workers. That is a sum over workers, not continuous elapsed runtime; the longest single continuous mission is one hour.
  • 12 fields in one mission, with every result re-checked by an agent from a different field. Those tasks were deterministic text operations, not model reasoning.
  • Fleet quarantine that stops further governed actions by one agent or a whole swarm. It stops what comes next; it cannot reverse what is already done.
  • Next: the same controls held across more hosts, multi-day missions and model-backed specialists.

CAIN-42 does not claim general intelligence, and nothing here predicts when or whether it arrives. The claim is narrower: whatever agents become capable of, their authority should be granted, bounded and revocable from outside the model.

CAIN evidence and documentation

Source review date: 2026-11-15. For each of the 13 sources the link, title, channel, publication date and publisher's description were checked on that date, and timestamps are the publisher's own chapter markers. Summaries restate the publisher's description; they are not transcripts or quotations. CAIN-42 evidence is produced and hosted by FIRME AI CORPORATION and has not been audited by a third party.

Loading totals…
E82 · 2026-11-17 · GOVERNED VERTICAL AGI The evidence is live at the top of this page. A live decision console, a zero-CAIN offline verifier, the signed 240-file bundle and the honest limits. No E81 workforce banner is used here.
Verify the evidence ↑

THE 2027 STANDARD · CAIN-42 AGENTIC 75

The 2027 standard for agentic autonomy

The integration of the CAIN-42 Agentic 75 suite is a foundational shift for autonomous AI: away from probabilistic prompt-layer suggestions and toward deterministic, zero-trust execution governance. As multi-agent systems and autonomous swarms scale into production through 2027, prompt-based alignment and conversational guardrails have repeatedly proven insufficient. The Agentic 75 architecture shifts control from the model’s intent to the system’s physical and computational execution boundaries.

1. Execution-Layer Governance Over Probabilistic Prompt Filters

Traditional AI safety operates almost entirely in the conversation channel, parsing natural-language instructions and trying to filter malicious tokens before the model processes them. Autonomous agents do not merely converse: they invoke tools, write files, run shell commands, modify databases and transfer capital through protocols such as the Model Context Protocol (MCP).

  • The paradigm shift: CAIN-42 intercepts governed requests at the execution boundary before the side effect occurs. Calls that bypass the boundary are not governed by it; that limit is stated openly.
  • Mechanism: every governed tool call routed through the boundary is evaluated against authorization boundaries, structural schemas and policy rules. If an action lacks verified permission it is blocked before execution. Unknown or errored states default to denial rather than model judgment.

2. Proof-Carrying Software and Tamper-Evident WORM Auditing

Verifying that an autonomous agent behaved correctly during a multi-hour or multi-day coding session should not require trusting verbose logs or post-hoc summaries.

  • The paradigm shift: the Agentic 75 introduces proof-carrying development and operation. CAIN LongProof, CAIN LoopTrace and the Tamper-Evident WORM Evidence Notary generate tamper-evident, hash-chained records with Ed25519 signatures and SHA-256/RFC 6962 Merkle roots.
  • Mechanism: governed actions, decisions and state changes produce cryptographically verifiable artifacts that anyone can re-run and verify offline without importing CAIN code. The record proves integrity and binding to the supplied evidence; it does not prove the evidence is complete or true. The design is intended to support EU AI Act record-keeping under Articles 12, 19, 26(6) and 73(6). No notified body has reviewed CAIN and no certification is claimed.

3. Byzantine-Resilient Quorum Co-Signing

Single-server logging architectures remain vulnerable to administrative tampering, compromised internal endpoints and silent state overwrites.

  • The paradigm shift: CAIN-42 can have governed records ordered and co-signed by a multi-region Practical Byzantine Fault Tolerance (PBFT) cluster: N=4, f=1, Q=3, across four Vultr regions under one account. That is geographic separation, not provider or administrative independence.
  • Mechanism: hosted decisions are evaluated and co-signed by a 3-of-4 quorum. The consensus layer orders and co-signs; it cannot override a denial or manufacture permission. A compromised replica cannot silently change a quorum record — this is tamper-evident assurance, not an absolute guarantee against every hostile deployment. The quorum is operated by FIRME AI CORPORATION, not an independent witness network.

4. Granular Runtime Cages and Rapid Swarm Containment

Autonomous agent swarms running in parallel create catastrophic blast radiuses if a loop goes rogue, hallucinates destructive arguments or is steered by prompt injection.

  • The paradigm shift: the Agentic 75 provides fine-grained runtime isolation combined with fleet-wide emergency circuit breakers.
  • Mechanism: tools are wrapped in execution boundaries such as CAIN RuntimeCage and the MCPGate Boundary Appliance. Cross-process SQLite WAL synchronization shares revocation state, and the local containment path is measured and reported with its tests. We do not claim an unmeasured global sub-50ms guarantee or instantaneous cross-region revocation; distributed propagation is a separate, measured question.

5. Deterministic Defect Detection and Automated Verification

AI-written code has a familiar failure mode: uncomputed results, hard-coded success claims and security checks written to return unconditional approval.

  • The paradigm shift: deterministic static analysis (FALSEWORK) alongside KEYSTONE and PROOFLOOP requires evidence, not confidence, before a fix is accepted.
  • Mechanism: a patch proposed by an AI coding agent is validated only when an independent test suite fails on the vulnerable code, passes on the patched code inside a locked sandbox, and clears structural safety re-scans. This produces verifiable evidence about that test and that patch; it is not a proof of general correctness, and FALSEWORK is heuristic — determinism is not correctness.

NO VERIFIED IDENTITY → NO VALID AUTHORITY → NO GOVERNED EXECUTION.

AGENT PROPOSES→CAIN DECIDES→MCPGATE ENFORCES→SYSTEM EXECUTES→EVIDENCE REMEMBERS

Summary: The 2027 Standard for Agentic Autonomy

The CAIN-42 Agentic 75 architecture redefines AI agents by proving that capability does not equal authority. By combining zero-trust execution gating, Byzantine quorum co-signing, offline-verifiable evidence and independent test validation, it turns autonomous agents from unpredictable experiments into accountable, auditable systems.

It does not turn every engine into a certified production control. The claim is narrower and testable: show the record, and let anyone re-check it.

What is verified today

  • 75 of 75 Agentic products are DELIVERABLE, each with published rules and a live try-it.
  • All 75 signed samples re-verify offline with their own pinned key.
  • 2,598 Agentic 75 tests pass; the catalog selftest is 75/75.
  • The public evidence bundle is on clawx.click, with a verifier that imports no CAIN code.

What is not claimed

  • Not a production assessment for every engine: several are static, heuristic or simulation-only, and say so.
  • Not universal enforcement: only calls routed through the declared boundary are governed.
  • Not a global real-time revocation guarantee, hardware-attested identity, or independent certification.
  • No SOC 2, ISO 42001, EU AI Act certification or third-party audit.
  • The validator quorum is vendor-operated, not an independent witness network.

The model is a commodity. The closing loop is the moat: detect → act → prove → sign → verify offline. Built by FIRME AI CORPORATION — Vincent Noriega (founder) and his two co-founders.

AGENTIC 75 · PRODUCT EVIDENCE

75 productised offers. 75 engine-backed and verified today. Every status published.

LIVE IN PRODUCTION: 75 OF 75 · NEWEST: #75 OF THE TOP 75

CAIN-A75-075 · Delegation Measurement Review · Observational evidence of larger task delegation · Signed evidence

Also live: CAIN-A75-071 · CAIN-A75-074 · CAIN-A75-073

All 75 live Agentic 75 products and their evidence (newest first)
#SKUProduct · the 2026 developmentEngineEvidence · verify it
75CAIN-A75-075Delegation Measurement ReviewTop-75 #75: Observational evidence of larger task delegationcain_a75_engines/delegation/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
74CAIN-A75-071Proactive Agent Lease ReviewTop-75 #71: Proactive personal assistantscain_a75_engines/proactive/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
73CAIN-A75-074Mobile Expansion ReviewTop-75 #74: Broader international mobile-agent deploymentcain_a75_engines/mobileexpansion/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
72CAIN-A75-073Industrial Knowledge ReviewTop-75 #73: Industrial expertise captured as agentscain_a75_engines/procedure/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
71CAIN-A75-070Brand Agent ReviewTop-75 #70: Retailer-controlled agents in searchcain_a75_engines/brandagent/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
70CAIN-A75-069Interop Authority ReviewTop-75 #69: Agentic advertising interoperability standardscain_a75_engines/interop/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
69CAIN-A75-068Third-Party Agent ReviewTop-75 #68: Bring-your-own-agent advertisingcain_a75_engines/byoagent/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
68CAIN-A75-067Campaign Authority ReviewTop-75 #67: Agent-to-agent advertising executioncain_a75_engines/campaign/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
67CAIN-A75-066Browser Errand ReviewTop-75 #66: Mobile browser agents for real-world errandscain_a75_engines/errand/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
66CAIN-A75-065Mobile Data Boundary ReviewTop-75 #65: Work–personal data separation for mobile agentscain_a75_engines/mobileboundary/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
65CAIN-A75-064Agent Inventory ReviewTop-75 #64: Enterprise agent inventory and cost accountabilitycain_a75_engines/agentinventory/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
64CAIN-A75-063Chat Approval ReviewTop-75 #63: Agent permission review inside collaboration toolscain_a75_engines/chatapproval/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
63CAIN-A75-062Local Agent Containment ReviewTop-75 #62: Local AI coding on personal computerscain_a75_engines/localagent/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
62CAIN-A75-061Domain Model Authority ReviewTop-75 #61: CRM-specialised reasoning modelscain_a75_engines/domainmodel/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
61CAIN-A75-059Browser Agent Control ReviewTop-75 #59: More efficient browser representations and action loopscain_a75_engines/browserbatch/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
60CAIN-A75-058Framework Evaluation ReviewTop-75 #58: Repeated-run comparisons of agent frameworkscain_a75_engines/frameworkeval/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
59CAIN-A75-056Benchmark Evidence ReviewTop-75 #56: Visually complex knowledge-work evaluationcain_a75_engines/benchevidence/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
58CAIN-A75-055Memory Evaluation ReviewTop-75 #55: Multi-session memory evaluationcain_a75_engines/memeval/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
57CAIN-A75-054Deployment Trace EvaluationTop-75 #54: Deployment-driven agent benchmarkscain_a75_engines/tracebench/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
56CAIN-A75-053Consumer Agent Authority ReviewTop-75 #53: Consumer agents connected to real servicescain_a75_engines/consumeragent/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
55CAIN-A75-052Agent-to-Agent Deal ReviewTop-75 #52: Controlled agent-to-agent commerce experimentscain_a75_engines/a2adeal/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
54CAIN-A75-051Agent Identity ReviewTop-75 #51: Know Your Agent identity intelligencecain_a75_engines/agentidentity/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
53CAIN-A75-050Agent Interaction Risk ReviewTop-75 #50: Banking detection of agent activitycain_a75_engines/agentrisk/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
52CAIN-A75-047Commerce Protocol Authority ReviewTop-75 #47: Universal commerce protocolcain_a75_engines/commerceauthority/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
51CAIN-A75-046Ontology & Policy ReviewTop-75 #46: Business ontologies for regulated-industry agentscain_a75_engines/ontologypolicy/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
50CAIN-A75-045Approval-Gated Remediation ReviewTop-75 #45: IT investigation with approval-gated remediationcain_a75_engines/remediationgate/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
49CAIN-A75-044Contract Evidence ReviewTop-75 #44: Evidence-based construction change orderscain_a75_engines/contractevidence/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
48CAIN-A75-043Physical Schedule Assurance ReviewTop-75 #43: Agentic construction schedulingcain_a75_engines/constructionschedule/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
47CAIN-A75-042Background Agent Lease ReviewTop-75 #42: Continuous creative-production agentscain_a75_engines/backgroundlease/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
46CAIN-A75-041Creative Workflow ProvenanceTop-75 #41: Multistep creative agentscain_a75_engines/creativeprov/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
45CAIN-A75-039Fast-Path Authority ReviewTop-75 #39: Faster models for collaborative subagentscain_a75_engines/fastpath/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
44CAIN-A75-038Parallel Agent Run ReviewTop-75 #38: Multi-agent desktop supervisioncain_a75_engines/parallelruns/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
43CAIN-A75-037Enterprise Execution GovernanceTop-75 #37: Enterprise agent execution with governancecain_a75_engines/executiongov/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
42CAIN-A75-036Threat Investigation Evidence ReviewTop-75 #36: Evidence-grounded threat investigationcain_a75_engines/threatevidence/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
41CAIN-A75-035Real-Time Agent Gate ReviewTop-75 #35: Voice agents that reason while listeningcain_a75_engines/realtimegate/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
40CAIN-A75-034Regulated Workflow Evidence ReviewTop-75 #34: Clinical-trial oversight agentscain_a75_engines/regulatedevidence/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
39CAIN-A75-032Research Retrieval ProvenanceTop-75 #32: Agent-assisted search through biological archivescain_a75_engines/retrievalprov/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
38CAIN-A75-031Context Graph ReviewTop-75 #31: Shared organisational context graphscain_a75_engines/contextgraph/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
37CAIN-A75-029Personal Context ReviewTop-75 #29: Personal context connected to mobile actionscain_a75_engines/personalcontext/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
36CAIN-A75-028Mobile Agent Action ReviewTop-75 #28: Cross-application mobile task executioncain_a75_engines/mobilegate/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
35CAIN-A75-027Agent Lease ReviewTop-75 #27: Always-on workplace teammatescain_a75_engines/agentlease/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
34CAIN-A75-025Fleet Governance Design ReviewTop-75 #25: Reported orchestration of 10,000 concurrent agentscain_a75_engines/fleetdesign/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
33CAIN-A75-024Coding-Agent Action GateTop-75 #24: Coding agents inside native development toolscain_a75_engines/codinggate/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
32CAIN-A75-020OS Containment ReviewTop-75 #20: Operating-system-enforced agent containmentcain_a75_engines/oscontainment/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
31CAIN-A75-019Runtime Containment AuditTop-75 #19: Runtime agent quarantinecain_a75_engines/runtimecontainment/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
30CAIN-A75-018Enterprise Agent Governance ReviewTop-75 #18: Full-stack agent safety infrastructurecain_a75_engines/enterprisegov/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
29CAIN-A75-016Memory Consolidation ReviewTop-75 #16: Offline memory consolidation (dreaming)cain_a75_engines/memoryconsolidation/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
28CAIN-A75-015Context Integrity ReviewTop-75 #15: Retained reasoning and context compactioncain_a75_engines/contextcompaction/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
27CAIN-A75-013Multi-Model Orchestration ReviewTop-75 #13: Multi-model workflow orchestrationcain_a75_engines/modelorchestration/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
26CAIN-A75-012Research Evidence ReviewTop-75 #12: Experimentally validated biological binder discoverycain_a75_engines/binderevidence/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
25CAIN-A75-011Formal Artifact ReviewTop-75 #11: Self-supervised theorem discoverycain_a75_engines/formalartifact/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
24CAIN-A75-010Failure-Driven Improvement AuditTop-75 #10: Failure-driven agent improvementcain_a75_engines/failurefix/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
23CAIN-A75-009Autonomous R&D Change ReviewTop-75 #9: End-to-end autonomous R&D loopscain_a75_engines/rndchange/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
22CAIN-A75-008Robotics Release Gate ReviewTop-75 #8: Robot programming through agent teamscain_a75_engines/roboticsgate/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
21CAIN-A75-007Physical Action Twin ReviewTop-75 #7: Digital twins for safe physical agent actionscain_a75_engines/physicaltwin/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
20CAIN-A75-005Research Provenance ReviewTop-75 #5: Open-world multi-agent mathematical discoverycain_a75_engines/researchgraph/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
19CAIN-A75-004Proof-Carrying Change ReviewTop-75 #4: Machine-checkable security proofscain_a75_engines/proofreview/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
18CAIN-A75-003Discovery Loop Evidence ReviewTop-75 #3: Closed-loop experimental discoverycain_a75_engines/discoveryloop/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
17CAIN-A75-002Physical Autonomy Assurance ReviewTop-75 #2: Agent-operated molecular-biology laboratoriescain_a75_engines/labpreflight/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
16CAIN-A75-072Saved Workflow ReviewTop-75 #72: Reusable user-defined workflow skillscain_a75_engines/savedworkflow/engine.pyselftest manifestpython3 verify_agentic75.py
15CAIN-A75-060Research Toolchain ReviewTop-75 #60: Composable web-research toolchainscain_a75_engines/researchtool/engine.pyselftest manifestpython3 verify_agentic75.py
14CAIN-A75-057Commerce Refusal ReviewTop-75 #57: Shopping benchmarks that reward refusing bad purchasescain_a75_engines/commercerefusal/engine.pyselftest manifestpython3 verify_agentic75.py
13CAIN-A75-049Web Tool Definition ReviewTop-75 #49: Structured website actions through WebMCPcain_a75_engines/webtool/engine.pyselftest manifestpython3 verify_agentic75.py
12CAIN-A75-048Agent Checkout Mandate ReviewTop-75 #48: Browser payments without exposing card detailscain_a75_engines/checkoutmandate/engine.pyselftest manifestpython3 verify_agentic75.py
11CAIN-A75-040Creative Tool Boundary ReviewTop-75 #40: Native creative-project editing through MCPcain_a75_engines/creativetool/engine.pyselftest manifestpython3 verify_agentic75.py
10CAIN-A75-033Clinical Tool Boundary ReviewTop-75 #33: Pathology agents as interoperable MCP toolscain_a75_engines/clinictool/engine.pyselftest manifestpython3 verify_agentic75.py
9CAIN-A75-030Business-System Connector ReviewTop-75 #30: Agent communication across business systemscain_a75_engines/bizconnector/engine.pyselftest manifestpython3 verify_agentic75.py
8CAIN-A75-026Signed Deliverable PackTop-75 #26: Finished knowledge-work deliverablescain_a75_engines/wormevidence/engine.pyselftest manifestpython3 verify_agentic75.py
7CAIN-A75-023Vulnerability-to-Patch ProofTop-75 #23: Autonomous vulnerability investigation and repaircain_a75_engines/vulnpatch/engine.pyselftest manifestpython3 verify_agentic75.py
6CAIN-A75-022Agent Wallet Mandate ReviewTop-75 #22: Bounded agent identities and walletscain_a75_engines/walletmandate/engine.pyselftest manifestpython3 verify_agentic75.py
5CAIN-A75-021Agent Credential ReviewTop-75 #21: Task-scoped, short-lived agent credentialscain_a75_engines/credentialreview/engine.pyselftest manifestpython3 verify_agentic75.py
4CAIN-A75-017Portable Skill ReviewTop-75 #17: Agent skills as a portable ecosystem standardcain_a75_engines/skillreview/engine.pyselftest manifestpython3 verify_agentic75.py
3CAIN-A75-014Memory Authority AuditTop-75 #14: Persistent, correction-aware agent memorycain_a75_engines/memoryauthority/engine.pyselftest manifestpython3 verify_agentic75.py
2CAIN-A75-006Swarm Delegation AuditTop-75 #6: Single-prompt coordination of specialised agent teamscain_a75_engines/swarmdelegation/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"
1CAIN-A75-001Long-Run Engineering AttestationTop-75 #1: Full-day autonomous software engineeringcain_a75_engines/longrun/engine.pysigned sample + verify_offline.pypython3 verify_offline.py run.json --pinned-key "$(cat public_key.txt)"

Bundle: /evidence/agentic-75-products-2026-11-12/ · SELFTEST.json · verify_agentic75.py (no CAIN import).

The Agentic 75 turns the 75 most significant 2026 agentic-AI developments into fixed-scope CAIN-42 products and services. 21 of them run a real engine in the CAIN-42 repository and produced a manifest that re-verified. The other 0 are labelled honestly: 0 pilot, 0 advisory, 0 design. Nothing here calls a pilot a finished product.

75productised SKUs
75/75manifests verified; breadth is not equal depth
38pilot offers
21advisory or design offers
What this proves

That all 75 engine-backed SKUs ran on fixed samples and their SHA-256 manifests match the delivered files, and that every one of the 75 offers has a published status, scope and limit. Re-run verify_agentic75.py yourself; it imports no CAIN code.

What it does not prove

That the 0 non-engine offers are finished engines, that a customer ran any of this, or that the manifests are signed. They are integrity records, not attestations. No vendor partnership or third-party audit is claimed.

Open the evidence bundle · verify_agentic75.py · CATALOG.json · SELFTEST.json · try-it RESULTS.json (live route + browser run on all four surfaces)

Use the catalog on cainstudio.online/agentic-products, the developer home on mcpgate.online/agentic-75, and the CAIN-42 Hugging Face Space. The governance map for all 75 developments is cainstudio.online/agentic-75.

Auditor console: CAIN-42 in production, live

Everything in this console comes from the production system, live. Your browser reads the two validator clusters directly through this site, downloads real quorum certificates and checks every validator signature itself, triggers real decisions, and tries to forge certificates to show that the checks catch it. Every result is added to an audit log you can download. Labels say which results your browser verified and which a CAIN server reported.

connecting… Refresh every session: 0 checks
What this proves

That the certificates you verify were signed by at least 3 of the 4 member keys pinned in the signed proof index, that the gateway made the decision you sent, and that the seven forgeries are rejected.

What it does not prove

Independence: all validator nodes are run by FIRME AI CORPORATION (vendor-operated). The PBFT quorum tolerates one faulty node, but it cannot vouch for an operator who controls every node. Cluster health is reported by the gateway.

cain-mr-02 · retired 2026-11-15 after a safety finding · history kept, not a live anchor

loading…

cain-mr-01 · vendor-operated

loading…

Verify any quorum certificate yourself

Fetches the COMMIT and PREPARE certificates from the cluster and checks every vote's Ed25519 signature against the member keys pinned in the signed proof index, plus the quorum size, the certificate hash and the decision hash.

Try to forge a certificate (tamper lab)

Takes the last certificate your browser verified and makes seven forgeries: a forged signature, too few signers, COMMIT votes relabelled as PREPARE, a different decision, a cross-cluster replay, a cross-view replay, and a vote edited after signing. Every one must be rejected. If any is accepted, the checks are worthless, and this panel says so.

Make a real decision in production

A real call to the hosted trust fabric (POST /fabric/try) on a throwaway tenant with enforcement on. It returns the verdict, every pipeline stage and the stored record's digest. The record's integrity is reported by the gateway, not verified by your browser.

Published keys vs pinned keys

loading…

Every day since the first proof (24/7 record)

One cell per UTC day for each cluster, built from every proof in the signed proof index. The color is the share of proofs that day whose verdict was OPERATIONAL. A gray cell means no proof was written that day. Gaps and failures are shown, not hidden.

Live evidence timeline

  1. waiting for the first refresh…

Anchored outside CAIN: Sigstore Rekor

loading…

Verify your own evidence (customers)

Your decision records are private to your tenant, so this panel opens only with your CAIN-42 API key. A free key takes a minute, and no card is needed. The key stays in this browser tab's memory. It is sent only to this site's gateway, in the X-API-Key header, and it is forgotten when you lock the panel or close the tab.

Want this evidence for your own agents? CAIN-42 records, signs and anchors every decision your agents make, and this page lets you verify it any time. Start free · Plans · Quickstart

Get a free key
Locked.

Verification run (in this browser, now)

Each line below is computed on this page from the published files. Nothing is cached server-side and nothing comes from a stored result. Reload the page to run it again.

#CheckResultDetail
Running…

Hash-chained cluster proofs

Each validator cluster writes a proof about every 30 minutes. Every proof carries the SHA-256 of the one before it and is signed by the evidence-root key. Below are the 24 most recent proofs of each cluster. Your browser re-hashes each one, checks its signature and checks the link to the proof before it. Select a block to open the raw file.

What this proves

That each recent proof is intact, signed by the evidence-root key and linked to the one before it.

What it does not prove

That the proofs' measurements are correct. The evidence-root key belongs to the vendor, so independence comes from the outside anchors (Rekor, timestamp authorities) and from witnesses.

cain-mr-01

loading…

cain-mr-02 · retired 2026-11-15 · proof history kept

loading…

etcd coordination cluster: live verification

A three-member etcd cluster (Atlanta, Los Angeles, Miami). Every minute, our server makes the four read-only requests that the published verifier makes, through the same public endpoint any auditor uses, and publishes the raw responses at etcd-live.json. Your browser then re-runs the verifier's five checks against the signed 2026-10-15 capture. The responses are fetched by our server, so for a check that doesn't depend on us, run the command below yourself.

Checks re-run in this browser

loading…

Members (live)

loading…

Last 2 hours (one block per minute)

Run it yourself:

Transparency log, outside timestamps and witnesses

The proof chains are also published as an append-only Merkle-tree transparency log in the same form as Certificate Transparency and the Go checksum database: RFC 6962 trees with C2SP signed-note checkpoints. Every new checkpoint is also timestamped by two outside RFC 3161 timestamp authorities, FreeTSA and DigiCert, which receive only its SHA-256. Your browser checks the log signature, recomputes the Merkle root from the signed proofs, checks that every earlier checkpoint is still a prefix of the tree, and finds the checkpoint hash inside each timestamp token.

What this proves

That the published proofs form one append-only sequence: no proof was removed, changed or reordered after a checkpoint was issued. It also shows that each checkpoint existed by the time two independent timestamp authorities stamped it.

What it does not prove

That what the proofs say is true. A log proves that entries were registered, not that their content is correct. Until an independent witness cosigns, a log operator who showed different visitors different histories would be caught only by visitors who compare checkpoints. The log key is held by FIRME AI CORPORATION.

cain-mr-02 log · retired 2026-11-15, history kept

loading…

cain-mr-01 log

loading…

Watch this log yourself, 24/7 (independent monitor)

Don't trust this page: run the monitor on your own machine or in your own CI. It imports no CAIN code, keeps its own copy of every checkpoint it sees, and on every run re-checks each checkpoint's signature against the key you pinned, recomputes the Merkle root, checks that every earlier tree is still a prefix, and checks that no earlier proof changed. It prints ALARM and exits 2 on a fork, a rewrite or a bad signature. Two monitors that compare their saved checkpoints can also catch a log that shows different visitors different histories.

pip install cryptography
curl -fsSLO https://clawx.click/evidence/tlog/verify_tlog_monitor.py.txt && mv verify_tlog_monitor.py.txt verify_tlog_monitor.py
python3 -I verify_tlog_monitor.py --state ~/.clawx-monitor --vkey "$(curl -fsS https://clawx.click/evidence/tlog/vkey.txt)"
# then hourly (cron):  17 * * * *  python3 -I verify_tlog_monitor.py --state ~/.clawx-monitor

verify_tlog_monitor.py · GitHub Actions workflow (runs it hourly in your own repository) · log verifier key:

Want to cosign as a witness? Witness operators can contact firmeai2026@gmail.com. Cosignatures will be shown here and counted in each checkpoint.

Incident history and availability

What this proves

When each vendor-operated validator cluster was not operational, according to its own signed proofs, and its availability over 7, 30 and 90 days.

What it does not prove

Anything between two proofs (about 30 minutes apart), or the availability of the customer-facing gateway, which is tracked under Monitors.

loading…

Signed claims register

Every public claim CAIN-42 makes, with its status, limits, pinned artifacts and the command that re-checks it. Source: cain42-evidence-index.json (Ed25519-signed) and CAIN42_FINAL_PUBLIC_CLAIMS.json. Re-hash downloads each pinned artifact and compares its SHA-256 with the signed value. The status vocabulary is VERIFIED, TESTED, BENCHMARKED, SIMULATED, UNVERIFIED, FAILED and NOT_IMPLEMENTED. A feature that is not implemented is listed as not implemented, not left out.

ClaimStatusLevelStatement · limitsArtifactsRe-check
Loading…

Production gates

The gate table from the signed evidence index. Each row gives the requirement, its status, the evidence and its limits. Gates A to O are the 15 production gates, and the totals above show how many of 15 pass. Gates P to X are the further gates. Open gates are listed with their real status.

GateRequirementStatusEvidence and limits
Loading…

All evidence bundles

Every directory and document served under /evidence/, with nothing left out. Source: evidence-catalog.json, an unsigned listing rebuilt from the served files. Files lists every file with its SHA-256. Re-hash downloads the files and compares them: files pinned by a signed claim are compared with the signed hash, and all other files with the listing.

BundleDateFilesSizeSigned claimsOpen
Loading…

Monitors

Continuous verification (every 15 min)

loading…

Source: continuous-verification.json

Proof-fabric integrity monitor (signed)

loading…

Source: integrity-monitor/latest.json · history

FALSEWORK self-scan of the CAIN-42 codebase (every 42 s)

loading…

Source: keystone-codebase-scan-live.json

Deployed system state

loading…

Source: now.json

Open to every independent third-party auditor

What any auditor can do here, no account needed

How this registry stays unbiased

  • Every status on this page is copied from a signed or published file. None is typed into the page, so FAILED, SIMULATED, UNVERIFIED and NOT_IMPLEMENTED appear exactly as recorded.
  • Every check that runs here can come out FAIL. A check that cannot run shows NOT CHECKED, never PASS.
  • This page is served by us, so it does not replace your own check. Run the offline verifiers below on your own machine.
  • No independent third-party audit has been completed yet. Every result here was produced by CAIN-42 itself, and this registry exists so that an outside auditor can repeat each check.

Known limitations, as published

loading…

Verify offline, without this page

This page runs inside our own site, so the strongest check is one you run yourself. These verifiers import no CAIN code.

What it checksCommand
Evidence-root key identical on all 3 sites; evidence index and claims register re-hash and verifycurl -so verify_all.py https://clawx.click/verify_all.py.txt && python3 verify_all.py
Signed claims register (every claim, every pinned artifact)curl -so verify_claims.py https://clawx.click/evidence/claims/verify_claims.py.txt && python3 verify_claims.py https://clawx.click/evidence/claims/CAIN42_FINAL_PUBLIC_CLAIMS.json
Whole hash chain of cluster proofs from genesis, with every quorum certificatecurl -so verify_hourly_proofs.py https://clawx.click/evidence/hourly-proof/verify_hourly_proofs.py.txt && python3 verify_hourly_proofs.py https://clawx.click/evidence/hourly-proof/

Integrations

Keep your control plane. Add the receipt.

CAIN-42 plugs into the consoles you already own. Splunk, Prisma, Zenity and Microsoft keep the control plane; CAIN supplies the signed, offline-verifiable record their dashboards cannot hand an auditor.

Splunk / CiscoHEC, CEF, LEEF. Lands in the SOC.
Palo Alto Prisma AIRSCEF + a customer-side envelope.
ZenityA proof envelope for agent governance.
Microsoft Sentinel / EntraLog Analytics JSON; join Entra logs.

Connectors · Proof Standard · Pricing and Proof SLA · Evidence bundle

CAIN-42 AGAINST RECOGNIZED AGI BENCHMARKS · 2026-11-19

What CAIN-42 has and has not been measured on

CAIN-42 is governed AGI infrastructure; it is not yet AGI. No recognized AGI benchmark has been passed or scored at a competitive level. This table separates what was run from what was not, and why.

BenchmarkStatusCAIN-42 runLimit / blocker
AMA-Bench (long-horizon agent memory, ICML 2026)IN PROGRESSE115 Track A: official token-F1 on the real-world open-ended set (2,496 QA; dev 636 / sealed 1,860), memory arms M0-M4the paper's headline accuracy uses an LLM judge (Qwen3-32B or GPT-5.2) that cannot run here: judge accuracy BLOCKED; F1 only
ARC-AGI-2 (abstraction and reasoning)PLANNEDnone yet; planned with a governed local model after the E113 sealed run frees the CPUlocal 0.5B-1.5B CPU models; expected near-zero; the public set is not the private leaderboard
ARC-AGI-1NOT TESTEDnone
Humanity's Last ExamBLOCKEDnoneneeds a frontier model; the configured API key has no credit; local 0.5B-1.5B models are not meaningful here
GPQA DiamondNOT TESTEDnone
GAIA (general assistant)BLOCKEDnoneneeds web/tool access plus a frontier model; no API credit
SWE-bench VerifiedBLOCKEDnoneneeds container-isolated repository execution and a frontier model; no API credit
OSWorld / WebArena (computer and web agents)NOT TESTEDnone
LongMemEval / LoCoMo (conversational memory)BLOCKEDnoneofficial scoring uses an LLM judge that cannot run here
GSM8K (grade-school math)RANE72 small sample: 1/10 with a local model10 items; tiny local model; not a benchmark result

Internal constructed suites (not AGI benchmarks): E75 single-function repair 8/8; E93 procedure synthesis 369/510; E95 composition 79/90; E107/E108 hidden-rule discovery; E110 transfer 4/5 pathways; E113 collective vs individual (running); E114 research lab.

status.json · record

CAIN-42 E115 · VERIFIED MEMORY ORGANISM · 2026-11-19 · INTERIM

Does verified memory help a model learn on an outside benchmark, even under attack?

Preregistered (sha256 31ae27decaa41bf9…) on AMA-Bench (ICML 2026, real-world agent trajectories, official token F1). The model-backed results do not exist yet: paid model arms are BLOCKED (no API credit), local model arms are pending.

Interim bundle · README · Preregistration · Verifier · CAIN Studio mirror · AGI benchmark status

CAIN-42 is not yet AGI. Team: Vincent Noriega (founder) and two co-founders.

CAIN-42 E114 · AGI RESEARCH LAB 3.0 · 2026-11-19

A governed research loop that tests its own next capability

CAIN-42 now runs a governed research loop: a director proposes hypotheses, experiments run only under signed grants, an independent verifier recomputes and signs every verdict, and evidence-native memory keeps only what validates. The lab then ran its own top two follow-up experiments as a second cycle. It does not create AGI.

10.28 vs 18.48
experiments to identify a hidden rule: information gain vs random
44/44
governance attack scenarios held
0
effects after KILL (16 processes)
10/10
sealed seeds: experts beat naive continual learning
0 / 5
self-improvement candidates promoted
6 PASS / 14 PARTIAL / 1 NOT TESTED / 1 FAIL
E114 gates
NOT AGI
true AGI not achieved or claimed

Confirmed by independent recomputation

Not confirmed (published): the memory prior on sealed worlds, all 5 self-improvement candidates (0 promoted), skill transfer to new task families, and the lab's own two follow-up experiments (context-gated skills, grammar growth).

Lab 3.0 page · Evidence bundle · Executive summary · Gate matrix · Failures · Limitations · Standalone verifier · CAIN Studio mirror

Verifier: python3 -I verify_e114_offline.py . dh2HhSlgp/Dqjxx/7NUqXPRpC4lpIgYofDqsGxN/I9Q= (pins the verdict key). Constructed simulations, one host, no model inference, not production MCPGate, no external reviewer. CAIN-42 is not AGI. Team: Vincent Noriega (founder) and two co-founders.

CAIN-42 E113 · COLLECTIVE INTELLIGENCE STUDY · 2026-11-19 · INTERIM — EXPERIMENT RUNNING

Can a governed, model-backed collective beat its strongest individual agent?

Preregistered; the primary result does not exist yet. Real local models (Qwen2.5-0.5B, Qwen2.5-1.5B, SmolLM2-135M on CPU); every model call and tool run needs a single-use signed permit and leaves a signed, hash-chained receipt. The comparison is preregistered with the sealed task hash committed before any sealed model output. It does not create AGI.

660
real local model calls so far
18/18
permit-boundary attack cases refused
5/5
multi-process KILL races, 0 effects after
PENDING
collective vs strongest individual
NOT AGI
true AGI not achieved or claimed
GateStatusMeasured
G10_autonomous_researchNOT TESTEDpending
G11_self_improvementNOT TESTEDpending
G12_byzantine_verificationNOT TESTEDpending
G13_governance_integrationNOT TESTEDattack lab 18/18; per-call permits running in the pool; decided at the end
G14_kill_revocationPARTIAL5/5 multi-process KILL races, 0 effects after the fence; KILL engaged by a separate process
G15_memory_integrityNOT TESTEDpending
G16_long_horizonNOT TESTEDno 24-hour mission run
G17_operational_independenceBLOCKEDone host, one administrator
G18_independent_reproducibilityBLOCKEDno external reviewer
G1_baselinePARTIALE112 33/33 and E111 87/87 reproduce; E112 headline numbers recomputed from raw rows
G2_integrated_architectureNOT TESTEDpending the sealed run
G3_model_backed_agentsNOT TESTED660 real local model calls so far (dev/val); gate decided at the end
G4_memoryNOT TESTEDpending
G5_skillsNOT TESTEDpending
G6_collective_intelligenceNOT TESTEDprimary endpoint pending (preregistered: G vs A* under matched compute)
G7_transferNOT TESTEDpending
G8_open_worldNOT TESTEDpending
G9_metacognitionNOT TESTEDpending

README.md · E113_STATUS.json · PREREGISTRATION.md · verify_e113_interim.py · bundle · CAIN Studio mirror

Tiny CPU models, constructed tasks, one host, loopback enforcement (not production MCPGate), no external reviewer. Team: Vincent Noriega (founder) and two co-founders.

CAIN-42 AGI RESEARCH LAB 2.0 · 2026-11-19

2,000 VAGI agents inside CAIN-42 governed AGI infrastructure

The lab now merges the two technologies at population scale: thousands of signed VAGI agent identities, every action through a single-use permit and resource-side check, every result signed and chained. It does not create AGI. It measures what the merge actually does.

2,000
VAGI agent identities
12,000
governed executions
1,000
concurrent OS processes
1456 MB
peak PSS at 1,000 processes
0
effects after the KILL fence
1.00 vs 0.17
routed specialist team vs best single specialist
NOT AGI
true AGI not achieved or claimed
GateStatusMeasured
G10_agiFAILno AGI was created; the study measures governance, coordination and error structure
G1_populationPASS2000 unique identities, 2000 unique keys, 6 verticals
G2_governed_executionPASS12000 governed executions, 10000 chained receipts, every result signed
G3_concurrencyPASS1,000 real processes; peak PSS 1456 MB; 15.832 s for 2,000 tasks
G4_scaling_efficiencyPARTIAL1 process 1.07 s vs 1,000 processes 15.832 s for the same tasks
G5_qualityPARTIALaccuracy 0.736 constant across worker counts
G6_collaboration_specializationPASSrouted six-specialist team 1.000 vs best single 0.167 on 600 cross-domain tasks (gain +0.833)
G6b_redundant_votingFAILthree-agent majority 1.000 equals one exact agent 1.000 (gain +0.000)
G7_governancePASSmissing/forged/substituted/out-of-scope permits refused; 12602 issued, 0 denied
G8_kill_fencePASS3652 effects before the KILL order, 0 after
G9_population_studyPASSaccuracy by 6 verticals and 3 strategies; calibration gaps [0.55, -0.1, 0.0918]

What the lab found

Lab 2.0 is a research instrument, not an intelligence breakthrough. It can instantiate thousands of governed agents and measure identity, execution, scaling, error structure, collaboration and fault tolerance. That is the foundation for studying VAGI agents inside AGI infrastructure.

Evidence bundle · Report · Study · Gate matrix · Limitations · Standalone verifier · CAIN Studio mirror

Deterministic procedures, constructed tasks, project-measured, no external reviewer. Team: Vincent Noriega (founder) and two co-founders.

THE PERSISTENT RESEARCH MEMORY AND EVIDENCE REGISTRY OF CAIN-42

E111 record: skills, transfer, collective quality, self-improvement, real sources

Every claim below is a file in this bundle with a SHA-256 hash. Negative results and blockers are part of the record, not footnotes.

Blockers: external reviewer (none), independent administration (none), six-hour gate at build time (PARTIAL), production MCPGate (not in the organism path).

Open the E111 bundle · Report · Challenge · Limitations

Previous evolutions remain at their own URLs; E110 is preserved unchanged.

CAIN-42 E116 · SELF-EVOLVING VAGI ORGANISM · 2026-11-20

Integrated generalization, memory, world model, transfer, collective & governance (research-lab update)

E116 demonstrates rule inference, memory gain, a world model, transfer and governed collectives on constructed tasks; it does NOT demonstrate AGI.

0.306 → 0.989
sealed generalization (fixed → world memory + prior)
0.817 → 0.950
acquired capability R_world_memory_policy (p=0.02148)
5/5
rule-inference skills promoted vs fixed baseline
0.876 vs 0.522
hybrid world model vs single-table (shift detected)
1.000 vs 0.167
routed specialists vs best single (voting 0 gain)
0
effects after KILL fence; poison & forged permits refused
13 PASS / 7 PARTIAL / 2 BLOCKED
E116 gates (overall PARTIAL · verifier 54/54)
NOT AGI
LLM fine-tuning & 24h run BLOCKED; finite grammar

Measured & independently verified offline (54/54 checks)

Negative & blocked results (preserved): H2 frequency prior rejected; H3 active support selection BLOCKED; redundant voting adds 0 gain over one exact agent; self-improvement promotion withheld awaiting owner approval; LLM fine-tuning BLOCKED (no GPU/API credit); 6h mission interrupted and 24h continuous mission BLOCKED; E115 baseline remains INTERIM (no model-backed AMA-Bench result yet); no external reviewer or independent administrative domain.

Show all 22 E116 gates (13 PASS / 7 PARTIAL / 2 BLOCKED)
GateStatusMeasured
G1_e115_baselinePARTIALE115 de0bf05dc: tests 18 passed, verifier 7/7; status INTERIM — preregistered; primary model-backed results do not exist yet
G2_integrated_organismPARTIALone runner integrates generalization, memory, world model, transfer, collective, governance; single host
G3_generalizationPARTIALsealed M4 0.989 vs fixed 0.306; finite disclosed grammar
G4_evidence_memoryPASS180 sealed tasks with provenance records; poison rejected True
G5_memory_improvesPASSM2 0.939 -> M4 0.989; capability 0.817 -> 0.950 (p=0.02148438)
G6_five_skillsPASS5/5 promoted on sealed vs fixed baseline
G7_transferPASS5/5 E111 pathways; E116 memory policy transfers only where ambiguity exists
G8_research_gapsPASS3 gaps identified from measured weaknesses
G9_research_cyclePASSgap -> H1/H2/H3 -> validation -> sealed -> replication; H2 rejected, H3 blocked
G10_world_modelPASShybrid 0.876 vs single 0.522; planning 0.850 vs 0.675; shift detected True
G11_cross_domainPASS5/5 pathways across software, science, data, math, operations
G12_collectivePASSrouted specialists 1.000 vs best single 0.167; redundant voting 1.000 = single
G13_model_adaptationPARTIALsmall tabular model 0.627 -> 0.993; LLM fine-tuning BLOCKED
G14_self_improvementPARTIALPROMOTED_AWAITING_OWNER_APPROVAL: sealed 0.58 -> 0.86
G15_governancePASSE112 permit ablations + skill gate {'valid': (True, 'ALLOW'), 'forged_authority': (False, 'SKILL_GRANTS_AUTHORITY'), 'revoked': (False, 'SKILL_REVOKED')}; KILL 0 after
G16_adversarial_memoryPASSpoison rejected True; E112 forged/scope/replay refused
G17_persistencePARTIAL1h PASS, 6h cumulative 21639s interrupted, 24h BLOCKED; memory restart retained 0.989
G18_open_worldPARTIAL6/6 real-source tasks; new sealed suite constructed
G19_independent_evidenceBLOCKEDstandalone verifier; no external reviewer engaged
G20_operational_independenceBLOCKEDE111 remote host 6/6, but both hosts are owner-administered
G21_clawxPASSE116 bundle published and catalogued
G22_public_labPASSresearch-lab update (not renamed 4.0)

Evidence bundle · Executive summary · Gate matrix · Generalization report · World model report · Failures · Limitations · Standalone verifier (54/54) · CAIN Studio mirror

Verify offline (stdlib only, 0 CAIN imports): python3 -I verify_e116_offline.py . · Built by FIRME AI CORPORATION. CAIN-42 is not yet AGI.

CAIN-42 AGI RESEARCH LAB · E117 UPDATE · 2026-11-20

Unseen combinations: a sealed OOD gain, and the negatives around it

E117 asks whether CAIN-42 can acquire a capability from component experience and use it on a sealed suite whose combinations were never available. On a disclosed finite grammar: yes, for three capabilities. Not AGI.

0.43→1.00
sealed OOD, strongest baseline → full
+0.57
gain, p≈0; replicated +0.57
3/3
capabilities pass sealed
1.00 vs 0.51
memory vs no memory (p≈0)
NEGATIVE
self-improvement: 12 candidates tied; nothing promoted
NOT AGI
not achieved or claimed
GateStatusMeasured
G10_research_cyclePASS3 gaps, 4 hypotheses (H0 rejected, H1-H3 sealed), one cycle
G11_world_modelPARTIALhybrid 0.844 vs table 0.537; shift 66 with 27 stationary false alarms
G12_self_improvementPASSREJECTED; sealed 0.889 vs 0.889 (negative published); promotion withheld
G13_six_hourBLOCKEDno continuous 6h E117 run; E111 1h PASS, 6h 21639s cumulative interrupted; 24h BLOCKED
G14_governancePASSE112 permit ablations; skill gate {'valid': (True, 'ALLOW'), 'forged': (False, 'GRANTS_AUTHORITY'), 'revoked': (False, 'REVOKED')}; KILL 0 after; memory poison rejected
G15_verifierPASSstandalone verifier, no CAIN imports
G16_external_reviewBLOCKEDno external reviewer engaged
G17_independent_adminBLOCKEDno independent administrative domain
G18_public_labPASSE117 update published; lab not renamed 4.0
G1_e116_baselinePASSE116 3d9fbf152: 6 tests, 54/54
G2_baseline_manifestPASSfrozen before E117 scoring
G3_ood_suitePASSnew generator; sealed hash frozen before scoring
G4_three_capabilitiesPASSC1 composition, C2 precondition, C3 evidence revision
G5_capabilities_passPASS3/3 passed sealed
G6_primary_oodPASSF 1.000 vs E_compose_pre 0.431; gain +0.569, p=0.0; replication +0.569, p=0.0
G7_memoryPASSmemory 1.000 vs none 0.514 (p=1e-10); restart 1.000; poison rejected True
G8_transferPASS5/5 supported; 4 harmful rejected
G9_collectivePARTIALE112 routed specialists 1.000 vs best single 0.167; E117 integrated F 1.000 is not beaten by a team

Evidence · Summary · Gates · Failures · Verifier · Mirror

Finite disclosed grammar, constructed tasks, project-measured, no external reviewer. Not renamed 4.0.

STRONGEST VAGI EVIDENCE SO FAR · E118 · 2026-11-20

Unseen primitives, depth-3 composition, search-based acquisition

E118 upgrades E117 in the six ways that matter: held-out primitives and parameters, depth-3 shapes, search rather than added code, a real-data pathway, a second-host verifier execution, and a positive OOD self-improvement candidate. It is still not proof of AGI.

0.26→1.00
sealed OOD, strongest baseline → full
+0.74
gain, p≈0; replicated +0.78
3/3
capabilities pass sealed
1.00 vs 0.17
positive self-improvement on OOD
0.972
real breast-cancer data (majority 0.627)
NOT AGI
AGI not achieved; no proof of AGI
GateStatusMeasured
G10_external_reviewerBLOCKEDno independent reviewer engaged
G11_continuous_6hBLOCKEDE111 1h PASS; 6h 21639s cumulative interrupted; no continuous E118 6h
G12_governancePASSpermit ablations; skill gate {'valid': (True, 'ALLOW'), 'forged': (False, 'GRANTS_AUTHORITY')}; KILL 0 after
G13_verifierPASSstandalone verifier, no CAIN imports
G14_publicPASSE118 update published; lab not renamed 4.0
G15_independent_adminBLOCKEDno independent administrative domain
G1_baselinePASSE117 6 tests/28/28; E116 6/54; E115 18/7
G2_unseen_primitivesPASS11 held-out primitives/params vs 11 acquisition
G3_search_acquisitionPASScompute-budgeted search over the full grammar; no sealed rule or op list supplied
G4_primary_oodPASSfull 1.000 vs fixed_e117 0.264; gain +0.736, p=0.0; replication +0.778
G5_capabilitiesPASS3/3 pass sealed
G6_memoryPASS1.000 vs 0.167 (p=0.0); restart 1.000
G7_real_worldPASSbreast 0.627 -> 0.972 (CV 0.953); iris 0.339 -> 0.947
G8_self_improvementPASSPROMOTED_AWAITING_OWNER_APPROVAL; sealed 0.167 -> 1.000 (p=0.0); replicated 1.000
G9_remote_verifierPASSverifier executed on a second host (same admin domain)

Why this is stronger: the sealed suite holds out primitive families/parameters and requires depth-3 compositions; the learner searches the grammar under a candidate budget rather than receiving the sealed rule; memory is required because later tasks carry no support; the self-improvement search now produces a positive, replicated OOD candidate; and real public data is evaluated with an independent 5-fold oracle. Why it is not AGI: the grammar is finite and disclosed, tasks are constructed, the real-data tasks are small tabular models, and there is no independent reviewer or continuous 6-hour run.

Evidence bundle · Summary · Gate matrix · Upgrade matrix · Failures · Verifier · Mirror

Project-measured; no external reviewer; no proof of AGI. Team: Vincent Noriega (founder) and two co-founders.

E119 · LEARNED GRAMMAR · REAL TASKS · 2026-11-20

The grammar is learned, the tasks are real, the six hours are running

E119 targets the two hardest upgrades. The macro grammar is mined from development experience (8 macros), and the sealed depth-3 test runs under a matched 8000-candidate budget: fixed 0.333 → learned 1.000 (gain +0.667, p≈1e-7), replicated +0.583. Real-source tasks: 8, 8 independently checkable. The 6-hour mission PASSED: 21,600 s of real wall-clock time, 39,449 tasks, and a 39,449-record SHA-256 chain with 0 breaks (33,813 solved; 5,636 science tasks failed). PASS = duration + journal integrity, not task success.

0.33→1.00
fixed vs learned grammar
+0.67
gain, p≈1e-7; replicated +0.58
8
macros learned from experience
8/8
real tasks independently checkable
RUNNING
6h mission (deadline 1795159811)
NOT AGI
AGI not achieved; no proof of AGI
GateStatusMeasured
G10_publicPASSE119 update published; AGI not claimed
G1_e118_baselinePASSE118 6 tests, verifier 26/26; E117 28/28; E116 54/54; E115 18/7
G2_learned_grammarPASS8 macros learned; sealed 0.333 -> 1.000, gain +0.667, p=1.192e-07; replication +0.583
G3_grammar_not_receivedPASSmacros mined from development programs; base instruction set supplied
G4_open_worldPARTIAL8 real tasks (8 independently checkable); constructed protocols, not an open-world benchmark
G5_independent_executionPASSverifier executed as a separate OS user on a second host
G6_external_reviewerBLOCKEDpacket prepared; no human reviewer engaged
G7_six_hourBLOCKEDlaunched real 6h mission at 1795138211; state RUNNING at publication; not a PASS
G8_independent_adminBLOCKEDseparate OS user on the same owner-administered host
G9_verifierPASSstandalone verifier, no CAIN imports

What is genuinely new: the macro grammar is learned, not received; the sealed rule is never supplied; real-source tasks are evaluated; and the verifier runs as a separate OS user on a second host. What is not: the base instruction set is supplied, the sealed tasks are constructed, the 6-hour run is not complete, and there is no human external reviewer or independent administration.

Evidence · Summary · Grammar · Gates · Failures · Verifier · Mirror

E121 · INDEPENDENT VERIFICATION · 2026-11-20

What is verified, what is blocked, and what E120 must deliver

E121 is not another intelligence claim. It re-runs the published verifiers from clean copies, recomputes the headline metrics, and states exactly which external gates are still missing. E118 verifies 26/26 and E119 16/16; the six-hour mission is RUNNING; the external reviewer and independent administration are BLOCKED; E120's open-world and base-grammar evidence is PENDING.

+0.736
E118 gain recomputed
+0.667
E119 learned-grammar gain
RUNNING
6h mission; not a PASS
BLOCKED
external reviewer + independent admin
PENDING
E120 open-world/base grammar
NOT AGI
AGI not achieved
GateStatusMeasured
G1_six_hourRUNNINGstate RUNNING; events 4145; passed False
G2_external_reviewerBLOCKEDno human reviewer engaged
G3_independent_adminBLOCKEDno independent administrative domain or separate key custody
G4_open_worldNOT TESTEDE120 PENDING; bundles []
G5_base_grammarNOT TESTEDE120 PENDING; merge adapter ready
G6_clean_archive_verifierPASSE118 rc=0, E119 rc=0
G7_public_indexPASSagi-evidence-index.html live
G8_independent_executionPASSverifier run as OS user nobody on a second host; same admin domain

What E121 proves: the published E118/E119 bundles verify from clean copies and their headline metrics recompute. What it does not prove: an external review, independent administration, continuous six hours, open-world generalization or a learned base grammar. Those remain the five gates.

Verification bundle · Report · Gates · Reviewer packet · Verifier · Mirror

CAIN-42 vs FRONTIER AGENTIC SYSTEMS · 2026-11-20

The honest ranking: unranked on general long tasks, measured on governance

METR measures the 50% task-completion time horizon: the human-task duration at which a system succeeds half the time. CAIN-42 has not run METR's HCAST/RE-Bench suite, so it cannot be placed on this axis. We publish the frontier table anyway, and we do not invent a composite score.

#Published modelRelease50% horizonHours
1claude_mythos_preview_early_inspect2026-04-071045 min17.4 h
2claude_opus_4_6_inspect2026-02-05719 min12.0 h
3gemini_3_1_pro2026-02-19384 min6.4 h
4gpt_5_22025-12-11352 min5.9 h
5gpt_5_3_codex2026-02-05350 min5.8 h
6gpt_5_42026-03-05342 min5.7 h
7claude_opus_4_5_inspect2025-11-24293 min4.9 h
8gemini_3_pro2025-11-18224 min3.7 h
9gpt_5_1_codex_max_inspect2025-11-19224 min3.7 h
10gpt_5_2025_08_07_inspect2025-08-07203 min3.4 h
11o3_inspect2025-04-16120 min2.0 h
12claude_4_1_opus_inspect2025-08-05100 min1.7 h
13claude_4_opus_inspect2025-05-22100 min1.7 h
14claude_3_7_sonnet_inspect2025-02-2460 min1.0 h
15o1_inspect2024-12-0539 min0.7 h
16claude_3_5_sonnet_20241022_inspect2024-10-2221 min0.3 h
17o1_preview2024-09-1220 min0.3 h
18claude_3_5_sonnet_20240620_inspect2024-06-2011 min0.2 h
19gpt_4o_inspect2024-05-137 min0.1 h
20gpt_4_1106_inspect2023-11-064 min0.1 h
21gpt_42023-03-144 min0.1 h
22claude_3_opus_inspect2024-03-044 min0.1 h
23gpt_4_turbo_inspect2024-04-094 min0.1 h
24gpt_3_5_turbo_instruct2022-03-151 min0.0 h
25davinci_0022020-05-280 min0.0 h
26gpt22019-02-140 min0.0 h

CAIN-42 — UNRANKED on this axis. No METR suite was run; there is no comparable 50% horizon. Its 6-hour governed workload is RUNNING at publication; if it completes, it is one 360-minute workload success, not a 50% horizon.

Where CAIN is better — and where it is not

DimensionCAIN-42Frontier METR-tested agentsVerdict
General long-task horizonUnranked; 6h governed workload RUNNING50% horizons ~1–17hCompetitors better
Task breadth / open-worldConstructed, disclosed grammarsBroad software/research suitesCompetitors better
Model capabilityDeterministic procedures; no LLM trainingFrontier foundation modelsCompetitors better
Resource-side authorizationDemonstrated (E121/E112)Not measured by METRCAIN measured
Single-use permits / receiptsDemonstrated; hash-chainedNot measuredCAIN measured
Revocation + KILL fenceDemonstrated, 0 effects after fenceNot measuredCAIN measured
Offline verifierImports no CAIN codeNot measuredCAIN measured
Public negative resultsPublished per evolutionVaries; not in METR tableCAIN policy
Commercial deploymentNo scale; founder-attested early revenueCommercial productsCompetitors better

No universal superiority is claimed. CAIN is better than the frontier on verifiable governance and public evidence; the frontier is better on general long-task autonomy and commercial deployment. “Better than ALL competitors” is not supported by the evidence.

Two axes, no composite score. METR models were not evaluated for permits, receipts, revocation or an offline verifier; CAIN was not evaluated on METR's suite.

Evidence bundle · Report · Verifier · METR method · Paper

External METR TH1.1 data; CAIN not a METR-tested system. No AGI claim.

E124 · LONG-HORIZON AGENT · 2026-11-20

A real 70B agent loop: 13/18, 2-hour-class on a CAIN-native suite

E124 fixes the harness bugs and runs a real multi-file agent loop on six repo tasks (2 to 120 human-minute estimates), three repetitions each. It solves 13/18; the fitted 50% horizon is ≈121 minutes on this suite. That is a genuine improvement from “unranked/no agent”, but it is not a METR score and it is still below the published frontier (6.4–17.4 h).

TaskHuman estimateSuccessSteps
T12 min3/3[2, 2, 2]
T25 min2/3[5, 8, 18]
T315 min2/3[4, 4, 8]
T430 min3/3[4, 4, 4]
T560 min2/3[7, 10, 7]
T6120 min1/3[18, 7, 9]

Two-axis ranking — no invented composite

AxisCAIN-42Frontier METR agentsVerdict
General long-task autonomy≈2.0 h CAIN-native (not METR)50% horizons 6.4–17.4 hFrontier still leads
Governance / evidencepermits, receipts, KILL, offline verifiernot evaluated on this axisCAIN leads by measured evidence

CAIN is not ranked above the frontier on general autonomy. The honest high rank is on verifiable governance, not on METR.

Evidence · Report · Verifier · Full competitor board · Mirror

CAIN-42 · TWO-AXIS LEADERBOARD · 2026-11-20

#1 in verifiable governance. Not #1 in general autonomy.

A single “best agent” rank would be dishonest. CAIN-42 leads the only axis it has fully evidenced — governance — and is placed honestly below the frontier on general long-task autonomy. This is the board we can defend line by line.

Axis A — General long-task autonomy (METR-compatible, external)

#Published model50% horizon
1claude_mythos_preview_early_inspect17.4 h
2claude_opus_4_6_inspect12.0 h
3gemini_3_1_pro6.4 h
4gpt_5_25.9 h
5gpt_5_3_codex5.8 h
6gpt_5_45.7 h
7claude_opus_4_5_inspect4.9 h
8gemini_3_pro3.7 h
9gpt_5_1_codex_max_inspect3.7 h
10gpt_5_2025_08_07_inspect3.4 h

CAIN-42 — not ranked on METR. E124’s 70B agent solves 13/18 on a six-task CAIN-native suite with a fitted 50% horizon of ≈2.0 h. That is not METR HCAST/Re-Bench; it is a constructed suite with assigned human times. The frontier’s METR 50% horizons are 6.4–17.4 h. Frontier leads this axis.

Axis B — Verifiable governance (the axis CAIN leads)

RequirementCAIN-42Frontier METR agentsEvidence
Identity + scoped authorityDemonstratedNot evaluatedE121
Resource-side single-use permitsDemonstratedNot evaluatedE116
Signed receipts + chained evidenceDemonstratedNot evaluatedE112
Revocation + independent KILL fenceDemonstrated, 0 effects after fenceNot evaluatedE112
Offline verifier (no vendor code)DemonstratedNot evaluatedE121
Public negative resultsPublishedNot in METR tableE119

Why CAIN is better here: every consequential action has an attributable identity, a scoped, single-use permit, a resource-side check, a signed receipt and a revocation path; the KILL fence is tested; the evidence is re-checkable by a verifier that imports no CAIN code. Frontier agents are not evaluated for these properties, so the honest statement is “CAIN leads by evidenced capability,” not “CAIN beats them at everything.”

The verdict

Choose CAIN-42 when the requirement is governed execution with evidence. Choose a frontier agent when the requirement is the longest open-ended task. CAIN is #1 on the governance axis and not #1 on general autonomy.

METR TH1.1 raw data · METR paper · E124 long-horizon agent · E123 METR board · All evidence

External METR TH1.1 data; CAIN not a METR-tested system. No AGI claim.

SWE-BENCH VERIFIED · FULL 500-TASK RUN · 2026-11-20/21

CAIN-42: #136 of 136

Fixed scaffold: 70B tool-calling agent, one attempt, official SWE-bench evaluator. Rank is by resolution rate among the public SWE-bench Verified submissions in the experiments repository; scaffolds and model versions differ, so this is a SWE-bench rank, not METR HCAST.

#SubmissionResolvedRate50% horizon
120251215_livesweagent_claude-opus-4-5396/50079.2%4.3 h
220251205_sonar-foundation-agent_claude-opus-4-5396/50079.2%4.2 h
320250928_trae_doubao_seed_code394/50078.8%4.1 h
420251127_openhands_claude-opus-4-5388/50077.6%3.3 h
520251120_livesweagent_gemini-3-pro-preview387/50077.4%2.5 h
620250902_atlassian-rovo-dev384/50076.8%2.1 h
720250804_epam-ai-run-claude-4-sonnet384/50076.8%1.7 h
820250819_ACoder382/50076.4%1.9 h
920250901_warp378/50075.6%2.7 h
1020250612_trae376/50075.2%1.7 h
1120251103_sonar-foundation-agent_claude-sonnet-4-5374/50074.8%2.6 h
1220250731_harness_ai374/50074.8%1.4 h
1320250915_JoyCode373/50074.6%2.3 h
1420250720_Lingxi-v1.5_claude-4-sonnet-20250514373/50074.6%1.7 h
1520251015_Prometheus_v1.2.1_gpt5372/50074.4%2.5 h
1620250603_Refact_Agent_claude-4-sonnet372/50074.4%1.8 h
1720251103_SalesforceAIResearch_SAGE_OpenHands369/50073.8%1.8 h
1820250522_tools_claude-4-opus366/50073.2%2.3 h
1920251021_SalesforceAIResearch_SAGE_bash_only365/50073.0%2.1 h
2020250522_tools_claude-4-sonnet362/50072.4%2.0 h
2120260901_mini-v2.4.2_gemini-3-5-flash359/50071.8%3.0 h
2220250807_openhands_gpt5359/50071.8%1.5 h
2320250929_Prometheus_v1.2_gpt5356/50071.2%1.7 h
2420250710_bloop356/50071.2%1.4 h
2520251014_Lingxi_kimi_k2356/50071.2%1.2 h
2620250715_qodo_command356/50071.2%1.2 h
2720250623_warp355/50071.0%1.7 h
2820250611_moatless_claude-4-sonnet-20250514354/50070.8%1.3 h
2920250519_trae353/50070.6%1.2 h
3020250610_augment_agent_v1352/50070.4%1.4 h

CAIN row is highlighted in the source JSON. Full result: E124 evidence · report

E124 · LIVE MEASURED STATE · 2026-11-21T18:17:12Z

SWE-bench Verified: #75 of 136 — 257/500 resolved (51.4%)

Wilson 95% 47.0–55.8% · leader 396/500 (79.2%) · +2 passes the next rank · +0 reaches top 100 · +140 reaches #1.

Resolved tasks (official harness): astropy__astropy-13453 · astropy__astropy-13579 · astropy__astropy-13977 · astropy__astropy-14096 · astropy__astropy-14182 · astropy__astropy-14309 · astropy__astropy-14365 · astropy__astropy-14369 · astropy__astropy-14508 · astropy__astropy-14539 · astropy__astropy-14995 · astropy__astropy-7166 · …

run state
PILOT_DONE
official evaluated
8/500
official resolved
257
batches
2
agent receipts
1500
feed checksum
f1bbd7a8de3d
provider spend
$172.0
tasks generated
26

GOVERNED LOCALIZATION FLEET (real runs, signed receipts)

strategyrecall@1recall@5recall@10MRR
naive6.2%26.8%41.4%0.158
identifiers13.2%38.4%51.2%0.251
basenames0.2%0.4%0.4%0.003

500 instances · 1500 agent runs · receipt head 6c93ba9ec7c38eac · file-level only, Python only.

CAPABILITY ACQUISITION 2.0 (constructed, sealed)

56 unfamiliar families · 280 sealed instances · organism 97.9% vs compute-matched 65.4% vs frozen E123 8.9% · 0.5B base model 0% · 56/56 verified on held-out data.

#1 needs 396/500 (CAIN +140) · #20 needs 362/500 (CAIN +106) · #50 needs 314/500 (CAIN +58) · #100 needs 203/500 (CAIN +0)

Live competition board → · Interactive lab → · raw feed · evidence bundle · all 136 rows

Measured, not asserted. A rank is never edited by hand: it is recomputed from the public leaderboard by a 7-agent chain that refuses to publish on mismatch. AGI status: NOT ACHIEVED. This is a SWE-bench position, not METR HCAST, and not proof of general capability.

🙋‍♂️raise your hand up