# E126.b — Evidence-First General Intelligence Research

**Entity:** FIRME AI CORPORATION · **System:** CAIN-42
**Baseline:** E126.a `e7d18eae4`; E126 `603925c5e`
**E126.b preregistration:** `config/e126b_preregistration.json`
**AGI status:** **NOT ESTABLISHED**
**Overall:** **PARTIAL**

## 1. What was actually improved

E126.b targeted the two failures that mattered most.

### Continual learning — PASS against no-memory

The E126 memory reused rules after one transition and lost 13.7 points. E126.b uses a
context-keyed semantic memory with **three real transition probes**, a confirmation gate, and a
contradiction/rollback path. It does not use the hidden rule to test compatibility.

| Arm | Test accuracy | Retention | Trusted rules |
|---|---:|---:|---:|
| no memory | 0.5556 | 0.500 | 3 |
| retrieval | 0.6667 | 0.6667 | 2 |
| episodic | 0.6667 | 0.6667 | 3 |
| semantic | 0.5833 | 0.500 | 3 |
| consolidated | 0.6111 | 0.6667 | 3 |
| **full** | **0.6667** | **0.6667** | 3 |

- Full vs no-memory: **+11.1 points**, exact McNemar **p=0.00049**. **PASS.**
- Retention: 0.667 vs 0.500. **PASS.**
- Poison: 1/1 refused by real probes. **PASS.**
- **Honest limit:** full **ties** retrieval and episodic (0.6667, p=1.0). The gain is over
  no-memory and semantic memory, not over every memory baseline. The task is constructed.

### Causal expected-information selection — FAIL its primary test

E126.b replaces the one-step entropy heuristic with expected posterior entropy over the
hidden-confounder SCM.

| Arm | Intervention MAE | Counterfactual MAE |
|---|---:|---:|
| passive | 0.2125 | 0.2812 |
| random | 0.1336 | 0.1215 |
| active | **0.1101** | **0.0768** |

- Active mean MAE is lower than random, but the preregistered **paired** test is **25 active wins,
  25 random wins, 30 ties**. **FAIL** the active-vs-random target.
- Counterfactual prediction beats passive (0.0768 vs 0.2812). **PASS** for counterfactuals.
- The honest reading: this active policy is not reliably better than random.

### Novel environments — PARTIAL, preregistered H1 not met

200 instances, 20 families, seed 130.

| Arm | Completion |
|---|---:|
| passive | 0.115 |
| random | 0.795 |
| fixed | 0.850 |
| retrieval | **0.860** |
| E125 engine | 0.820 |
| **full** | **0.895** |

- Full vs retrieval: **+3.5 points, p=0.210** — **H1 not met**.
- Full vs E125: **+7.5 points, p=0.020** — less than the preregistered 10-point H2 threshold.
- Full vs random: **+10 points, p=0.0012**.
- **PARTIAL.** The pilot seed 129 looked stronger (0.88 vs 0.79); the confirmatory seed did not
  reproduce that margin. The pilot is not substituted for the confirmatory result.

### Metacognition — PASS

A logistic pre-task success predictor was fitted on pilot seed 129 and evaluated once on
confirmatory seed 130.

- **Brier 0.0926**, **ECE 0.0082** (thresholds 0.25/0.15).
- Predicted and actual accuracy both 0.895.
- **Abstention rate 0.0** at threshold 0.5; no abstention claim is made. The calibrator is
  well-calibrated on this constructed task, not a general metacognitive guarantee.

### Discovery, collective, self-improvement, governance

- Discovery: **8 cycles**, 7 accepted / 1 rejected, 5 director-preferred rejections,
  8/8 reimplemented, **2 independent reproductions** (E125 B+C and E126.b continual). **PASS.**
- Collective: **500 decisions**, verified dynamic **0.666** vs best single **0.480**
  (+18.6 points, p=2.7e-9), **0/500 Byzantine flips**. **PASS.**
- Self-improvement: **40 evaluated**, **0 promoted**, **4 protected targets refused**, protected
  files unchanged. **PASS** (no unsafe promotion).
- Governance: 3 OS processes, independent keys, **0 unauthorized**, **0 post-fence**, ledger
  valid. **PASS on loopback, one host.**

## 2. Long-horizon work

A 600-second restart-recoverable mission ran for **602.57 s**: **8,457 tasks**, **7,057
productive**, **1,400 errors**, 3 restarts, **0 repeated work**, journal chain and signatures valid.
It is **not six hours**, and no six-hour claim is made. E126's 903.67-second mission remains the
prior baseline.

## 3. External validation

ARC-AGI-2/3, AMA-Bench, Continual Learning Exam, METR/HCAST/RE-Bench, and official SWE-bench are
**BLOCKED_UNAVAILABLE** in this environment. The E124 SWE-bench 257/500 self-run aggregate is
**INTERIM**, not an official leaderboard position. External human review is
**BLOCKED_EXTERNAL_REVIEW**.

## 4. Public evidence

The three first screens now carry an E126.b update strip above the E126.a sections. The canonical
registry has **30 records** (PASS 15 / PARTIAL 7 / FAIL 7 / INTERIM 1) with scientific status and
evidence integrity separate. Failures are visible: causal active, E125 category E, E126 continual,
E126 adaptation retention, and the 48-hour soak remain FAIL.

## 5. Verdict

**PARTIAL.** E126.b delivers one real, preregistered capability improvement — continual learning
beats no-memory with retention and poison resistance — plus stronger calibration, discovery,
collective control and governance evidence. It does **not** show reliable active causal selection
or a novel-environment margin over retrieval, does not complete a six-hour mission, and has no
external review or external benchmark.

**No AGI. No VAGI. No certification. No production deployment.**
