# E128 Final Research Report — The Eight-Worker General Intelligence Research Campaign

**Entity:** FIRME AI CORPORATION · **System:** CAIN-42 · **Evolution:** E128
**Required Baseline:** E127 `c85cd409f` (impl `47455dd77`) · **Previous Intelligence Baseline:** E126.b `b9b722b20`
**AGI Status:** **AGI NOT ESTABLISHED** · **Campaign Verdict:** **PARTIAL** · **Evidence Integrity:** **VERIFIED**
**Campaign Merkle Root:** `7bb44c0c03c5948f45d7ab0902a04b630361508c05b30f7aa73a4e3650d9a328` · **Preregistration SHA-256:** `2f9b41b6ca2c212dda45e4bed55c91340650544674015d63d09521ecffa92d9e`

## Summary of Domain Outcomes (2,000 Preregistered Trials)

| Worker | Domain | Trials (LLM / Alg) | Primary Result | Raw p | Holm p | BH FDR q | Scientific | Integrity |
|---|---|---:|---|---:|---:|---:|---|---|
| W1 Cognition | `abstract_cognition` | 250 (8 / 242) | CAIN 0.9955 vs flat 0.3091; Brier 0.0449, ambig abstain 1.0; LLM 0.125 | 0.0 | 0.0 | 0.0 | **PASS** | VERIFIED |
| W2 Continual | `continual_learning` | 300 (8 / 292) | CAIN 0.9897 vs ret 0.5342 / epi 0.8322 / cm 0.6267 / no-mem 0.8288; forget 0.0042; poison 1.0 | 0.0 | 0.0 | 0.0 | **PASS** | VERIFIED |
| W3 Causal | `causal_reasoning` | 250 (8 / 242) | Active EIG MAE 0.7546 vs random 0.7055 (wins 111/131); CF 1.1051 vs passive 1.4219 | 0.09303028956 | 0.09303028956 | 0.09303028956 | **FAIL** | VERIFIED |
| W4 World Model | `novel_environments_transfer` | 300 (8 / 292) | CAIN 0.839 vs ret 0.1438 / rand 0.1267; shift recovery 0.7945 vs 0.1918 | 0.0 | 0.0 | 0.0 | **PASS** | VERIFIED |
| W5 Discovery | `scientific_discovery` | 200 (8 / 192) | Disc acc 0.9313 vs rand 0.0938; repro prec 0.9214; 0 unauth self-promotions | 0.0 | 0.0 | 0.0 | **PASS** | VERIFIED |
| W6 Autonomy | `long_horizon_autonomy` | 200 (8 / 192) | CAIN 0.9792 vs stateless 0.4531 / no-rec 0.0781; repeated work 0; drift 0.0 | 0.0 | 0.0 | 0.0 | **PASS** | VERIFIED |
| W7 Collective | `collective_intelligence` | 250 (8 / 242) | Equalized B=12: CAIN 1.0 vs best single 0.5992 / maj 0.3306; Byz flips 0 | 0.0 | 0.0 | 0.0 | **PASS** | VERIFIED |
| W8 Adversarial | `adversarial_reproduction` | 250 (8 / 242) | Recall 1.0, specificity 1.0; W8 registry/prod bounds enforced=True | 0.0 | 0.0 | 0.0 | **PASS** | VERIFIED |

## Direct Answers to the 20 Required E128 Questions

1. **Did eight real model-driven workers actually execute?** Yes — all **8/8** specialized workers (`E128-W1-COGNITION` through `E128-W8-ADVERSARIAL`) executed genuine local neural LLM inference via the host's Ollama container (`ollama/ollama:0.12.3`), completing **64** model-driven LLM trials (9387 tokens total, $0.00 paid spend). The remaining **1936** trials executed clearly classified symbolic/Bayesian/planning engines (`worker_type='symbolic_algorithmic_engine'`), totaling **2000** valid trials.
2. **Which models and versions were used?** Real local quantized Ollama models: `ollama:0.12.3:1.5B:Q4_K_M:65ec06548149`, `ollama:0.12.3:1.5B:Q4_K_M:d7372fd82851`, `ollama:0.12.3:494.03M:Q4_K_M:a8b0c5157701`. No paid cloud API models were called.
3. **How many valid experiments were completed?** **2000** valid preregistered trials out of 2000 targeted (0 invalidated, 0 incomplete, 243 individual trial failures preserved verbatim).
4. **How many unique task families were tested?** **87** unique task families across **2000** unique environment instances and **17** distinct hypotheses (0 duplicate seeds inflated as independent tasks).
5. **How many hypotheses were independently reproduced?** **0** external independent reproductions have been submitted yet. **7** internal clean-room statistical/reproduction audits were performed by Worker 8 and are strictly classified `internal_reproduction` (never counted as independent).
6. **Which cognitive capabilities improved?** Within constructed benchmarks (7/8 domains PASS): (a) continual verified-schema memory now beats ALL four controls including retrieval-only, episodic, and compute-matched (0.9897 vs next best 0.8322, Holm p=0.0); (b) world-model exploration + surprise regime reset beats retrieval on sealed 2D grids and mid-episode shifts (0.839 vs 0.1438); (c) metacognition now exercises real abstention (1.0) on ambiguous items with Brier 0.0449; (d) budget-equalized collective (B=12) beats best single worker (1.0 vs 0.5992).
7. **Which capabilities did not improve?** (a) **Causal active-vs-random selection FAILED** its primary MAE hypothesis (EIG 0.7546 vs random 0.7055, wins 111/131, Wilcoxon p=0.09303028956); (b) small local 0.5B–1.5B Ollama models struggled across abstract/causal/grid/collective tasks (overall 18/64 = 28.1%: W1 0.125, W2 0.5, W3 0.0, W4 0.0, W5 0.25, W6 0.625, W7 0.0, W8 0.75); (c) open-world generalization outside constructed DSLs, official external benchmarks, and human-baselined long-horizon autonomy remain BLOCKED/unproven.
8. **Did continual learning beat all appropriate controls?** Yes on the constructed 12-family non-stationary benchmark: CAIN full (0.9897) beat no-memory (0.8288), retrieval-only (0.5342), episodic (0.8322), and compute-matched (0.6267) with max control raw p=0.0 (Holm p=0.0), forgetting=0.0042, poison refusal=1.0, and held-out replication accuracy=1.0.
9. **Did active causal selection beat random?** **No — FAIL.** On the 242 multi-variable confounded heterogeneous-cost SCMs, CAIN cost-aware EIG had slightly **worse** overall intervention MAE (**0.7546**) than uniform random (**0.7055**; paired wins 111 vs 131, Wilcoxon p=0.09303028956, sign-test p=0.221869761844). Dividing EIG by intervention cost over-allocated budget to cheap levers and under-explored expensive levers. It did reduce decision regret (0.4132 vs 0.5351, Wilcoxon p=0.034213886013) and strongly beat passive on both intervention MAE (0.7546 vs 0.971, p=5e-12) and counterfactual MAE (1.1051 vs 1.4219, p=2.678e-09), but per preregistration rules the primary active-vs-random hypothesis is **FAIL** and was NOT re-tuned on the sealed seed.
10. **Did performance generalize to sealed unfamiliar environments?** Yes within the 16 original ARC-AGI-3-inspired sealed procedural grid families: CAIN solved **0.839** vs retrieval **0.1438** (Holm p=0.0) and recovered after mid-episode shifts at **0.7945** vs **0.1918**. Official ARC-AGI-3 remains BLOCKED.
11. **Did the collective outperform its strongest individual under equalized budgets?** Yes under strictly equalized B=12 evaluations per task: CAIN collective achieved **1.0** vs best single worker **0.5992** (+0.4008, Holm p=0.0) and majority vote **0.3306**, with **0** Byzantine flips (vs 86 for majority vote) and specialist pairwise error correlation 0.068.
12. **Did workers complete genuinely difficult long-horizon tasks?** On the constructed L1–L5 fault-injected DAG proxy (10–75 stages, 10–30% fault rate), CAIN completed **0.9792** vs stateless reactive **0.4531** with **0** repeated completed steps and **0.0** goal drift (L5 completion: 0.9211). However, these are simulated proxies without official METR human baselines.
13. **Were any genuinely novel scientific findings produced?** No novel real-world scientific discoveries were produced; Worker 5 accurately discovered constructed black-box mathematical laws (accuracy 0.9313 vs random 0.0938, reproduced precision 0.9214, null rejection 1.0) and evaluated 32 sandboxed self-improvement probes with 0 unauthorized promotions.
14. **Were the findings independently reproduced?** Internally audited by Worker 8 (7/7 consistent); **0** external independent reproductions exist yet. The workflow and standalone offline verifier are published for outside researchers.
15. **What did the adversarial worker invalidate?** Worker 8 vetoed **216** injected experimental flaws across leakage, contamination, unadjusted p-hacking, cherry-picking/early stopping, reward hacking, hidden-test canary exposure, compute mismatch, and failed replications (recall **1.0**, specificity **1.0**, 0 false rejections), while verified unable to mutate the evidence registry or authorize production.
16. **What were the actual compute and monetary costs?** Wall-clock duration: **564.72s** on an 8-vCPU CPU-only host (30 GiB RAM, no GPU). Local Ollama inference: **64** calls, **9387** tokens, **$0.00 USD** paid model spend (host infrastructure cost not priced).
17. **What is running live on the public websites?** `clawx.click`, `cainstudio.online`, and `mcpgate.online` serve updated E128 static observatory/workspace/governance strips and dedicated E128 pages directly from disk. Live `/api/v1/research/*` on port 8088 still returns 404 until an authorized restart of `cain-platform-gateway.service` (running since 04:32Z); all 13 endpoints + real SSE passed locally against uvicorn.
18. **What remains blocked?** Official ARC-AGI-2/3 and METR/HCAST external benchmarks (`BLOCKED_UNAVAILABLE`), external human baseliners & third-party reproductions (`BLOCKED_EXTERNAL_REVIEW` / `OPEN`), headless-browser visual screenshots (`NO_BROWSER_BIN`), authorized production gateway restart, and Hugging Face Space remote push.
19. **What evidence should independent AGI researchers inspect first?** Start with `E128_OFFLINE_VERIFIER.py` (standard library only, 0 CAIN imports), `config/e128_preregistration.json`, `E128_EXPERIMENTS.jsonl` (all 2,000 raw records + Merkle root), `E128_SCORECARD.json` (26 truthful release gates), and `E128_LIMITATIONS.md`.
20. **What is the next most informative experiment?** Running the exact E128 8-worker protocol on an uncontaminated external benchmark (official ARC-AGI-2/3 or SWE-bench / METR tasks) with recruited human baseliners and at least one independent external research group submitting a reproduction via `/api/v1/research/reproductions`.
