Attributed research checkpoint
E9: at a fixed 32k vocabulary a 20% science slice leaves 0.3997 of the Polish-only arm's pooled English-science excess fertility tax for a 0.0215 Polish token cost; three separator pieces change the Polish token count by -0.000469.
Assignment, status and accepted plans remain authoritative on each experiment.
Reported state
- Reported progress
- E9 completed with three retrospective plan versions (design, Amendment 1, Amendment 2): five tokenizers trained, audited and measured on the fertility atlas; every gate passes, the determinism gate only under Amendment 2.
- Open issues
- The design, both amendments and the results were committed together. Amendment 2 was written after the determinism gate had failed, its passing re-run was not logged, and the atlas re-run the gate requires has no record. The training text skips the first 2,000 documents of the unshuffled Polish and English streams, which does not exclude E5's shuffled evaluation holdout: one Polish holdout document is fully and 43 partly contained in pl.txt, which matters wherever pl.txt trains a model scored on that holdout. Per-corpus intervals are unadjusted. Fertility only: no model was trained on the arms.
- Suggested next action
- Test whether the science-slice tokenizers' fertility gain carries over to model capability by transplanting them into a small model (open follow-up: allocation to capability).
- Access needs
- Training text, arm inputs and trainer files are restricted materials: ask the Room owner, or rebuild them at the pinned revisions. The atlas needs the gated reference tokenizers and four restricted corpora.
Exact references
Exact record · H24
d1a46254-542b-43a9-b64b-3536a37e2efdH1, the headline prediction on the science-slice arm.
Exact record · H26
6b80001d-30ca-46bf-8aa3-74928a619665H3, the separator pieces on the Polish-only arm.
Exact record · P219
fd49148b-f337-4f5a-a26d-4c8227d57f1eThe locked verdict: H1 and H2 both confirmed for the science-slice arm.
Exact record · P214
07fd3382-a184-4c36-9047-709faa73f7d6H2: pooled cost below 0.05 and every per-corpus cost below 0.07.
Exact record · P215
1739ca18-1299-45d9-9178-d0042dc8ccbaWhitespace arm's Polish cost, read by the verdict only if the science-slice arm fails.
Exact record · P222
c9f4171b-b3e5-4e73-96ba-c3776d52e2fePer-corpus secondary endpoint of H1, without a bin.
Exact record · P217
8a4b7bae-a4a9-406e-ab8f-3ef8d0766aafH3 criterion (a): one token per character between digits.
Exact record · P218
493bb320-c790-443f-8ce9-b03ff790cd8fH3 criterion (b): pooled Polish cost below 0.005.
Exact record · P226
23b1d16c-84ca-464f-a2d6-421d0f9e7f26Informative check of the Polish-only arm against APT4.
Exact record · P232
2a78e81e-96a3-4141-aeb3-c5453c325fb0Instrument check: determinism gate G3 after Amendment 2.
Exact record · P233
4a8e6714-d6e6-4792-aef5-94a67c29090eInstrument check: the atlas reproduces the reference values exactly.
Experiment · E9
Open experiment →Accepted plan
Exact plan →Accepted plan
Exact plan →Accepted plan
Exact plan →