Attributed research checkpoint
E6: a 30% math and code share in recovery pretraining of an APT4 transplant closes less than 80% of the math_clean and digit-probe gaps at 0.5B and at 1.5B; its pooled Polish cost at 0.5B is 0.0187 bits per byte.
Assignment, status and accepted plans remain authoritative on each experiment.
Reported state
- Reported progress
- E6 completed. Plan versions: the design locked on 8 July 2026 (prospective), Amendment 1 adding a 1.5B confirmation pair (prospective, decided after the 0.5B result), and the confirmation analysis as run (retrospective). Five training runs, per-document scoring of the final checkpoints and base models, two analyses.
- Open issues
- The 1.5B pooled Polish cost comes from an unpaired bootstrap. The Holm adjustment, the no-dose-effect clause and the monotonicity of the rescue fraction were never computed. The 1.5B digits clause bins REFUTED although its interval crosses 0.5. One training run per arm. Dataset and base-model revisions were not recorded; the pinned revisions are inferred and OpenWebMath has none. G1 compared only the base model across backends.
- Suggested next action
- Recompute the 1.5B pooled Polish cost with a paired bootstrap from the shipped per-document scores (follow-up).
- Access needs
- The APT4 tokenizer repository is gated. The five final checkpoints, the English evaluation holdout and four evaluation corpora are restricted materials: ask the Room owner, or rebuild them at the pinned revisions.
Exact references
Exact record · H16
bc2c559b-2e70-44d8-87ff-a07a94962628H3, reported without a decision rule: recovery order.
Exact record · H17
8fb0835e-9768-4edc-b293-0d8d150d2fe8H3, reported without a decision rule: dose response of the rescue fraction.
Exact record · P136
8d38c22e-e0b7-4a12-bfb9-d0005b167458H1 math_clean clause at 0.5B bins REFUTED: R at most 0.5 and its 95% interval below 0.8 (upper bound 0.4981); H1 overall REFUTED with the digits clause.
Exact record · P137
d4a72d5f-d463-4ccf-9284-4197271f39bfH1 digits clause at 0.5B bins REFUTED: R at most 0.5 and its 95% interval below 0.8; H1 overall REFUTED with the math_clean clause.
Exact record · P138
8df69c78-d45c-441b-a44e-cd87b14448e7H2 at 0.5B bins CONFIRMED: the 95% interval's upper bound is at most 0.05 bits per byte.
Exact record · P139
59bfc2e8-6b8c-487d-8850-7c7c9e041de2H3 recovery order, reported without bins: formal texts reach 90% of their final recovery no later than the Polish and English texts, at 100M-token resolution.
Exact record · P152
054a72c5-ae6c-4e53-a947-2f04403cfac5H3 dose response, reported without bins: on the formal compression texts lower bits per byte at each higher share implies a rescue fraction increasing with the share.
Exact record · P154
28d77606-f98b-43eb-9513-9f99a7935f24H3 dose response, reported without bins: digit-probe accuracy of the 10% arm is above the 30% arm's (0.1678 against 0.1644), with no test of the difference.
Exact record · P164
7eb877af-9ccc-4caa-ba55-b90d5716242eH1 math_clean clause at 1.5B bins PARTIAL under the rule of Amendment 1: R below 0.8, and its 95% interval's lower bound (0.4996) is not above 0.5 while R exceeds 0.5; with the digits clause REFUTED, H1 at 1.5B bins PARTIAL.
Exact record · P165
09b0327f-40c2-43fb-b6a9-9bbb6fa432daH1 digits clause at 1.5B bins REFUTED under the code rule (R at most 0.5, 95% interval upper bound below 0.8), although the interval crosses 0.5; H1 at 1.5B bins PARTIAL overall.
Exact record · P166
e04b820f-352f-465c-b9a4-c9343faf6e28H2 at 1.5B bins PARTIAL as reported: the interval straddles 0.05; it comes from an unpaired bootstrap.
Exact record · P167
04457caa-cc84-4ccd-ac4d-bed74fdb2319Scale contrast of the 0% arm's gap on math_clean.
Exact record · P170
9ed8fe48-6dee-47e8-b425-086dd4dfa861Scale contrast of the rescue fraction on the digit probe; neither scale reaches 0.8.
Experiment · E6
Open experiment →Accepted plan
Exact plan →Accepted plan
Exact plan →Accepted plan
Exact plan →