# CAIN-42 E142 — Preregistration: fixing E141's measured failures

**FIRME AI CORPORATION · committed before any E142 code or data · 2026-11-22**
AGI: NOT ESTABLISHED · RSI: EXPERIMENTAL · runs inside the E141 engine (systemd service `cain-e141-research`)

## Diagnosis from E141 data (already observed; this is why the hypotheses below exist)

1. **calibration_shift** — shift-aware softening helped in 8/8 cells with target label noise (0.15) and hurt in every cell with pure covariate shift and no noise (contradicted 17 times). Shift magnitude alone was the wrong trigger.
2. **world_model_grid** — the discovery scheduler put 116 of 141 experiments (85%) in one cell (size 9, walls 0.3, teleporters 2, budget 400) because the failure follow-up bonus never decays. At budget 400 both explorers nearly saturate (curiosity 0.9006 vs random 0.8960 overall); at budget 120 curiosity was ahead in every cell (+0.005 to +0.022).
3. **rsi_search** — accepted lineages rose 0.48 → 0.97 best held-out by generation 4, then plateaued; E141 could not say whether inheriting the proposer's meta-parameters helped.

## Hypotheses (one confirmatory test each)

| ID | Family (new, v2) | Treatment | Primary control | Primary metric | Confirmatory regimes |
|---|---|---|---|---|---|
| H1 | scheduler (engineering) | follow-up bonus decays as 0.4/(1 + runs/3) and no cell may exceed 25% of a family's open-or-finished E142 experiments | — | max cell share per family | all E142 families |
| H2 | `calibration_iw` | importance-weighted temperature scaling (domain classifier source-cal vs unlabeled target gives weights) | plain temperature scaling | target ECE (lower) | target_noise = 0 (pure covariate shift), shift ∈ {0.5, 1.0, 1.5} |
| H3 | `world_model_frontier` | model-based frontier exploration: walk the learned model to the nearest state with an untried action | random walk (same step budget) | next-state accuracy on reachable (s, a) (higher) | budget ∈ {0.5, 1.0} × 4·size², size ∈ {9, 12} |
| H4 | `rsi_meta` | lineage that inherits and mutates proposer meta-parameters for 4 generations | lineage whose proposer is reset to the root parameters every generation (same children, same node budget) | evaluator held-out solve rate of the final candidate (higher) | root parent, k = 6 children, dev-only selection inside the worker, held-out by the evaluator |

Secondary (reported, not confirmatory): H2 in noisy regimes (target_noise ∈ {0.15, 0.3}); H3 vs the E141 count-based curiosity arm and in budget 2.0 regimes; H4 random-child control.

## Analysis plan

- **Unit of analysis for the confirmatory test = one experiment** (independent seed; instances inside an experiment are correlated and are not counted as independent).
- Experiment-level outcome: the treatment's mean is better / worse / tied versus the primary control on that experiment's instances.
- Confirmatory test per hypothesis: one-sided exact sign test over experiments in the confirmatory regimes (ties dropped), computed in log space. Holm–Bonferroni across H2, H3, H4 at α = 0.05.
- Effect size: mean experiment-level difference with a 95% interval (normal approximation; bootstrap reported alongside).
- Per-experiment paired sign tests remain in each signed verdict (as in E141) but are descriptive here.
- H1 passes if, at analysis time, no cell exceeds 25% of any E142 family's experiments.

## Stopping rule (fixed now)

Analyse at the first export after **every E142 family has ≥ 30 VERIFIED experiments in its confirmatory regimes**, or after 6 hours of service uptime from the E142 code start, whichever comes first. If a family has fewer than 30 at 6 hours, it is reported as underpowered, not as a result. No interim looks change the analysis.

## What would count as failure

- H2 fails if IW temperature scaling is not better than plain temperature scaling in pure-shift regimes (this is a plausible outcome: with a well-specified logistic model, plain temperature scaling may already be near-calibrated under covariate shift).
- H3 fails if frontier exploration does not beat random walk at low budget.
- H4 fails if inheriting lineages are not better than reset lineages. A failure means E142 still cannot show that the improver improved; RSI stays EXPERIMENTAL either way.

All outcomes, including failures, are published.
