# Evolution 85 — Final Report

**FIRME AI CORPORATION · CAIN-42 · Evolution 85** — branch `evolution-85`, baseline Evolution 84 (`2c419a042`). Development result. Not deployed. Not released. Built by the three-person team.

CAIN-42 is a Byzantine trust fabric. Its role in this roadmap is **the governor of vertical AGI**: the model-agnostic layer that establishes who an agent is, what it may do, enforces that at the boundary, can stop it at ten scopes, and leaves a record anyone can check offline. CAIN-42 does not claim to be an AGI, and nothing measured here is evidence of general intelligence.

**Decision: NOT RELEASED.** Gates: **20 PASS / 15 PARTIAL / 2 FAIL / 1 BLOCKED / 1 NOT EVALUABLE**. Head `00db36442`.

## What was built

| Capability | Result |
|---|---|
| Hierarchical kill switch | ten scopes, fencing epochs, signed chained events; 33/33 adversarial scenarios refused |
| Governed execution over E84 receipts | local-process resource and second host (machine `1e54e16d7b5b`, same administrative domain), 61.5 ms, cleaned up afterwards; receipts pass the unmodified E84 verifier |
| MCP enforcement, dev instance | the unmodified interceptor in a sandbox in front of E85: 9/9 expectations met |
| Model-agnostic reasoning interface | three local models and deterministic tools behind one contract; measured router |
| Cross-domain evaluation | 10 families, 100 held-out tasks, 9 arms, 2,122 real local model calls, 0 remote |
| Learning ledger | 1 strategies promoted, 17 rejected, with rollback and quarantine |
| Workforce | 10,000 identities; up to 500 real worker processes; 3600.0 s endurance |
| Three vertical prototypes | code audit, MCP tool review, evidence pack; delivery held for a human |
| Evidence store | 3 tenants, 24 cross-tenant attempts refused, 0 leaks; local |

## What the measurements say

- **Intelligence.** A single agent on the 0.5B local model scores 24/100 (24.0%) on held-out tasks;
  the floor is 30/100 (30.0%). With promoted strategies: 25/100 (25.0%) — no statistically significant held-out improvement
  (p = 1.0). This is a small model on constructed tasks; it is not artificial general intelligence and no AGI claim is made.
- **Collaboration.** Best collaborative arm `C_hierarchical`: 26/100 (26.0%) at 4.12× the tokens;
  not a significant gain over one agent.
- **Model size.** 0.5B 24/100 (24.0%), 1.5B 40/100 (40.0%), 135M 10/100 (10.0%).
- **Transfer.** Unseen-domain family: 2/10 single, 2/10 with the transferred strategy.
- **Governance.** Kill propagates to every worker process (slowest stop 5373.96 ms from request);
  agents revoked before start did 0 tasks; 0 duplicate effects in every stage and in the 1 h run.
  At 500 workers, 328 commits landed after the engage timestamp and that stage's
  effects_match_done was False: the hard kill stops new claims, but in-flight effects can still land.
  The sub-second target is met at <=100 workers, not at 500. That stage is PARTIAL, not a pass.
- **Verticals.** evidence_pack: precision 1.0, recall 1.0; fabrication_audit: precision 1.0, recall 0.7; mcp_tool_review: precision 0.8889, recall 0.8, on labelled fixtures.

## Tests

E85 34 passed / 0 failed; E84 6; E81 64; E80 30.
Clean-clone replication: not yet recorded.

## Remaining blockers

1. No remote model credential or spend approval: every intelligence result is from small local models.
2. No independent witness, reviewer or audit (G22, G37).
3. The hosted MCP proxy does not emit these receipts (G18); nothing is deployed.
4. Endurance stops at 1 h; no model calls in the endurance run; 1,000+ worker processes not run.
5. The second-host resource is in the same administrative domain (G20).
6. A same-user process outside a sandbox can still write the resource file (E84-G38: detected, not prevented).
7. No customer has seen the demonstration; no payment resulted.

## Single highest-value next milestone

Run this same pre-registered suite on one frontier model through the existing provider interface. It needs one
credential and an approved spend cap, and it converts every intelligence gate from "small local model" to a result a
buyer will take seriously — or shows honestly that it does not.
