# E126.b Limitations

- **Constructed tasks.** Environment, continual-learning, causal and metacognition tasks are
  constructed; the same team wrote the environments, agents, controls and verifiers.
- **Novel-environment H1 not met.** 200 confirmatory instances: full 0.895 vs retrieval 0.860,
  p=0.210. The pilot seed 129 looked stronger; it is not substituted for the confirmatory result.
  Full vs E125 is +7.5 points, below the preregistered 10-point H2 threshold.
- **Continual learning is bounded.** Full beats no-memory (+11.1 points, p=0.00049) and retains
  more, but **ties** retrieval and episodic (0.6667). The memory is a symbolic transition table,
  not a neural memory. Acquisition has a longer horizon than test; disclosed.
- **Causal active selection failed.** Mean MAE is lower than random (0.1101 vs 0.1336) but the
  preregistered paired test is 25/25 with 30 ties. Active is not reliably better than random.
  Counterfactuals are scored against a simulator, not real outcomes.
- **Metacognition has no abstention.** Brier 0.0926 / ECE 0.0082, but abstention rate at 0.5 is
  zero. The calibrator uses five observable features on one constructed task.
- **Collective uses rule-based agents and a truth-probe oracle.** 22 agents were banned, including
  some honest ones.
- **Transfer and adaptation are E126 baselines, not re-run.** Transfer +12.5 points and 0 false
  accepts; adaptation still catastrophically forgets (control 0.375 → 0.0).
- **Mission.** A 600-second restart-recoverable run; **not six hours**. No 24/72/168-hour claim.
- **External benchmarks and external human review are BLOCKED.** No official leaderboard position.
- **Deployment.** Local pages and bundles only; live production and Hugging Face deployment were
  not performed.
- **No claims.** No AGI, VAGI, certification, SOC 2, hardware attestation, production readiness,
  customer revenue or patent.
