Round 3 (pre-registered 2026-10-04, before any round-3 result exists)

Rounds 1-2 (full/, full2/, cmpE/, cmpR/): once both sides use the same two-hot expectation (D22), every training statistic matches (entropy seed ranges overlap in 10/10 windows in both pairs; S2 and world-model losses match). The pre-registered eval statistic S1 still fails in both pairs, driven by the 20000-row checkpoint, where all 6 reference runs score >= 399 (stock 405, 466, 400; exact 470, 495, 469) against 1/6 at 18000 and 0/6 at 16000; no Ajax run shows this. Over all ten checkpoints the eval sums are ref 1871 / 1824 vs Ajax 2117 / 1892.

Hypotheses for the 20000-row concordance: - H1 end-of-run artifact on the reference side: the reference jumps at its final checkpoint, whatever the run length. - H2 correlated reference runs (D2): stock make_replay gives every reference run selectors.Uniform(seed=0), the same replay index stream, so reference runs are not independent and can share good/bad stretches. - H3 real late-learning difference: the reference improves faster than Ajax late in training.

Design: reference = 29eb964 + exact two-hot expectation (D22) + replay sampler seeded per run ([seed, 0x5EED]; D2), i.e. Ajax's documented choices except D1 (asynchrony, not removable); run_reference_round3.py. Ajax = 5cb0738 (frozen worktree), unchanged. Both: seeds 0 1 2 3 4, 24000 rows, evaluation every 400 rows (10 episodes, PROTOCOL.md section 4), all else as PROTOCOL.md.

Pre-registered statistics (per seed, then over seeds): - LATE = mean eval return over the checkpoints in [12000, 24000] (31 points). Primary. Rule: LEARNS = mean Ajax LATE >= min reference LATE; WITHIN_RANGE = min <= mean Ajax LATE <= max; plus Welch's t-test on the 5 vs 5 per-seed LATE values (two-sided p). - S1 as in round 1 (16000, 18000, 20000), for continuity, same rule. - Concordance: number of runs per side with eval >= 400 at 20000 and at 24000, and the per-side rate of eval >= 400 over all checkpoints in [16000, 24000].

Readings (decided now): - H1 if >= 4/5 reference runs score >= 400 at 24000 while the reference rate at 20000 is no higher than its [16000, 24000] rate. - H3 if Welch p < 0.05 with reference LATE > Ajax LATE. - H2 (or noise) if neither: no reference concordance at 20000 beyond its late-phase rate and no significant LATE difference. - If Ajax LATE > reference LATE with p < 0.05, that is a difference too and is investigated.

Amendment (2026-10-04, before any round-3 evaluation existed: the runs had

not reached their first 400-row checkpoint; found by a synthetic self-test)

The H1 reading above would also fire when the policy simply reaches the cap late in training (high at 24000 because learning progressed, rate at 20000 at or below the late rate). H1 is an anomaly of the final checkpoint relative to the reference's own late phase, so the reading becomes: H1 if >= 4/5 reference runs score >= 400 at 24000 AND the reference's rate of eval >= 400 over the checkpoints in [16000, 24000] is at most 0.4. Everything else is unchanged.