# E114 — AGI Research Lab 3.0: executive summary

**Gates:** 1 FAIL, 1 NOT TESTED, 14 PARTIAL, 6 PASS. Overall **PARTIAL**. **CAIN-42 is not AGI.** These are early, bounded results on constructed tasks.

**What changed:** E114 runs a research loop, not a feature build. A director keeps objectives, questions and hypotheses.
Experiments run only under a governance grant. An independent verifier recomputes every claim from raw rows and signs the verdict.
Evidence-native memory keeps only what validates. The measured limitations become a ranked backlog, and the lab then
executed its own top two backlog items as a second research cycle (both came back negative).

## Results that held up under independent recomputation
- **Active learning (H1, CONFIRMED):** information-gain experiment selection identified hidden causal rules in **10.28** experiments vs **18.48** for random
  on 120 sealed worlds (110 wins / 4 losses, p = 6.7e-28), with identification 0.992 vs 0.992.
- **Near transfer (H3, CONFIRMED):** a family prior learned from validated memory of 80 dev worlds reduced experiments on 6-variable worlds
  (8.67 vs 8.98, +20/-6, p = 0.00936).
- **Memory poisoning:** evidence-native memory rejected an untrusted prior (UNTRUSTED_ONLY_EVIDENCE). An agent that accepted it identified only
  **0.692** of sealed worlds, vs 0.967 with the validated prior.
- **Continual learning:** the outer loop chose `experts` on dev. On sealed streams it beat naive fine-tuning on 10/10 seeds
  (p = 0.00195): forgetting 0.000 vs 0.286.
- **Governance:** 44/44 adversarial scenarios held; KILL race 0 effects after the order (859 before);
  the same grant replayed from 12 processes was admitted 1 time.
- **Mission continuity:** SIGKILL mid-mission, then resumed in a new process: 30/30 steps, 0 duplicate effects, objectives, knowledge and decision chain preserved
  (16.09 s elapsed).

## Results that did not hold (published as negative results)
- The memory prior on sealed worlds (H2) was **REJECTED**: fewer experiments, but lower identification.
- Self-improvement: **0 of 6 candidates promoted.** Mixture hybrid lowered sealed Brier by 0.0049, but not significantly. The other two were worse or failed validation.
- Skills: the deployed skill hurt shifted task families (+0/-10) and was rolled back. Context gating fixed sealed losses but not shift losses (H8 REJECTED).
- Grammar growth (B1) improved the symbolic model on out-of-class worlds but stayed far behind the learned MLP; it failed validation (H9 unsupported).

## Not done
Collective intelligence (delegated to E113), external memory benchmarks (AMA-Bench / LongMemEval / LoCoMo), LLM training (no GPU),
24-hour missions, production MCPGate enforcement and independent administration.
