303 — The names that stayed

2026-09-27, distribution lane. DISTRIBUTION D7: the two small items JOURNAL/301 left.

The traction check ran first: 0 moved, 0 unreadable, 6 current, 2 not checked, exit 0, 3 requests.

BM1's summary

The stored summary in the zod BM1 results file still said the pack would need "about 28 fact-dependent tasks" to repay itself. JOURNAL/301 fixed that unit on every rendered surface. This paragraph is not rendered anywhere, but it is served in index.json and in the results file, so an agent reading the JSON still got the old unit.

The figure is 28.5. It counts tasks in the run's mix, where 12 of the bare arm's 36 cells were null tasks. The summary now says that. Because this is prose in a results file, the change is not silent: a bracketed, dated note at the end of the paragraph says what it used to say.

Reading the paragraph for that one phrase turned up a second problem, the same one JOURNAL/297 and /301 found in the drafts. It said "so the pack single measured effect is not context bulk", which states B3 as a finding. It now says B3 was a registered prediction that held, not a significance test, and the note records this change too. No cell and no derived figure moved.

The per-task names

Three fields are named *_per_task but hold per-cell values: mean_chars_per_task, mean_retry_chars_per_task and pack_break_even_tasks.bare_retry_chars_per_task. Renaming them would break every consumer that reads them. The instruction was to document the unit wherever an agent reads these fields:

• Every completed benchmark in index.json now carries field_units. It gives the unit of all three fields and says what pack_break_even_tasks.tasks counts: tasks in the run's mix, which is the registered rule, and not fact-dependent tasks.
• The MCP server's measured_effect now includes mean_chars_per_task_unit, quoted from the index rather than written a second time, so the two cannot drift apart. The selftest checks it: 97/97.
• llms.txt has no field notes, and its benchmark lines already say "cell", so it did not change.
• A comment in build-index.mjs said "per task". It now says per cell.

claims-check is OK on all three drafts.

What is left

Nothing in D7 except the posting itself, which is asks d7-1, d7-2 and d7-3 for Sam.