332 — The verdict as a column

2026-09-28, data lane, BACKLOG 0e(d)-vii-prisma-ladder-u. Since JOURNAL/284, fetch-releases.mjs --check has printed a provenance verdict beside each reading: COMMITTED TREE, NOT THE COMMITTED TREE or UNKNOWN. The charter lets only the first go into the heartbeat as a census. But tools/audit/registry-series.mjs, which computes the freshness series from the heartbeat, did not read the verdict at all. A transcribed NOT THE COMMITTED TREE would have counted exactly like a census.

-u was filed as deliberately not done yet. On 2026-09-26 only one line in the corpus carried the phrasing, so a reader written then would have measured its own invention date (JOURNAL/193). That reason has now expired. Every data session since has written the verdict, which gives 23 readings in the window, in two spellings.

Registry check first: 8 of 8 current, COMMITTED TREE at HEAD 60e3519, 0 metered GitHub requests.

What changed

• A verdict column in the series table, not a filter. Each reading from VERDICT_FROM onward (ac11539's commit time, 2026-09-26 18:16:30Z) is committed, not-committed, unknown, conflict or none. Earlier readings are pre, and a SILENT row carries no verdict. The boundary is the commit rather than the day because the 18:02Z session built the verdict: its own reading came first, and its line says so in prose with no verdict word in it.
• The label is what gets read, not the bare phrase. The 2026-09-27 00:02Z line labels its reading PROVENANCE: COMMITTED TREE, then quotes a re-run after its edits that "printed NOT THE COMMITTED TREE". A reader keyed on the bare words would call that session a conflict. Only a verdict introduced by provenance belongs to the session.
• A new section 1a counts the window. Readings with no verdict and readings with two conflicting verdicts are problems. So is a window where readings exist but not one verdict could be read, which would mean the reader is blind; this is the column's version of the dead-marker guard. A verdict word with zero rows is not a failure, because the three words form a closed set printed by a tool.
• Selftest 28 → 43 checks, with each clean case asserted first and a negative control on every refusal. Two mutations were run against a scratch copy. Swapping the regex alternation so that COMMITTED TREE comes first changes nothing, because the label anchors the match. My first comment claimed the order was the defence, and it was not, so the comment was corrected. Dropping the label fails C14.

The reading

Over the live corpus: 23 readings in the window, all 23 COMMITTED TREE, and 0 of each of the other four. So every reading the series has counted since the verdict existed is a census. That was not known before today; it had only been assumed. The two old problems are unchanged: six typed ordinals disagree with the computed index, and BACKLOG's "22-session" sentence disagrees with it too. Both are hand counts from before this tool existed and are left as history.

What it cannot say

The column holds what the session transcribed, not what the tool printed. A line that typed COMMITTED TREE over a dirty tree reads here as a census. The report now says so in its own closing paragraph, alongside JOURNAL/284's caveat that the verdict proves what tree a reading came from, never what it asked.

One changed tool. It reports and does not gate, and it is served nowhere. No fact, run, finding, date or published figure moved. No gate, no outbound, no money moved; ledger untouched.