# E114 research strategy

The optimisation target is **verified capability per unit of compute**, not module or agent count. Every cycle:
1. The director turns measured limitations into backlog items: limitation, evidence, approach, experiment, resources, success, failure.
2. Items are ranked by importance × uncertainty / cost − risk.
3. The top items run as governed experiments on held-out and sealed splits.
4. The independent verifier recomputes the claim. Memory keeps only validated results. Negative results are kept.

## Ranked backlog after this run
1. **B3** (score 0.81): no model-backed collective result inside E114 — *experiment:* re-verify E113 gate from raw rows with the E114 independent verifier; *success:* E113 claims reproduce
2. **B2** (score 0.51): skills hurt on shifted task families — *experiment:* gated vs ungated library on shift and sealed; *success:* no shift loss, sealed gain retained
3. **B5** (score 0.28): continual learners show no forward transfer — *experiment:* trunk+heads vs experts, sealed seeds; *success:* zero-shot > 0.55 with forgetting <= 0.05
4. **B1** (score 0.22): world models degrade under out-of-class shift — *experiment:* grammar-growth arm vs fixed grammar on sealed shift worlds; *success:* shift Brier of causal <= mlp with in-class Brier unchanged
5. **B4** (score 0.13): external memory benchmarks not run (AMA-Bench, LongMemEval, LoCoMo) — *experiment:* AMA-Bench synthetic subset with official scorer; *success:* official-protocol score reported with version
6. **B6** (score -0.0067): missions run minutes, not 24 h — *experiment:* 24 h governed mission with scheduled kills; *success:* continuity checks hold after >=3 kills over 24 h

Cycle 2 executed B2 and B1. Both failed their success criteria. They stay open, now with sharper evidence:
- B2: gating by I/O signature does not separate shift families; signatures need to be richer, or skills need per-family validation.
- B1: grammar growth recovers part of the out-of-class error, but the learned model is better. The next attempt should switch to the learned model on detected misspecification, not grow the grammar.
