# Leg 2b — the runtime-dividend A/B (window 2026-08-21, ~07:40–07:55 UTC)

Prereg: AB-LEG-PREREG.md (sha pinned in PREREG-INDEX.txt, unamended).
DISCLOSED DEVIATION: no docker/podman exists on the box, so Arm OLD ran the
official v0.32.9 release binary as an isolated second instance (own port
11436, own scratch model store copied blob-for-blob, 18.0 GB) instead of
the registered container — same substance, different vehicle. First start
attempt hit a foreign listener on the registered-adjacent port 11435
(unattributable cross-user socket; per house law we curled it, got nothing,
and sidestepped to 11436 rather than pid-guessing).

## Result (gemma4:26b, robber Q1, temp 0/seed 0/ctx 32768/npred 220, n=10 + warm-up per arm, empty card)
| arm | decode tok/s median | range | determinism |
|---|---:|---|---|
| OLD — ollama 0.32.9 | **154.09** | 153.82–154.22 | 10/10 byte-identical |
| NEW — ollama 0.32.13 | **203.00** | 201.37–203.63 | 10/10 byte-identical |

**P1 CONFIRMED: +31.7% from the runtime alone** (>25% registered threshold).
P2 (archive's ~145 for the OLD era): measured OLD = 154.09, within the
registered ~15% band of ~145 — P2 CONFIRMED as registered.

> CORRECTION (2026-08-21, recon pass, RECON-B/C in this dir): this note's
> original P2 paragraph attributed the OLD-era gap to "a 70B co-resident"
> and decomposed the archive jump into runtime + contention. The recon
> contradicts both props: no shipped kit publishes a ~145 gemma4:26b
> median (nearest shipped figure: production p50 138.600, n=1,982 live
> rulings, 08-15), and the 08-14 archive record reads resident models
> ['gemma4:26b'] — solo; the only 70B residency on record is
> 08-16T01:27–01:33Z, AFTER every ~140-class reading. The decomposition
> is RETRACTED; what survives is the controlled pair itself (154.09 →
> 203.00, +31.7%, this run) and the honest bracket across designs
> (1.32×–1.54×, see RECON-C item 15). Also note: the two arms generated
> DIFFERENT text (eval 128 vs 130, different shas, each 10/10 stable) —
> the rate comparison stands; "same answer, just faster" would be false.

Old instance SIGTERMed and verified stopped; scratch store retained on the box for the article's audit trail until after publication.
Raw rows: ab-runs.json (this directory).

OLD-ARM VERSION RECEIPT (added 2026-08-21 after recon flagged the gap):
ab-runs.json carries only new_version; the old arm's served version WAS read
at staging — the run log's line 3 reads verbatim `old runtime up: {'version':
'0.32.9'}` (an /api/version call against the isolated instance's port before
any scored call). So the 0.32.9 identity rests on a served response, not on
the binary's filename alone.