prereg(H_9988): does the recombination wall move with DATA DIVERSITY? — exposure held FIXED, corpus scaled 500x on a public natural corpus

[dancinlife]

TRIGGERED BY MEASUREMENT, not by preference. H_9987's ladder ran 9/9 and the recombination wall did
not move: rho-weave best_distinct read 0-1 against a bar of >=2 at every rung, and pressing the
capacity pressure 10x harder (0.28x -> 0.03x) produced no trend. Fluency passed 9/9, so the
FLUENCY-BOUND escape is closed — the models could speak; they could not recombine. The ratio
isomorph is dead, which leaves exactly the hypothesis H_9984 pre-committed as its successor:
"not ratio but DIVERSITY, and that is bought with scale money."

This is also the one axis the repo's own root rules flag as never opened: natural corpus at LLM data
scale. Parameter scale was measured as an amplifier-not-lever, but on the SAME small corpora, so it
never bounded the data axis.

THE DESIGN'S ONE IDEA: HOLD EXPOSURE FIXED. Growing a corpus normally moves quantity and diversity
together, and then a positive cannot be attributed. Fixing total training bytes (steps) across every
rung means a larger corpus is simply seen with FEWER repeats, so the only moving variable is how
much unique content exists:

D0 60 MB (en_general, the baseline) 1x
D1 600 MB 10x diversity, same bytes seen
D2 6 GB 100x
D3 30 GB 500x

Rungs are NESTED (D1 in D2 in D3, shards accumulated) so they differ in RANGE, not in distribution.
Model size is pinned at --d 1200 (~30M): the capacity axis was already swept by H_9987, and moving
both at once would cross-contaminate.

PROCUREMENT IS MEASURED, not assumed: fineweb-edu sample/10BT, 14 shards x 2.15GB = 28.5GB, public
natural EN (CommonCrawl educational filter), ~63 MB/s observed, 726,000 documents per shard at ~4KB
each (~2.9GB of text per shard), disk 239G -> 216G. D1 (600.0MB, 131,200 docs) is already built, so
the first rung can fire without waiting on the remaining shards.

THE HONEST COST. There is NO $0 pre-gate for unlabelled web text: --earned needs T annotated
outside the token stream and ngram-audit needs labels too. So this branch pays a training run per
rung — which is part of why the axis has stayed shut.

Five-cell table frozen before firing (DIVERSITY-OPENS · DIVERSITY-TREND · DIVERSITY-KILL ·
FLUENCY-BOUND · REVERSED); the ARCHITECTURE node lands PENDING with every numeric slot empty.

SCOPE: 30GB is the ceiling. The LLM regime (~10^12 tokens, TB-scale) is unreachable on this disk, so
no result here may be extended to "and so at LLM scale". What this opens is the GB–30GB band, i.e.
directly above the MB–GB band H_9968 closed.

no-verify-ok pre-existing doc-gate debt only: ZERO new violations (ARCH-BIG-CELL 148, unchanged),

6 new leaves max 268 chars, JSON validated before commit.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com