Mixture arms: the anchors, acted on

docs/research/2026-09-27-domain-anchors.md measured the training mixture and left it unbuilt. This is the build. Full record, with every byte count, every window and every anchor:

/mnt/e43497ab-0ff2-45b4-b45f-28de3339a53e/aria_data/pretrain/mixture/MANIFEST.md

Three arms on disk as directory trees of symlinks (12–22 GB of mixture at zero bytes copied). The trainer drains files one at a time under a per-epoch shuffle, so bytes are the weight — --data <arm dir> is the whole interface. No config change, no new flag.

arm	total	what it is
`arm-a-prose`	21.86 GB	control: prose only, bytes ∝ corpus size
`arm-b-anchored`	17.91 GB	priced by 5-gram BPB × unseen rate, + code/math/genomics
`arm-c-heavy`	11.99 GB	same price, code+math doubled, genomics 12.8%

The price is one line: share ∝ g × q, the 5-gram BPB times the share of held-out contexts unseen in train. g alone cannot separate prose (every text domain is 2.25–2.76 BPB); q alone over-rewards small files. The product ranks on “hard to predict and not locally memorizable”.

What re-deriving the anchors changed

Ten of ten inherited text rows reproduce to three decimals at the inherited 2 MB window. What that window hid:

math 2.446 is a JSON container, not math. mix/math sorts its two .jsonl files (10.9 GB) before 13,708 .txt shards, so any 2 MB window is pure JSON. Same window, same size: jsonl 2.446, txt 2.757. The mixture takes the txt only. Same failure mode that killed the “books is where the language lives” claim.
books has no stable bar. 2.482 / 2.434 / 3.095 / 2.557 at 2 / 4 / 6.3 / 10 MB — one 2.15 GB parquet, so the window walks documents.
genomics is bimodal, not a number. 0.16–0.19 BPB on low-complexity windows, 2.16–2.19 on coding windows, 2.909 on held-out species. The inherited 2.010 prices the hard half only.
1.5 MB of train text is too small a sample. Every sharded text domain is 0.07–0.19 BPB pessimistic at 2 MB. The table is a valid ranking — which is all a mixture needs — and not an absolute bar.
web (103 GB) is the largest text domain on disk and was missing from the table. Measured 2.458–2.537. The real_filtered_v2 corpus actually being trained on reads 2.849 and its held-out tail 2.911: DCLM filtering doubles the unseen-context rate, because the filter removes exactly the boilerplate a counter lives on.
qa reads 262 KB, not 2.6 GB. The passage lives in document.html, inside a parquet struct; the reader extracts only top-level StringArray columns and drops it silently. The 3.245 row measures id + question stems. qa gets 0% of every arm until that is fixed (one downcast in dormouse-data, not done here).

One number to not look at

arm	blended, all domains	blended, text only
A	2.505	2.505
B	2.405	2.612
C	2.317	2.628

C wins globally and is worst on text — the global number is counting the genomics share, whose counter bar is 0.19. Every arm is judged on its own per-domain ladder.

Gate

verify_mixture.py — 8 checks, 53,889 assertions, exit 0. It resolves every symlink, checks every extension against collect_files (an unaccepted one is skipped silently and the arm trains on nothing while looking fine), proves no link reaches an eval region, re-asserts leaked_ids == [], and reconciles every byte total. All five fault injections — dangling link, wrong extension, eval leak, emptied domain, extra file — were caught. anchors on all 26 arm/domain directories returns the full 2 MB, so nothing is silently empty.

Cost, and the order

5,000 steps = 25.6 M tokens = 5.67 h/arm at the canary’s 4,081 ms/step; 17.0 h for all three. A run consumes 0.12–0.21% of an arm, so no share repeats.

Run B, then A, then C. B is the arm that tests the hypothesis and the only one that can be falsified cheaply: if the pricing is right it beats A on code, math, reasoning and genomics at once. A is only interpretable beside B. C is the expensive bet and is the one that can be cut to zero without losing the answer.

Caveat: 4,081 ms/step was measured on a single-file corpus; these are 1,333–3,127-file trees. Time the first 30 steps (--timers) before committing.