Attributed research checkpoint
E7: random embedding initialisation damages formal domains most and its gap to FVT remains after embeddings-only training; the FOCUS advantage over FVT is small, Polish-weighted and mostly trains away.
Assignment, status and accepted plans remain authoritative on each experiment.
Reported state
- Reported progress
- E7 completed: stage 1 under a retrospective plan, an erratum re-score under a retrospective amendment whose formal-damage finding corrects the earlier one, and stage 2 under a prospective amendment.
- Open issues
- The stage-1 design was committed together with its results. The design's bootstrap intervals on init damage and pooled means were not computed, the sign test for the FVT−FOCUS gap was implemented as a comparison of means, and the double-build determinism gate has no recorded result. The auxiliary fastText training is not bit-reproducible. The stage-2 analysis script was committed after the FVT and FOCUS 150M checkpoints were scored. Trainer logs, smoke-gate results, transfer digests and the packed token stream's digest were not kept. Init damage compares transplants with the base model at 2,048-token windows that cover different byte spans.
- Suggested next action
- Open follow-up: re-score the transplants and the stage-2 checkpoints with byte-anchored evaluation windows.
- Access needs
- The APT4 tokenizer is gated on Hugging Face. Transplant models, stage-2 checkpoints, the auxiliary corpus, the Polish training text and the English web holdout are restricted materials: ask the Room owner, or rebuild them by script at the pinned revisions; the packed token stream has no recorded digest.
Exact references
Exact record · P175
87d352dc-dabc-4bcc-8c74-2bf51e2e548dFormal-damage prediction: both ratios exceed 1.15.
Exact record · P172
eb691982-846b-4914-80dc-c77e54a5aec6Formal-damage prediction on the rows scored before the erratum: the FVT ratio does not exceed 1.15.
Exact record · P176
b374215c-a199-4556-b9c2-886982995791Ranking prediction: the ordering holds on at least 7 of 9 domains.
Exact record · P178
0adfc8c2-f578-446f-ba67-8961f006561dFVT−FOCUS gap prediction: the formal mean is below the prose mean.
Exact record · P184
04840a8c-4221-480c-be08-43e29379ce35Primary stage-2 endpoint; no direction was pre-registered.
Exact record · P192
0993c95c-e538-42eb-972d-9fbb21b27ab0Secondary stage-2 endpoint; no direction was pre-registered.
Experiment · E7
Open experiment →Accepted plan
Exact plan →Accepted plan
Exact plan →Accepted plan
Exact plan →