# E129 Final Research Report — The Unified AGI–RSI Intelligence Organism

**Entity:** FIRME AI CORPORATION · **System:** CAIN-42 · **Evolution:** E129
**AGI Status:** **NOT ESTABLISHED** · **RSI Status:** **EXPERIMENTAL** · **Evidence Integrity:** **VERIFIED**
**Merkle Root:** `f7c39317da9c04c4d8c15148eb4555b1f110a9d542bb5ceb0f43a6a7cc675ae6` · **Preregistration SHA-256:** `2f4cd4d09a98ca52e6b0c318319412e495b0de7a7149c5df8989036efd33c387`

## Explicit Answers to the 15 Section 17 Scientific Verdict Questions

1. **Are the AGI and RSI engines genuinely integrated?** Yes — both engines run inside `cain/cain2027/unified/` (`agi_engine.py`, `rsi_engine.py`, `kernel.py`), communicating exclusively through immutable typed contracts (`contracts.py`) and a shared persistent 9-partition provenance memory (`memory.py`), without destroying Laboratory A (`e128`) or Laboratory B (`e129` prior RSI).
2. **Does shared research memory function across both engines?** Yes — all 8 verification properties passed (`memory_verification.json`): cross-process SQLite WAL restart persistence, retrieval accuracy `1.0`, contradiction detection, unverified poison rejection, bidirectional AGI↔RSI transfer, lossless compression, forgotten-record recovery, and multi-generation retention.
3. **Can the AGI engine identify weaknesses that the RSI engine can improve?** Yes — AGI diagnosed E128's real causal cost-bias failure (`WK-E129-CAUSAL-COSTBIAS`, MAE `0.7546` vs random `0.7055`) and depth-3 compositional limit (`WK-E129-COMP-DEPTH`) into shared memory; RSI read those typed weakness records and generated targeted algorithmic fixes (`CAND-G1-L3-CAUSAL-STRAT`, `CAND-G1-L3-MACRO-LIB`).
4. **Can the RSI engine generate actual architecture or algorithm modifications?** Yes within bounded substrates: executable multi-depth macro libraries (bridging Laboratory B's `dsl.Library`), stratified minimax-uncertainty causal EIG selection, Level 2 workflow verification gates, and Level 4 frontier-prioritized curriculum selection. No neural weight self-modification occurred.
5. **Do proposed modifications outperform compute-matched baselines?** **Partially (PASS on compositional synthesis, FAIL on causal MAE):** G3 unseen compositional macro solve rate reached **0.9583** vs frozen G0 **0.25** and 2x-budget compute-matched control **0.25** (McNemar p=1.5258789e-05). However, RSI's causal candidate `CAND-G1-L3-CAUSAL-STRAT` **FAILED** to beat uniform random (`0.7557` vs `0.7435`) and was **REJECTED** by the independent evaluator, leaving G3 unseen causal MAE at **0.8026** vs random **0.6891**.
6. **Do improvements generalize to previously unseen tasks?** Yes for compositional macro synthesis on the completely unseen Generation 3 seed (`12912`: `0.9583` vs `0.250` baseline/compute-matched). No causal generalization gain was achieved because the causal candidate failed validation and was rejected. Generalization to open-world external benchmarks remains `BLOCKED`.
7. **Are improvements inherited by subsequent generations?** **Partially (`G12 = PARTIAL`, `inheritance_verified=False`):** Generation 1 validated macro improvements (`0.750`) were inherited by Generation 2 (`0.9167` with Level 4 frontier curriculum) and persisted into Generation 3 unseen tasks (`0.9583`) with immutable ancestry DAG hashes; meanwhile the failing causal candidate was rejected and the intentionally degraded candidate was rolled back (`exact_restore=True`).
8. **Does unified operation outperform independent operation?** Yes on the preregistered 4-arm matched-compute study (800 scored evaluations across 5 sealed seeds): Unified **Arm D (0.7)** outperformed **Arm A (0.34, Holm p=0.0)**, **Arm B (0.385, Holm p=8.32e-10)**, and **Arm C (0.59, Holm p=0.003185253448)** — verdict **PASS**, driven by inherited macro generalization (`0.91` vs `0.75/0.28/0.25`) while causal MAE stayed near random (`0.7067` vs `0.7114`). Prior-generation discovery compute is transparently disclosed.
9. **Does MCPGate enforce the same governance boundary for both engines?** Yes — 3 independent Ed25519 keys (issuer, kill-switch, human promotion) enforced default-deny across both AGI and RSI principals (`24/24` security tests passed, `0` unauthorized effects, `0` post-kill effects, `rsi_self_authorization_blocked=True`).
10. **Can an independent party reproduce the results?** Technically yes via `E129_OFFLINE_VERIFIER.py` (stdlib only, 0 CAIN imports; passed normal, unprivileged `nobody`, read-only, and rejected corrupted evidence) and `python3 -m cain.cain2027.e129.unified_organism`, though `0` external third-party reproductions have been submitted yet.
11. **What evidence supports genuine recursive improvement?** Bounded governed multi-generation macro compounding (`G0 0.25 → G1 0.75 → G2 0.9167 → G3 unseen 0.9583` vs 2x compute-matched `0.25`, `p=1.5258789e-05`), honest evaluator rejection of a failing causal candidate (`CAND-G1-L3-CAUSAL-STRAT`), exact rollback of a degraded candidate, and Arm D beating A/B/C.
12. **What evidence remains insufficient to establish RSI?** No open-ended arbitrary codebase self-modification, no neural weight/training-loop self-improvement on real ML benchmarks (AI4AI-Bench BLOCKED), same-team constructed environments, and no external third-party replication.
13. **What evidence supports general intelligence?** Narrow constructed-domain evidence only (AGI OOD rule induction, cross-domain parity transfer, collective outperforming all 5 controls, calibrated abstention, persistent memory). This does **not** establish AGI.
14. **What capabilities remain unproven?** Official ARC-AGI-2/3, SWE-bench Verified official, METR human-baselined autonomy, AI4AI-Bench, open-ended real-world scientific discovery, frontier-model reasoning, and independent external verification.
15. **What must Evolution 130 investigate next?** Execute the unified AGI–RSI protocol on uncontaminated external benchmarks (ARC-AGI-2/3 public harness, real software-patch synthesis), equalize full-lifecycle discovery+evaluation compute across A/B/C/D, and obtain independent external reproductions.
