CAIN-42 · Coming next

CAIN Agent Lab

A safe lab to try agent ideas before they touch anything real.

Last reviewed 2026-10-01

In development   Product lines in development · Product line

Where it fits

Part of Coming next: Product lines in development. Tested code, not yet sold or hosted. Every CAIN-42 product runs behind the same rule: an AI agent's action is checked before it runs (identity, authority, policy, risk), decided as allow, hold for a human, or block, and recorded as signed evidence. Unknown or error never becomes allow.

Use it

• Evidence bundle https://clawx.click/evidence/e35-governance-intelligence-2026-09-30/

Evidence behind it

What E35 claims, and how far each claim goes

• C42-E35-INTELLIGENCE-NOT-AUTHORITY TESTED: the whole intelligence fabric refuses to be authority; every component records grants_authority=false and a recommendation is only a recommendation
• C42-E35-EPISTEMIC TESTED: OBSERVED, INFERRED, PREDICTED, SIMULATED, COUNTERFACTUAL and UNKNOWN are never merged; a counterfactual is never an observed fact
• C42-E35-REVIEW TESTED: a learned recommendation becomes a deployed control only after a deterministic review and a canary; the review result still passes the E33 kernel and its E8 binding
• C42-E35-SELFHEAL TESTED: self-healing, the red team, degradation and self-governance can never expand authority; degradation only reduces it
• C42-E35-BENCH TESTED: 3544 of 3544 scenarios across 68 categories held; 799 of 799 invariants; mutation 15 of 15
• C42-E35-CLEANROOM TESTED: an independent verifier that imports none of CAIN re-derives the E35 invariants and rejects every crafted forgery
• C42-E35-HONEST-SCOPE TESTED: the intelligence layer is NOT in the trusted root; the governance cloud is NOT DEPLOYED; predictions are modelled over synthetic features and are not calibrated against real incidents

Known limits of E35

• An in-process TESTED library; not hosted. The intelligence layer is deliberately NOT in the trusted authorization root; it can only produce recommendations.
• Predictions, counterfactuals and simulations are modelled in-process over synthetic features; they are not calibrated against real production incidents because none are ingested here.
• The red team, blue team, research lab, tournament and marketplace run in-process; there is no external ecosystem, no third-party publisher and no customer.
• The technology radar and research graph are seeded from a small set of items; provenance is recorded but coverage of the literature is incomplete.
• Conformance, deployment-readiness and governability scores are internal multi-dimensional views; they are not certifications and never collapse into a single number.
• No learned component may be placed in the trusted root; the deterministic kernel and E8 remain authoritative.

Tested guarantees in this area

Every rule in these niches has its own page with its recorded result.

• Autonomy, control loops & recovery: 407 tested invariants — Agents acting on their own, and how CAIN keeps them in bounds and recovers.
• Transactions, markets & economics: 664 tested invariants — Agents that buy, sell, bid and pay, within limits.

Related

• CAIN Agent Builder — Build new agents with governance designed in from the start.
• CAIN Agent Conformance — Check an agent against a published conformance standard.
• CAIN Agent Evaluation — Deep evaluation of agents before they are trusted with real work.
• CAIN Agent Factory — Produce many governed agents from one approved template.
• CAIN Agent Hosting — Host your agents on CAIN, governed by default.
• CAIN Agent Labor — Rules for agents doing paid work: who did what, and who gets credit.
• CAIN Agent Marketplace — Find and hire governed agents built by others.
• CAIN Agent Organization — Teams of agents with roles, a constitution and recorded decisions.
• CAIN Agent Passport — A signed passport that says who an agent is and what it may do.
• CAIN Agent Procurement — Let agents buy things for you, within limits you set.