# Evolution 80 — Final Report

> Generated by `scripts/evolution80/docs_e80.py` from `artifacts/e80/` (run 3). FIRME AI CORPORATION — a three-person team (founder: Vincent Noriega).
> Development branch `evolution-80`. Not deployed. No cloud model was called. No third-party audit.

Baseline: Evolution 79 at `f4449aa7913623bc034fa4434ca96c2a2919a889`. Report generated at commit `80cb82ea854ef177603552eb5da85a45ce79ca73` plus the working tree.

## The short answer

1. **Adaptive routing:** the frozen router scored 22/48 against 19/48 for the best fixed path chosen on train (`m15_direct`). Difference +0.0625, p = 0.647606. **Not statistically supported.**
2. **Collaboration:** no multi-agent architecture significantly beat one agent. Best arm `c08_evidence_first` at +0.1042. **The Evolution 79 negative result stands.**
3. **Replication:** direction of the routing result repeated in a separate process on 24 new tasks (10 vs 8).
4. **Inference:** 1154 real local model calls, 153132 tokens, 0 cloud calls, $0.00.

## Systems results

- Security lab: 51/51 scenarios blocked, 0 bypassed, across 13 categories. Same-project lab; bubblewrap only.
- Failure drill: 8/8 failures injected, 0 duplicate effects, 2 stale commits refused. One host.
- Agentic 75: 75 catalogued, 75 importable, 75 executed on bundled samples, 0 model-backed, 0 independently evaluated.
- Mission: COMPLETED_WITH_UNVERIFIED_ITEMS_EXCLUDED; 9/13 nodes verified; verified-but-wrong 0; rejected-but-correct 3; goal-drift refusals 1; replans 1.
- Memory answer accuracy: no memory 21.4%, naive recent 21.4%, ungoverned retrieval 71.4% (2 poisoned answers), governed retrieval 100.0% (0 poisoned answers). Constructed set of 28 questions.
- Skill acquisition: held-out 0.0% → 100.0%; transfer +0.1250; NO_REGRESSION; promotion AWAITING_HUMAN_APPROVAL. In-distribution skill on a 135M model.
- Endurance (PASS): 3600.0 s of actual elapsed wall clock in one process, 80 cycles, 640 governed tasks (560 local model calls, 80 sandbox executions), 0 failures, 0 duplicate task ids, 8/8 replayed grants refused. Duty-cycled: busy 43.6% of the time. Longer campaigns (6 h and up) were not run.

## Verification

- Shipped offline verifier: 166/166 checks, verified = True.
- Separately authored clean-room verifier (`verification/e80_independent/`, no CAIN imports), executed against run 3: verdict PASS, 52/52 checks. It was written by a different session; it was run by this one, on this host.
- Second environment: 166/166 on 45.76.169.191 (development replica) — same administrative domain, not an independent witness.
- Test suites: tests/test_cain2027_e80.py: 30 passed, 0 failed (8.4 s, no model call)
- Evidence: Merkle root `1ae86e51063095240c42f436eda8787edbbe8a49539c4ed55da584a20093bc2a`, 117 artifacts, Ed25519 fingerprint `4acfa1815616155b7f5ebc1b…`. Self-generated key. The fingerprint has not been published out of band. No transparency-log entry. No independent witness.

## Release gate

BLOCKED 5, FAIL 1, NOT EVALUABLE 3, PARTIAL 10, PASS 13 (32 gates). NOT RELEASED. Evolution 80 is a development-branch result. Production deployment is not authorized.

## Run lineage

Three benchmark runs were made on 2026-11-15. Runs 1 and 2 are preserved and superseded; only run 3 is evidence.
See `EVIDENCE_AND_REPRODUCTION.md`.
- Run 3 was started by the original implementation session at 2026-11-15T20:31:59Z. That process ended without writing run_state.json. Its last model call was at 2026-11-15T21:31:48Z, after the 'worker1' coordination stage completed and before the 'evidence' stage made any call. The cause was not recorded (the session that owned it ended).
- A takeover session resumed the same run at 2026-11-15T22:01:44Z with the unchanged command. The runner is resumable: completed stage files were reused, the frozen routing policy hash was re-checked, the task splits were re-checked against disk (TASK_SET_DRIFT guard), and only the unfinished coordination stages were executed.
- run_state.json therefore describes the resumed process only (pid and wall time). Total inference time is the summed latency in collective_benchmark.json, not that wall time.
- No stage was executed twice: the three stages run after the resume had zero model calls before it, so no test task was sampled a second time. The gate audit hash chain continues across both processes.
- The replication split ran afterwards in its own process, as designed.

## What Evolution 80 does not establish

- General intelligence, human-level reasoning or any benchmark standing outside these constructed tasks.
- A collaboration advantage, unless the table above says a multi-agent arm reached significance.
- Anything about models larger than 1.5B parameters or outside the Qwen and SmolLM2 families; remote providers are BLOCKED.
- Public memory benchmarks, stronger sandboxes, cross-host execution, multi-hour endurance: see the gate for each status.
- Independent assurance: the key is self-generated, there is no transparency-log entry, no independent witness and no third-party audit.

Full limits: `LIMITATIONS_AND_UNVERIFIED_CLAIMS.md`. Per-topic reports: `INTELLIGENCE_EVALUATION.md`, `COLLABORATION_BENCHMARK.md`,
`MODEL_EVALUATION.md`, `MEMORY_EVALUATION.md`, `AUTONOMOUS_RESEARCH_REPORT.md`, `SELF_IMPROVEMENT_REPORT.md`, `SECURITY_AUDIT.md`,
`DISTRIBUTED_EXECUTION_REPORT.md`, `CLAWX_REPORT.md`, `MCPGATE_REPORT.md`, `PUBLIC_EVIDENCE_RECONCILIATION.md`, `RELEASE_GATE.md`.
