CAIN-42 CLAWX
CAIN-42 Proto-VAGI is the infrastructure governor of AGI. Early, measured signs of intelligence; every governed action comes with a signed receipt. Evidence · Auditor index · VAGI Agents · not yet AGI
E81 · 2026-11-16

Evolution 81 — Sovereign autonomous intelligence workforce

30 disciplines / 90 specialist profiles / 99 durable identities · killed worker's lease recovered, late commit STALE, 0 duplicate effects · 60/60 adversarial scenarios blocked · 1,475 governed local model calls (0 remote, $0 model spend) · collaboration honestly rejected (single 10/24 vs best arm 10/24, delta 0.0, p = 1.0) · 1 h endurance (3,600 s, 114 cycles, 0 failures, 2 recoveries) · 90/90 tests · 26 PASS / 12 PARTIAL / 2 BLOCKED / 0 FAIL (40 gates) · NOT RELEASED, not deployed.

Evidence bundle · ZIP · CAIN Studio

E80 · 2026-11-16

Evolution 80 — Adaptive Routing & Coordination: the collaboration advantage did not replicate

1,154 real local model calls · 48 pre-registered unseen tasks · router 22/48 vs best fixed path 19/48 (McNemar p = 0.65, not significant) · no multi-agent arm beat one agent (single 11/48, best evidence-first 16/48, committee 10/48 worse) · fixed two-model composition 27/48, better than the router · 24-task replication repeated the direction (10 vs 8) but not significance · 166/166 offline checks, 52/52 independently authored · security lab 51/51 · 1 h endurance (3,600 s, 640 tasks, 0 failures) · 13 PASS / 10 PARTIAL / 1 FAIL / 5 BLOCKED / 3 NOT EVALUABLE · NOT RELEASED, not deployed.

Evidence bundle · ZIP · CAIN Studio

E79 · 2026-11-15 · final

Real Model-Backed Collective Intelligence (measured)

986 real local model calls · 15 signed specialists · 104-task paired benchmark · single 34.4% vs four 25.0% vs eight 31.2% · 57/57 adversarial scenarios blocked · 144/144 offline checks recomputed from raw rows · cloud BLOCKED_BILLING · not deployed.

Full evidence page · Signed bundle · ZIP · CAIN Studio

← ClawX · Changelog

CAIN-42 Evolution history

Preserved development milestones. Each entry links to its evidence; later entries do not rewrite earlier evidence.

E79 · 2026-11-15

Real Model-Backed Collective Intelligence

1,186 real local model calls; 15 signed specialists; 100-task benchmark. Four specialists + calculator reached 29% vs 1% single direct and 11% single calculator. Held-out replication 40% vs 0%. 22/22 offline checks. MemoryAgentBench BLOCKED; cloud BLOCKED_BILLING; not deployed.

Evidence bundle · ZIP · CAIN Studio

E78 · 2026-11-15

Collective Organizations, Research & 1-Hour Endurance

15 organizations, 1,000 control-plane records, 100 local experiments, real cross-host task and partition drill, 3,600.043 s campaign with 3,578 tasks verified by a second agent, 10-level kill switch, 15 local ClawX views. C+ (6.94).

E78 evidence · changelog

E77 · 2026-11-14

Infinite Mission Swarm Fabric

75/75 Agentic75 mission integration, 1,000,411 durable rows, 410 tasks verified by a second agent, real SmolLM2 inference, AES-256-GCM memory, bubblewrap sandbox, 27/27 offline checks. C+ (6.58).

E77 evidence

E74–E76

Governed AOS, Cognitive Generalization and Supreme Fabric

Preserved baselines; see the changelog and the evidence registry for their signed bundles.