session 2026-09-10: the scoping sitting — REVIEWED-138/-139, the child-gate FIX, PENDING-182/-183, and six register defects

Session record, ledger, MEMORY.md rotation (prior Active Session demoted to
MEMORY-reference.md), OWED-6 queued against the frozen ladder, KG triples, and
skill-harvest proposal D.

N-now re-measured for REVIEWED-123 condition 2: 59, split 31 real / 28 mumble.
⚠ Not commensurable with the banked 36.9%, which used the one-prompt predicate.
The ladder pointer's 'direction has REVERSED / rising fast' clause is superseded
— it was a dated finding read as a status, and it contradicted the new figure in
the same sentence.

PENDING-archive.md

@@ -3521,6 +3521,14 @@ The reason is that (a) and (b) both answer *"which population did this field mea
 > A dry run shows what the tool *intends*. It cannot show what the tool will do to a file whose shape the tool misreads. On 2026-08-17 a 16-file stamping pass dry-ran clean, applied clean, and silently orphaned the frontmatter of the 3 files whose shape differed — and the post-hoc check reported `malformed: none`, because it asked whether each file *began with frontmatter and a banner* (true) rather than whether the stamp had *preserved the record's keys* (false).
 > **The check that worked was the one that could not be written from intent:** recover each file's pre-change text from git, extract its key set, and assert no key was lost. That control cannot be satisfied by a file the tool mangled, because its subject is the *difference* rather than the result.
 
+**OWED-6 — classify transcripts by the TWO-prompt first-user-turn signature, and never compare a composition figure computed with one predicate against a figure computed with another.** · **Authorizing ruling: NONE — earned 2026-09-10, queued unruled.** · **Ladder claim-class: population and denominator claims.**
+
+> **Earned by the rule of three, counted forward rather than noticed afterwards.** The transcript real-vs-mumble classifier has now been written from scratch three times — PENDING-179 gate 2a (2026-09-03), the rebuild that found the two-prompt defect (2026-09-09), and today's N-now composition split. Each writing re-derived the same predicate; the second found a defect in the first, which is the only reason the duplication was productive.
+>
+> ⚠ **The rule has two halves and the second is the load-bearing one.** *(a)* There are **two** generator prompts — the mumble's `You are writing ONE line as Tarbuckle.` and the seam/wrap/invoke's `You are Tarbuckle. Your character, filed and unalterable:` — and a predicate that knows one **undercounts**. `human_turns()` is not a discriminator at all (PENDING-179). *(b)* **A composition figure computed with a different predicate is not comparable to a banked one.** Today: N-now 59, split 31 real / 28 mumble (47.5%) by the two-prompt predicate, against a banked 41/24 (36.9%) computed by the one-prompt signature. **The apparent rise is partly a change of instrument, and reporting it as a trend would have been the week's own defect** in the figure the ladder trial reports at every wake.
+>
+> ⚠ **NOT ADDED TO THE LADDER.** REVIEWED-123 condition 1 freezes `reference-verification-ladder.md` against additions from any source whatever its authorization. Queued here as condition 3 directs. ⚠ **Note the coupling OWED-5 already flagged:** this row is a rule about *denominators*, queued behind the grading of a *ladder-retrieval* trial, and today's N-now measurement is itself an input to that trial's bound.
+
 **OWED-5 — a must-detect control must report its denominator, and a denominator of zero is a FAIL.** · **Authorizing ruling: NONE — steward-stated 2026-09-04, queued unruled.** · **Ladder claim-class: gate-design claims.**
 > **A control that finds nothing to test reports success indistinguishable from a control that tested everything and found no fault.** `PASS (0/0)` is not a pass; it is the instrument saying it never ran, in the voice of an instrument that did. Print the denominator on every must-detect line, and make an empty one fail loudly.
 >