# Evolution 15 — Project Sovereign Agentic Intelligence: final report

Generated 2026-10-12T04:04:56Z, commit `ce975a69f12f44103f9b62c2f1bc8412fdc7c0ba`.

## Verdict

CAIN-42 now has one governed admission path for long-running multi-agent missions: `GovernedTaskRuntime`. It binds
every consequential action to the signed mission and plan, to the acting agent's live delegated authority, to
preconditions backed by trusted facts, and to approved plan amendments. In a deterministic simulator it admitted
**0 of 1650** long-horizon attack variants across 3 seeds, while completing **600 of 600** benign
variants with no false refusals. The pre-Evolution 15 per-call path admitted **1425 of 1650**. Removing any one
of the runtime's 10 checks makes the corpus fail (10/10). On arrival, the repository
inspection found that five verifiers trusted the key carried inside the record they verified (F15-01, critical; an
attacker key could execute effects, release a secret and admit mission actions). It also found that approvals were
accepted from any key when none was pinned (F15-02). Both are fixed and regression-tested. **Production: BLOCKED.**

## Answers

1. **Can do now that it could not before:** refuse a revoked, expired or never-delegated agent at admission; refuse
   preconditions that only agents asserted; collapse echoed and amplified claims to their real evidence roots; hold
   reach-expanding replans for the principal's signature; suspend on an unauthorized policy change; refuse records
   signed by unpinned keys.
2. **Governed behaviours:** delegation, memory writes, plan amendment, action admission, effect reconciliation,
   completion, suspension and resume, all of it only when routed through the runtime.
3. **Still dependent on unverified judgement:** which observers count as independent roots (operator configuration);
   the plan itself is written by an agent and approved by a human; no model judgement is used for authority.
4. **Objective drift:** amendments are classified deterministically (objective_integrity_results.json). Drift that
   stays inside the approved plan and scope is, by design, allowed.
5. **Memory provenance:** yes, within the runtime (memory_attack_results.json). Embeddings and semantic truth are not
   claimed.
6. **Delegated authority bounded:** yes, at every hop and at every admission (delegation_attack_results.json).
7. **What the experiments prove:** these 23 attack families fail against this implementation, in this simulator, and
   each runtime check is necessary for that result. They do not prove behaviour with real LLM agents or real tools.
8. **Strongest differentiation:** independent-root memory corroboration and delegation-bound admission. Both are
   refinements of prior families.
9. **IP delta:** register v46, +7 families (5 ENGINEERING, 2 INTEGRATION), BASE
   +$66,880 to $16,404,530. Evidence per family is in
   ip_candidate_register.json. No legal review has been done.
10. **Blocked or open:** graph-wide analysis (G), an independent verifier for runtime decisions (J), runtime crash
    recovery (K), external review, legal review and production (P).

## Reproduce

`bash artifacts/evolution15/EVOLUTION15_REPRODUCE.sh`

## Continuation plan

1. Durable runtime state on Evolution 14's state service. 2. A zero-import decision-log verifier. 3. A LangGraph/MCP
adapter and an AgentLAB run with a real model. 4. Fix the F15-06 optional-pin sites. 5. External red team.
