# Evolution 86 — AGI Infrastructure Foundation

**Decision: NOT RELEASED.**
**Branch:** `evolution-86` · **Baseline:** E85 `06da643ff` · **E84:** `2c419a042`
**Tests:** 40 E85/E84 baseline passed; **10 E86 tests passed**
**Evidence:** **14/14** offline checks; **4/4** tamper mutations rejected; Merkle
`see MANIFEST.json`
**Gates:** **30 PASS / 2 PARTIAL / 1 NOT EVALUABLE / 1 BLOCKED**

CAIN-42 is the governed infrastructure layer for autonomous intelligence. E86 builds the
foundation planes; it does **not** claim that AGI has been achieved, that a third party audited it,
or that it is deployed.

## What was built

| Plane | Implementation | Measured evidence |
|---|---|---|
| A. Intelligence runtime | `providers.py` — 3 local model adapters, deterministic tool, remote registered but BLOCKED | 3 local providers healthy; 0 remote calls |
| B. Mission intelligence | `mission.py` — durable states, evidence-gated completion | mission `COMPLETED_VERIFIED` only after external evidence; event chain valid |
| C. Memory | `memory.py` — 8 types, provenance, verification, tenant principal, poison quarantine, hash chain | 0→20/20 retrieval; poison 0; cross-tenant 0; authority grants 0 |
| C. Learning | `learning.py` — propose, validate, external approve, promote, rollback | promotion and rollback measured; proposer cannot approve |
| D. Workforce | `workforce.py` — capability assignment, delegation, revocation, OS process probe | 10/50/100 real processes (160 total); revoked child refused |
| E. Verticals | `verticals.py` — software repair, security review, evidence pack | 3/3 objectively verified |
| F. Governance/evidence | `gates.py`, signed bundle, standalone verifier | 14/14 checks, 4/4 tamper rejected |
| Interop | `interop.py` — external agent 10-step flow | register→authority→mission→execute→receipt→verify→complete→revoke |

## What was measured

- **Cross-domain evaluation:** 20 held-out tasks, 10 families, **199 real local model calls, 0 remote**.
  - Floor 4/20 (20%), single 5/20 (25%), learned 5/20 (25%), specialist+verifier 6/20 (30%),
    hierarchical 7/20 (35%).
  - Hierarchical vs single: +2 tasks, exact McNemar p=0.7266 — **not significant**; 4.16× tokens.
  - **No intelligence breakthrough is claimed.** This reproduces E85's negative result at smaller
    scale.
- **Memory:** without memory 0/20; with typed retrieval 20/20; poisoned item quarantined;
  cross-tenant read/write refused; every hit `grants_authority:false`; chain valid.
  This measures memory plumbing, not model learning.
- **Learning:** one candidate promoted on a measured validation gain by an external approver, then
  rolled back by monitoring; proposer cannot self-approve; event chain valid.
- **Workforce:** 10/50/100 real OS processes, 160 total, unique ids; revoked agent refused;
  delegation expansion refused; budget exhaustion refused.
- **Verticals:** software defect (failing test → patch → passing independent test), security review
  (precision 1.0, recall 1.0 on one fixture), evidence pack (tamper detected).
- **Interop:** a local external agent registered, obtained a signed grant, submitted a mission,
  executed one sandbox write, produced a valid receipt, submitted evidence, completed verified, was
  refused a different tool, and was revoked.

## What did not improve

- No general intelligence: small local models, constructed tasks, 25% single accuracy.
- No statistically significant learning or collaboration gain.
- Kill switch at 500 workers remains **PARTIAL** (E85: 328 post-engagement commits; in-flight
  effects still land). E86 does not claim an immediate global stop.
- Budget enforcement is covered by unit tests but has no separate artifact: **PARTIAL**.
- No remote model: no credential and no spend approval: **BLOCKED**.
- No independent witness, external reproduction or third-party audit: **NOT EVALUABLE**.
- No production deployment: **BLOCKED**.

## Gates

| Status | Count |
|---|---|
| PASS | 30 |
| PARTIAL | 2 (G05 budget artifact, G24 kill-switch in-flight) |
| NOT EVALUABLE | 1 (G33 independent verification) |
| BLOCKED | 1 (G34 production) |

Full matrix: `artifacts/e86/acceptance_matrix.json`.

## Evidence and reproduction

```bash
python3 -m pytest tests/test_e86_foundation.py -q          # 10 passed
python3 scripts/evolution86/verify_e86_offline.py artifacts/e86/bundle   # 14/14
```

The verifier imports no CAIN code (stdlib + cryptography). The bundle contains the raw eval rows,
memory/learning/mission/interop/vertical artifacts, the gate matrix, tests and source.

## Honest limitations

- Same-project tests, no third-party audit or independent witness.
- Self-generated Ed25519 key; no out-of-band publication.
- Evaluation uses 20 constructed held-out tasks and 0.5B/1.5B/135M local models.
- Memory retrieval is lexical and deterministic; not model learning.
- Vertical fixtures are small and synthetic.
- Interop is local; not a production MCPGate/CAIN Studio integration.
- No hardware attestation, no TEE.
- The E85 500-worker kill switch is not an immediate stop of in-flight effects.

## Single highest-value next milestone

Run the same pre-registered suite on **one approved frontier model** through the E86 provider
interface, under a hard dollar ceiling, and publish the paired result against these local baselines.
That is the next honest test of whether the intelligence layer improves beyond small-model noise.
