# E129 Laboratory B — RSI Research Engine Findings (Levels 1–5)

**Entity:** FIRME AI CORPORATION · **RSI Status:** EXPERIMENTAL (Bounded Governed Algorithmic/Library RSI)

Builds cleanly atop Laboratory B's preserved prior RSI E129 executable macro library (`docs/evolution-129/E129_RSI_LAB_PRIOR_REPORT.md`, `clawx-site/evidence/e129-rsi-2026-11-21/`) without modifying any prior artifact.

## Multi-Generation Inheritance Results (G0 → G1 → G2 → G3)

| Generation | Sealed Macro Solve | Unseen G3 Macro Solve | Causal MAE (Unseen) | Status |
|---|---:|---:|---:|---|
| G0 Frozen Baseline | 0.25 | 0.25 | E128 `0.7901` / rand `0.6891` | FROZEN |
| Compute-Matched Control (2x budget, no inheritance) | 0.25 | 0.25 | Random `0.6891` | CONTROL |
| G1 Weakness-Driven (L1+L3 Causal + Macro) | 0.75 | — | Sealed `0.7557` | PROMOTED |
| G2 Inherited + L4 Frontier Curriculum | 0.9167 | — | Inherited | PROMOTED |
| G3 Unseen Generalization Verification | — | **0.9583** (p=1.5258789e-05) | **0.8026** vs rand 0.6891 (p=0.025709910326) | PARTIAL |

- **Rollback Verification:** Intentional degraded candidate `CAND-G1-L1-DEGRADED-PROBE` was rejected by the independent evaluator (`FAIL`) and rolled back to exact pre-candidate SHA-256 (`exact_restore=True`).
- **Honest Separation:** All RSI gains here are executable macro-library, research-policy, active-curriculum, and causal-algorithm improvements — **NOT** neural weight training and **NOT** unrestricted self-modification.
