# CAIN-42 E142 — Open-World Preregistration (OW1–OW7)

**FIRME AI CORPORATION · CAIN-42 · committed before any open-world campaign data · 2026-11-22**
AGI: NOT ESTABLISHED · RSI: EXPERIMENTAL · paid API budget $0.00 · no model calls · no weight updates

This complements `docs/unified-apex/E142_PREREGISTRATION.md` (H1–H4, commit `025499a15`), which is unchanged.

## What exists before this commit (disclosed)

- Code: `cain/cain2027/unified/e142/` (`worlds`, `agent`, `scoring`, `rsi_ow`, `engine`) and the engine hooks in `cain/cain2027/unified/e141/` (`families`, `worker`, `evaluator`, `discovery`). Tests: `tests/unified_apex/test_e142_open_world.py`.
- **Development observations already made** on public *dev* seeds (`seed_for("dev", …)`), used to fix the design. They are not results and will not be reported as results:
  - exploration strategies were close to each other (dev means 0.74–0.78 model accuracy, 100 instances each);
  - change detection helped at L6 (0.76 vs 0.39) and cost ~1–3 points on stationary levels;
  - a rule prior **without** an evidence gate hurt the target at 5–10% budgets (negative transfer); with the gate it was ≈ scratch (+0.02 at 5% for same-family configs). The gate was added because of this observation.
- Engine experiments draw fresh random 31-bit seeds per experiment (never the dev seeds). `ow_rsi` sealed instances come from the evaluator's secret salt (`seed_for("sealed", …, salt)`); the sandboxed worker cannot read it.

## Environments and levels

Ten constructed families (grid navigation, lights toggle puzzle, causal switchboard, 1-D physics, crafting, ARC-like 3×3 transforms, modular-arithmetic oracle, stack machine, lever chain with delayed consequences, two-agent coordination) × ten levels (1 new configuration · 2 new knob combination · 3 unseen primitive: permuted actions + a compound action · 4 hidden predicate goal, ~20% unsolvable · 5 one feature masked · 6 rule change at half budget · 7 larger instance · 8 two-stage goal · 9 permuted actions and features · 10 random mix of 3/5/6/8/9). The agent receives only `observe / step / reset / n_actions` and a step budget. These are constructed environments, **not** external benchmarks.

## Confirmatory hypotheses

| ID | Family | Treatment | Control | Metric (evaluator-computed) | Regimes |
|---|---|---|---|---|---|
| OW1 | `ow_explore` | frontier (model-based) | random | next-observation accuracy over all reachable (state, action) | all cells |
| OW2 | `ow_model` | full (rules + tested revision + history) | facts-only | same | all cells |
| OW3 | `ow_transfer` | gated aligned prior | scratch | target accuracy | budget_mult = 0.05 |
| OW3b | `ow_transfer` | gated aligned prior | ungated aligned prior | target accuracy | all cells |
| OW4 | `ow_goal` | full agent | e141_best (count curiosity, tabular facts) | goal completion executed by the evaluator (unsolvable instance: correct abstention) | all cells |
| OW5 | `ow_rsi` | inherit_mem final config | G0 root config | sealed score = ½ accuracy + ½ completion | all |
| OW5b | `ow_rsi` | inherit_mem | random_search (same 12 candidate evaluations) | same | all |
| OW6 | `ow_rsi` | inherit_mem | inherit_nomem | same | all |
| OW7 | `ow_rsi` | inherit_mem | reset_mem (proposer reset each generation) | same | all |

## Analysis plan (fixed now)

- **Unit = one experiment** (independent seed). Instances inside an experiment are correlated and are not counted as independent.
- Experiment-level outcome = sign of (treatment mean − control mean) over that experiment's instances; ties dropped.
- One-sided exact sign test across experiments, computed in log space (p is never displayed as 0). **Holm–Bonferroni across the nine tests above at α = 0.05.**
- Effect size: mean experiment-level difference, 95% percentile bootstrap interval (10,000 resamples, seed 142), Cohen's d_z. **Practical significance:** |mean difference| ≥ 0.02 absolute; a statistically significant effect below that is reported as "significant but small".
- Effective sample size = number of experiments in the regime (and the number of non-tied experiments used by the sign test).
- Retries count once. Replications are reported separately and are not added to the confirmatory count. REJECTED experiments (e.g. adversarial audits) are excluded and counted.

## Stopping rule (fixed now)

Analyse at the first export after **each of `ow_explore`, `ow_model`, `ow_goal`, `ow_transfer` has ≥ 30 verified experiments in its confirmatory regime and `ow_rsi` has ≥ 20**, or after **6 hours** of E142 instance uptime from the OW campaign start, whichever comes first. A hypothesis with fewer experiments at that point is reported as **underpowered**, not as a result. No interim look changes the analysis.

## Secondary (descriptive only)

Per-family and per-level breakdowns; all eight exploration strategies vs random; goal-hypothesis accuracy; abstention accuracy and calibration (ECE, Brier) of `p_solvable`; L6 revision counts; transfer types A (same family) / S (permuted interface) / C (cross-domain); compressed prior; memory arms rootcause / retrieval / compressed / xfer_mem; per-generation sealed curve G0→G3 of the treatment lineage; fresh-process re-execution of final configurations (persistence after restart).

## What counts as failure

Each hypothesis fails if its Holm-adjusted test is not significant in the predicted direction. Plausible failures we expect from dev data: OW1 (strategies were close), OW3 (gated ≈ scratch), OW6/OW7 (memory and proposer inheritance may not matter over three generations). A failure of OW7 means E142 still cannot show that the improver improved; **RSI stays EXPERIMENTAL in every outcome.** All outcomes are published.
