# Evolution 80 — Autonomous Research 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.

One end-to-end project: question, literature review, hypotheses, design, authorization, isolated execution, raw data,
separate-process replication, statistics, conclusion.

- **Question:** Does a routing policy fitted on train beat the best FIXED execution path on unseen test tasks, and does any multi-agent architecture beat one agent at this model scale?
- **Literature review:** `docs/evolution-80/RESEARCH_REVIEW.md`. Depth: abstracts and summaries read; full papers not read; no public benchmark reproduced.
- **Design:** train 40 / test 48 / replication 24, hash-disjoint; analysis: exact two-sided McNemar on paired per-task outcomes; Wilson 95% intervals; paired bootstrap for the difference; stopping rule: no peeking: test rows are written once; no re-run with altered prompts counts as this experiment.
- **Frozen policy:** `b98b26485b018bb6b3c2a9f3225f310e8f3d93e5dbb4fbd5d341d4e7264172fb`
- **Pre-registration:** written at 1794775662.477; first test or replication call at 1794775678.4; written first: **True**.
- **Authorization:** every model and sandbox call held a signed single-use gate grant.
- **Isolation:** bubblewrap sandbox for all code; local CPU inference; no network use by the experiment.

## Verdicts

| hypothesis | supported | detail |
|---|---|---|
| H1_router_beats_best_fixed_path | False | {"a_correct": 19, "b_correct": 22, "bootstrap_95": [-0.1042, 0.25], "diff": 0.0625, "n": 48, "p_two_sided": 0.647606} |
| H2_some_multi_agent_arm_beats_single | False | {"arms_worse_than_single": ["c07_specialist_committee"], "best_delta": 0.1042, "best_multi_agent_arm": "c08_evidence_first", "significant_arms": []} |
| H3_direction_replicates | True | {"a_correct": 8, "b_correct": 10, "bootstrap_95": [-0.2083, 0.3333], "diff": 0.0833, "n": 24, "p_two_sided": 0.774414} |

## Replication

Ran: True; separate process: True (main pid 163832, replication pid 203378);
same frozen policy: True; direction replicated: **True**.
Limit: same host, same weights, same code, same operator: a re-run on new instances, not an independent laboratory.

## Claim classification

| class | claims |
|---|---|
| HYPOTHESIS | why a given architecture helps or hurts (no causal ablation was run) |
| MODEL_GENERATED_SPECULATION | none |
| VERIFIED_FROM_RAW_ROWS | per-arm accuracy; paired differences; exact McNemar p-values |

Scientific discovery: **NOT_CLAIMED**. Statistical power: 48 test tasks and 24 replication tasks: only large effects are detectable.

## Limitations

- constructed closed-form tasks from E79 generators (same templates, unseen instances)
- three small CPU models of one family plus one tiny model
- prompt formats were fixed after a smoke run that touched a few test-split questions (format repair only; see LIMITATIONS)
- no external replication
