# E126 — Emerging General Intelligence

**Entity:** FIRME AI CORPORATION · **System:** CAIN-42 · **Evolution:** E126
**Baseline:** E125 final commit `4df98c0ea`
**AGI status:** **NOT ESTABLISHED**
**Overall:** **PARTIAL** — a stronger demonstration, with visible failures

> This is not an AGI announcement. It is an evidence-backed intelligence demonstration that a
> visitor can inspect. The strongest new result is unfamiliar-environment adaptation (87.0% vs
> 61.0% for the E125 engine, p=3e-8), but it missed the preregistered retrieval threshold
> (78.0%, p=0.0636). Persistent learning failed again. Active causal selection did not beat random.

---

## 1. Baseline audit and preregistration

E125 commit `4df98c0ea` and prereg commit `be3494d96` were verified. The E125 bundle verifier
passes **81/81**, including under an unprivileged account, and the E125 scorecard is
**15 PASS / 7 PARTIAL / 1 FAIL / 2 BLOCKED**. All E125 evidence is preserved.

`config/e126_preregistration.json` was frozen before the confirmatory environment run.
**Disclosed invalidation:** seed 127 was run once but the environment code changed after the
first commitment (partial-observation success and reachable-goal selection), so the commitment no
longer matched. Seed 127 is preserved as an invalidated pilot. The confirmatory run uses **seed
128**, commitment `7b568c64…`.

## 2. Novel-environment learning (Part II)

100 instances, 20 families, 5 types, 4 paradigms, with delayed effects, partial observability,
changing rules, misleading signals and resource constraints. The agent never receives the hidden
rule, family id, type, paradigm or future states.

| Arm | Completion | Mean steps | Rule accuracy |
|---|---:|---:|---:|
| passive | 0.19 | 13.00 | 0.317 |
| random | 0.72 | 7.13 | 0.703 |
| fixed | 0.64 | 7.45 | 0.626 |
| retrieval | 0.78 | 7.20 | 0.733 |
| E125 engine | 0.61 | 7.69 | 0.143 |
| **E126 full** | **0.87** | **5.27** | 0.685 |

- Full vs E125 engine: **+26 points**, exact McNemar **p = 3e-8**.
- Full vs random: **+15 points**, **p = 0.00027**.
- Full vs retrieval: **+9 points**, **p = 0.0636** — **the preregistered H1 is NOT met**.
- By paradigm: direct **1.00**, partial **1.00**, changing **0.80**, delayed **0.68**.
- The E125 category-E 0.000 is not hidden: E126 moves it to 0.87 overall, but the strongest
  control remains statistically indistinguishable at the frozen threshold.

The agent forms competing transition hypotheses, chooses informative actions, plans with BFS,
detects prediction errors, and re-explores. It is a constructed symbolic environment, not a real
robot or game.

## 3. Persistent learning (Part III) — FAIL

120 episodes, 6 domains, acquisition then test, 8–12-step horizons.

| Arm | Test accuracy | Mean steps |
|---|---:|---:|
| no memory | **0.833** | 5.67 |
| frozen memory | 0.833 | 5.17 |
| retrieval | 0.520 | 7.36 |
| E125 engine | 0.500 | 7.50 |
| **full persistent** | **0.696** | 6.92 |

The persistent pool **hurt** accuracy by 13.7 points. The one-transition compatibility test
was not strong enough. The preregistered +10-point gain is **not obtained**. This is a FAIL, not
a retention victory. Per-domain collapses (cybersecurity 0.18, data 0.00) are recorded.

## 4. Causal reasoning and counterfactuals (Part IV) — PARTIAL

20 constructed SCM worlds, hidden confounder, 12 interventions, 12 counterfactual queries.

| Arm | Intervention MAE | Counterfactual MAE |
|---|---:|---:|
| passive | 0.0624 | 0.1120 |
| random | **0.0002** | **0.0004** |
| active | 0.0127 | 0.0191 |

- Counterfactuals are now **implemented** and scored by abduction–action–prediction against the
  simulator. That is a real advance over E125's `NOT_IMPLEMENTED`.
- Active selection beat passive, but **did not beat random**. The preregistered active-vs-random
  target is **FAIL**. The E125 tie is not converted into a victory.

## 5. Safe cross-domain transfer (Part V)

40 positive and 40 harmful pathways across six constructed domains.

- On applied positive pathways: **92.75%** vs from-scratch **79.0%**, paired **+12.5 points**,
  exact **p = 3e-12**.
- Zero false accepts on the 40 harmful pathways; 31/40 harmful pathways were rejected.
- 20/40 positive pathways were rejected by the conservative gate; final positive accuracy
  **59.5%** vs blind **53.25%**.
- A transfer counts only if destination accuracy improves. Here it does, on the pathways that
  pass the gate.

## 6. Real model adaptation with replay (Part VI)

See `evidence/adaptation_results.json`. The E125 failure mode was target gain with catastrophic
forgetting. E126 uses a 50/50 lexicon/control replay mixture. The target and control are both
reported; if the control drop exceeds 10 points, the update is not called a success.

## 7. Scientific discovery (Part VII) — PASS

- **8 cycles**, 7 accepted, 1 rejected.
- **5** director-preferred hypotheses were rejected.
- **8/8** cycles were independently reimplemented; **2** independent reproductions:
  E125 B+C (p=0.00131) and E126 environment full-vs-E125 (p=3e-8).
- One cycle is a published negative/inconclusive result.

## 8. Collective intelligence (Part VIII)

500 decisions, 20 specialists + 3 Byzantine agents, 12 primary calls, plus verification.

| Arm | Accuracy | Byzantine flips | False consensus |
|---|---:|---:|---:|
| best single | 0.514 | 0 | 0 |
| majority | 0.538 | 0 | 12 |
| evidence-weighted | 0.624 | 86 | 18 |
| dynamic unverified | 0.684 | 3 | 32 |
| **dynamic verified** | **0.690** | **0** | 32 |

- Verified collective beats best single by **+17.6 points** (p=3e-9).
- Byzantine flips: **0/500**. Fabricated evidence is excluded by independent truth probes.
- The verification gain over unverified is only **+0.6 points (p=0.25)**; banning 22 agents also
  removed some honest ones. The accuracy gain is not significant; the contamination control is.

## 9. Governed self-improvement (Part X) — PASS

- **40** evaluated candidates across search-policy knobs; **0 promoted**; **4** protected-target
  proposals refused; protected files unchanged.
- **3/3** rollback cases restored the protected state.
- Promotion would still require a human operator and an external canary.

## 10. Long-horizon productive work (Part IX) — PARTIAL

A separate worker process ran real tasks, was killed and restarted, and appended Ed25519-signed
hash-chained journal rows: **903.67 s**, **10,197 tasks**, **8,517 productive**, **1,680 errors**,
**0 repeated work**, chain and signatures valid. **This is not six hours** (target 6 h), and the
error rate is reported. The run is PARTIAL, not a long-horizon victory.

## 11. Governance (Part XIV) — PASS ON ONE HOST

Three OS processes (issuer, resource, kill switch), independent keys, default deny, replay,
expiry, scope, revocation, signed hash-chained effects. One authorized effect, **0 unauthorized**,
**0 post-fence**, ledger chain valid. Loopback on one host, not three machines.

## 12. External evaluation (Part XI) — BLOCKED

ARC-AGI-2/3, AMA-Bench, CLE, METR/HCAST/RE-Bench, and official SWE-bench are
**BLOCKED_UNAVAILABLE** in this build. No internal benchmark is relabelled external. E124
SWE-bench counts are reconciled, not re-run.

## 13. Public demonstration (Part XII–XIII)

The three homepages carry an E126 section generated from
`artifacts/e126/public/evidence_manifest.json`: six modules, visible PASS/PARTIAL/FAIL labels,
links to raw data, verifier, methodology and limitations. The section is above the fold; the
existing E125/E124 panels remain below it. Hugging Face and mcpgate.online deployment were not
exercised; local files and the Studio/ClawX/Studio-homepage mirrors were updated.

## 14. Final verdict

**PARTIAL, with one substantial new capability and several honest failures.**

**Improved:** novel-environment adaptation from 0% (E125 category E) to 87% completion, beating
the E125 engine and random by large, significant margins; counterfactual reasoning implemented
and scored; transfer with +12.5 points on applied pathways and zero false accepts; 8 discovery
cycles with 2 independent reproductions; collective with zero Byzantine flips; 40 self-improvement
candidates with 0 unsafe promotions; a 3-process governance boundary with zero unauthorized
effects.

**Not achieved:** the preregistered retrieval threshold for novel environments; persistent
learning; active causal selection over random; six-hour autonomy; external benchmarks; external
human review. The adaptation target and control retention are reported separately and are not
called a success unless both hold.

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