# Superseded: the first dynamic-boost pass's LEG 1 results (2026-09-21, 20:24–20:56Z)

These files are the language leg of the FIRST dynamic-boost pass. They are kept as receipts and
are NOT the pass the fold reads. Why: the system `ollama` instance had `nomic-embed-text:latest`
resident on the card (610 MiB, keep-alive forever, loaded at 19:51:10Z by ONE local /api/embed call
on the loopback whose caller the journal does not name — NOT the archive's hourly embed job, which
points at a remote seat; an earlier draft of this note guessed that job and was wrong) for the whole
of this leg. The 95 W pass ran with the card otherwise EMPTY (its idle
receipt reads 15 MiB; this pass's reads 633 MiB), and the render leg's own gate refused to start
on a card another process held (`CardBusy`, 20:56:19Z). To measure the boost posture under the
same condition as the 95 W pass, the tenant was unloaded (`ollama stop`, it reloads on its next
call), the hourly embed timer was paused for the window and put back, and the WHOLE pass was
re-run — both legs — into fresh files. The pre-registration's §8 names the embedder a disclosed
tenant for the language leg; the render leg's instrument requires an empty card, so the re-run
is the like-for-like measurement and this directory is the replication with the tenant present
(it agrees with the re-run within nine-tenths of a token a second at every gemma4 window — the worst
rung, 32,768, differs by 0.897 tok/s; the first note here said 'half', which was wrong).
Nothing in these files was edited. The pass log above the re-run in `logs/rung-dynboost.log` is
this run's log.