# CAIN-42 UNIFIED AGI–RSI INTELLIGENCE ORGANISM — APEX EVOLUTION REPORT

**Entity:** FIRME AI CORPORATION · **System:** CAIN-42 · **Source Commit:** `15bc7d25f`
**AGI Status:** **NOT ESTABLISHED** · **RSI Status:** **EXPERIMENTAL** · **Evidence Integrity:** **VERIFIED**
**Merkle Root (1,200 3-domain 4-arm evals):** `52592af27c5e16d6cf7fa9bd1d3da2467c98b29c012cd0e3b03254846f3f8e8d` · **Wall Clock:** `972.22s` · **Paid Spend:** `$0.00`

## Executive Summary of All 5 APEX Breakthroughs

| # | Breakthrough | Measured Outcome | Key Control / Statistical Test | Verdict |
|---|---|---|---|---|
| 1 | **Bayesian Mixture Deconfounded A-Optimal Causal Engine** (200 worlds, 6 & 8 levers) | APEX MAE **`0.4119`** | vs naive random `0.7709` (`Holm p=0.0`), vs estimator-matched random_mixture `0.4632` (`Holm p=0.000609259728`), vs E128 `0.7832` | **PASS** |
| 2 | **AST-Verified Sandboxed Python Code-RSI** (`qwen2.5-coder:1.5b`) | Unified Code-RSI **`0.5`** | vs G0 zero-shot `0.375` & blind best-of-2 `0.375`; AST escape attacks blocked `4/4` | **PARTIAL** |
| 3 | **6-Generation Compounding (`G0 → G5`) + Rollback** | Unseen G5 **`0.875`** (depth `8`) | vs frozen G0 `0.25` & 2× compute-matched `0.25` (`p=6.1035156e-05`); rollback exact=`True` | **PASS** |
| 4 | **Full-Lifecycle Compute-Equalized 3-Domain 4-Arm Unification** (`1200` evals, 5 seeds) | **Arm D `0.8133`** | vs Arm A `0.14`, Arm B `0.3667`, Arm C `0.3433`; Holm p=`{'D_vs_A': 0.0, 'D_vs_B': 0.0, 'D_vs_C': 0.0}` | **PASS** |
| 5 | **Neuro-Symbolic Hybrid + 10-Vector Adversarial Immune System** (`qwen2.5-coder:1.5b`) | Hybrid **`1.0`** vs Raw `0.0` | OOD abstain `1.0`, Brier `0.0144`; Immune `10/10` blocked; MCPGate `0` unauth / `0` post-kill | **PASS** |

## 1. Causal Breakthrough Detail (Solving E128/E129 Active-vs-Random)

- **Mean Intervention MAE (200 sealed 6-lever & 8-lever confounded heterogeneous-cost SCMs):**
  - `passive`: `0.8787`
  - `e128_eig_div_cost`: `0.7832`
  - `random_naive`: `0.7709`
  - `random_mixture` (same Bayesian mixture estimator, uniform random choice): `0.4632`
  - **`apex_bayesian_mixture_a_optimal`**: **`0.4119`**
- **Counterfactual MAE:** APEX `0.5606` vs passive `1.3355` (Holm p=`0.0`).
- **Paired Wins vs Estimator-Matched Random:** `129` wins vs `71` losses (sign-test p=`4.9729735e-05`).

## 2. Full-Lifecycle Compute-Equalized 3-Domain Unification Breakdown (1,200 Scored Evaluations)

| Arm | Composite Success | 95% CI | Macro Synthesis | Causal (MAE ≤ 0.45) | Causal Mean MAE | 2D Shift World Model |
|---|---:|---|---:|---:|---:|---:|
| **Arm A** | **0.14** | [0.1053, 0.1838] | 0.25 | 0.1 | 0.8081 | 0.07 |
| **Arm B** | **0.3667** | [0.3141, 0.4226] | 0.34 | 0.57 | 0.443 | 0.19 |
| **Arm C** | **0.3433** | [0.2919, 0.3987] | 0.79 | 0.15 | 0.7629 | 0.09 |
| **Arm D** | **0.8133** | [0.7654, 0.8534] | 0.88 | 0.71 | 0.381 | 0.85 |

## 3. Honest Limitations (Never Hidden)

- **AGI Status remains NOT ESTABLISHED; RSI remains EXPERIMENTAL.**
- All multi-domain benchmarks (compositional macros, 6/8-lever SCMs, 2D shift grids, 8 algorithmic Python tasks) are constructed/internal; official ARC-AGI-2/3, SWE-bench Verified, METR, and AI4AI-Bench remain `BLOCKED_UNAVAILABLE`.
- Hardware is CPU-only (`8 vCPUs`, no GPU, `$0.00` paid spend); local neural calls used `qwen2.5-coder:1.5b`.
- Zero independent third-party external reproductions have been submitted yet; Hugging Face Space remains unpushed.
