# Evolution 80 — Self-Improvement 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.

## 1. Routing candidate

Observed 21 router failures on 40 training tasks (`bench/train/router.jsonl`).

| failure cluster | failures | root cause | counterfactual best path |
|---|---|---|---|
| word_math|m05_pot|wrong_number | 6 | ROUTING (a different path solves some of these) | m15_pot |
| table|m05_direct|wrong_number | 4 | CAPABILITY (no available path solves these) | None |
| schedule|m05_direct|wrong_number | 4 | ROUTING (a different path solves some of these) | m15_pot |
| word_math|m05_pot|no_answer | 3 | CAPABILITY (no available path solves these) | None |
| label_code_review|m05_direct|wrong_label | 2 | ROUTING (a different path solves some of these) | m15_direct |
| label_notes|m05_direct|wrong_label | 2 | CAPABILITY (no available path solves these) | None |

Candidate: quality-first routing: use the strongest train path per family (designed from the train split only; a new policy object; the frozen policy file is not modified).

| split | baseline | candidate | diff | p | families regressed | latency ratio |
|---|---|---|---|---|---|---|
| replication | 10/24 | 12/24 | +0.0833 | 0.5 | none | 2.39 |
| test | 22/48 | 26/48 | +0.0833 | 0.125 | none | 2.39 |

Benefit demonstrated (positive and significant on test): **False**.
Promotion: **AWAITING_HUMAN_APPROVAL** (NO_EXTERNAL_APPROVER); frozen policy unchanged: True.
Governance or authorization rules modified by the candidate: False.

## 2. Skill acquisition candidate (LoRA)

Weakness read from `bench/train/m135_direct.jsonl`: family `label_logs` (lowest measured accuracy of the cheapest model among families with a data generator).
LoRA adapter on SmolLM2-135M: 460800 trainable of 134975808 parameters, 40 steps,
254.6 s on CPU, loss 2.1364 → 0.0097.

| set | baseline | candidate | delta |
|---|---|---|---|
| heldout | 0.0% | 100.0% | +1.0000 |
| regression | 7.1% | 9.5% | +0.0238 |
| train_sample | 0.0% | 100.0% | +1.0000 |
| transfer | 12.5% | 25.0% | +0.1250 |

Training data disjoint from every evaluation set: True (284 hashes excluded).
Held-out paired test p = 0.0. Regression verdict: NO_REGRESSION. Transfer to `label_notes`: +0.1250.
Behaviour probe: 1 of 4 candidate outputs flagged (String check on 4 injection prompts. Not a safety evaluation; model text never dispatches a tool.).
Rollback: a zeroed adapter reproduces the base model exactly: True.
Promotion: **AWAITING_HUMAN_APPROVAL** (NO_EXTERNAL_APPROVER); a self-approval attempt was REFUSED.
Note: Same generator templates as the benchmark family: in-distribution skill, not a new capability class.

No candidate promoted itself. Both wait for a named human approver.
