RP002A second frozen confirmatory protocol

English publication copy, prepared after the completed run. The authorization quote below is translated into English for publication; its original wording remains in docs/research/CONFIRMATION_002_APPROVAL.json. Numerical and operational text is unchanged. The experiment used the original pre-outcome protocol and source-bound configuration; that file remains unchanged. A publication copy does not redefine the experiment.

Date: 2026-10-10. Run ID: confirmatory-20261010-002. Version: confirmation-2.0. Frozen before confirmatory outcomes. User authorization: fresh explicit instruction: “Try step 4 again”. Exact numerical file: CONFIRMATORY_002_CONFIG.json; full procedure and interpretation rules: STATISTICAL_PIPELINE_PROTOCOL.md. Historical CONFIG_PROPOSAL.json remains unchanged.

Exact scope

Question: for these finite, discrete, stochastic, passive, partially observed generators, when do the specified trained history-based algorithms improve next-observation prediction over a current-observation conditional-count algorithm? E0 p=.1,q=0; E1 p=.1,q=.5; E2 p=.5,q=.5. Initial hidden-bit probability .5, stationary within independent episodes. No action, policy, intervention, self-modification or privileged realized hidden-state access. Prior-art and analytic limits are retained in COLLISION_REVIEW.md and CONFIRMATORY_DESIGN_REVIEW.md.

The frozen configuration selects 20 fresh training replicates per world, 60 units and 120 learned fits. Each unit uses 4096 training, 1024 validation and 2048 held-out assessment episodes, 129 observations/128 prediction pairs. B0 counts, B1 window-eight MLP (579 parameters), B2 GRU state 12 (651 parameters), B3 observed-history Bayes filter. Training: Adam .001, batch 32, gradient clip 1, cap 300 epochs, patience 20, minimum-validation checkpoint. The exact fixed-budget procedure is the target, not an optimally trained algorithm. All trace warnings are retained.

Fresh confirmation-2.0 seeds separate every world/replicate/split and model initialization/order; no calibration, variance or earlier data/checkpoint reuse. World lexicographic, B1 then B2; reset recurrent state every episode. CPU two threads, pinned environment, one 14400-second attempt including data/training/assessment/artifacts. No retry, extension, discretionary seed replacement or optional stopping. An incomplete attempt has no complete family decisions.

Pre-outcome count selection

Synthetic calibration qualified n=20,30,50 only under its fixed fixtures. Five independent target-size exploratory replicates per world supplied the seven contrast SDs. The largest observed SD was 0.000450571; the simultaneous normal-theory upper-SD planning bound was 0.001805443, factor 4.007010. Apply the predeclared .005-nat half-width rule to the qualified counts:

Selected n=20; target met under the declared assumptions: True. Count selection uses exploratory SDs only, not effects or confirmatory observations. Five-replicate chi-square bounds assume normality and are unstable; this is a precision rationale, not an 80% power guarantee. Calibration does not prove actual-loss or distribution-free coverage. Dependency hashes bind all inputs and the selection record.

Primary decisions and limits

Equal-weight training-replicate paired episode-mean natural-log-loss differences, clip 1e-8. Seven fixed two-sided Bonferroni intervals at alpha .05: mean +/- 2*t(1-.05/14,n-1)*s/sqrt(n). Independence and normal-theory assumptions remain explicit. The preselected factor two was stress tested before the variance stage; no normality test switches methods after confirmation. Zero observed SD is an analysis failure.

E1:B1-B0 and E1:B2-B0 require an entire interval below -.01 for benefit. E0/E2 B1-B0 and B2-B0 require an entire interval strictly inside [-.01,+.01] for equivalence. E1:B2-B1 requires the benefit rule, but is solely an implementation comparison. Other outcomes are indeterminate; nonsignificance does not establish equivalence. The .01 margin is a fixture-specific engineering tolerance, about one quarter of E1's ideal .03981-nat current-to-history gain, not a universal substantive threshold. The ideal window-eight/full-history gap is ~.000012 nats: this experiment cannot establish recurrence necessity.

Secondary outputs are descriptive: multiclass Brier score, current-erasure/visible conditional log losses, fixed top-label confidence bins, clipping fractions, unclipped loss if finite and gap to B3. Empty strata/bins remain null. All raw datasets, selected weights, validation traces, original-precision predictions/losses and model-level secondary metrics are retained losslessly; saved predictions and analysis must replay without training.

The user has authorized execution, but an exact source/baseline/config approval record must be committed before launch. No automatic scientific Claim promotion or Cycle closure follows a completed benchmark.

Separately authorized technical retry

The first attempt remains INCOMPLETE in confirmatory-20261009-001; its immutable artifacts and CONFIRMATORY_001_REPORT.md are retained. That attempt's no-retry rule governed its original authorization. On 2026-10-10 the user explicitly authorized this separate second attempt. Partial outcomes from the first attempt were accessible; they are not used to choose counts, margins, algorithms, budgets or analysis. Every inferential and training parameter above matches the original freeze. The old configuration is an additional hashed dependency; only run identity, seed namespace and approval binding differ. No data, seeds, checkpoint or partial results are pooled/reused. This is a disclosed technical retry, not a claim that the earlier observations were unseen.

The repaired timer expires when either civil wall time or monotonic elapsed time reaches 14400 seconds, checked before phases, at every training batch and after validation/unit completion. timing.json retains start/end UTC and both elapsed clocks. A forward civil clock discontinuity stops conservatively; a backward adjustment cannot extend the monotonic allowance. Suspension counts against civil wall time. The command runs under /usr/bin/caffeinate -i for its lifetime to prevent idle system sleep, without persistent power-setting changes. Manual sleep/lid closure is still possible; the timer remains the stop rule. Artifact finalization on failure preserves the incomplete attempt and does not authorize further training.

Execute once from clean, source-bound approval: caffeinate -i python3 scripts/rp002a-statistical-pipeline.py confirmation --config research/RP002A/CONFIRMATORY_002_CONFIG.json --execute. The separately committed docs/research/CONFIRMATION_002_APPROVAL.json binds this exact configuration, baseline and reviewed source. No automatic third attempt, extension, seed replacement, pooling, partial-outcome inference or method switch is authorized.