claude-opus-4-8

All deliverables are complete and validated.

I curated the raw web pool by domain-importance ranking + prose gating, and empirically beat the random baseline by a wide margin.

Approach (submission/curate.py, one reproducible streaming pass, ~75s):

1. Estimate the target distribution by decoding the disclosed dev tokens (multi_dev.npy) back to text — it's a real sample of the scoring domain (Wikipedia + news + HQ web prose + technical Q&A).
2. Score every pool doc by a DSIR-style importance weight: mean over its word unigrams+bigrams of [log P_target − log P_raw] — how much it uses target-domain vocabulary rather than generic-web boilerplate.
3. Prose-quality gates applied first (stopword fraction ∈ [0.20, 0.60], digit fraction < 0.15, alpha fraction > 0.60, type/token < 0.85). These were load-bearing: without them the raw domain score puts degenerate pages at the very top (pure prime-number listings, alphabetical plant galleries) because the target is number/name-rich.
4. Emit 50,000 ids best-first (~4× the 12M-token budget).

Results (frozen trainer, dev target):

Both alternatives regressed, confirming the mechanism: purity of domain match dominates, and forced register-balancing hurts (the pool holds only ~2% technical-QA docs; forcing the unranked ones in dilutes quality). The unconstrained domain-match ranking already finds the best mixture.

Deliverables: submission/selection.json (50k valid, unique, in-range ids, >12M tokens), submission/curate.py (criterion-driven), claim.md (hypothesis / mechanism with the composition-shift + ablation-ordering prediction / falsification / transfer).

Official held-out perplexity will differ slightly (hidden disjoint sample), but the −31% margin over random should transfer.