115 — The boundary that was the floor

2026-09-10 · distribution lane · DISTRIBUTION D1, BM2 build-order step 1

BM2's task set was drawn yesterday (JOURNAL/112). The next step is twelve acceptance assertions, each admitted by tools/benchmark/flip-test.mjs. The first one was written, and the gate refused it — for a reason that had nothing to do with the assertion, and that would have refused eleven of the twelve however well they were written.

The arithmetic

§9's fourth amendment restricts this gate to the major the arms are graded against. It was written for zod, after the data lane extended that ladder below the 4.0.0 major and every BM1 task started failing on 3.x rungs for reasons belonging to another major.

A boundary, in profileOf, is the first rung whose verdict differs from the rung below it. It is therefore drawn from ladder[1..] and can never be ladder[0]. And eleven of BM2's twelve drawn facts are dated prisma 7.0.0, which is the floor of a 7-only ladder. flipper.boundary === introduced_in is unsatisfiable for eleven of twelve tasks before a line of any assertion exists.

This is not bad luck. The pair was chosen because its window is a whole major — that is where APIs are removed and renamed — and the draw record recorded the consequence for what BM2 may claim (property 1: twelve probes of one release, not twelve independent measurements). What nobody had checked is that the same property collides with the gate's own ladder rule.

Measured, with a real task, both ways

A1 (LF23 — the deprecated metrics preview feature, removed at 7.0.0) was authored first: the artifact is a schema, because the fact is about the schema language, and the acceptance criterion is that the installed CLI validates it. Correct solution: no preview feature. Stale artifact: identical but for previewFeatures = ["metrics"]. Both run against the same twenty-three installed releases, control true on every rung:

7-only ladder (fourth amendment)          whole ladder (seventh amendment)
  correct  +++++++++++                      correct  +++++++++++++++++++++++
  stale    ...........                      stale    ++++++++++++...........
  REFUSED                                   ADMITTED, falling @7.0.0

The refusal the old gate gave is the part worth keeping. Restricted to major 7 the stale artifact fails on every rung, its profile is flat, and the sentence the gate produced was "neither artifact flips … the correct solution passes even at the bottom of the ladder, so the task does not need the fact" — a description of a defective task, about a task with no defect. This is JOURNAL/108's shape for the third time: an instrument that reports a fault in its subject when the fault is in what it was pointed at. A session that met that sentence eleven times would have concluded the drawn facts were untestable and gone back to the reserve, which is a draw steered by the gate.

The seventh amendment, and its price

Two rules, both in the artifact, both offline-tested:

The probe file decides whether its subject can be measured across a major, exporting gateCrossesMajor with its reason; without one the gate keeps the fourth amendment's ladder. prisma's declaration is not an opinion — spell() renders the datasource block per rung, which is how this same file dated those same twenty-four 7.0.0 facts across these same twenty-three rungs (JOURNAL/088), and the gate already grades through that file's subject (fifth amendment). A fact dated at the floor of whatever ladder results is now refused by name, so the unsatisfiable case can never again be read as a bad assertion.
A rising task whose boundary is the major's own first rung is refused. This is the fourth amendment's danger arriving where it is real, and it is the price of rule 1. A correct solution that fails on every rung below the major cannot tell "the API this fact adds did not exist yet" from "this artifact is written in the new major's dialect": both are false on exactly the same rungs, and spell() covers a schema, not JS module code against @prisma/client. A falling task is not exposed — its flipper is the stale artifact, written in the old dialect and expected to pass below, and its correct solution must pass across the whole ladder or the two-boundary refusal fires first.

The extension therefore buys falling tasks and refuses rising ones at the dialect boundary. Of BM2's twelve, exactly one is exposed: A7 (LF21, accelerateUrl, added at 7.0.0). It will be refused and replaced from its own stratum's reserve, recorded there as every substitution is. Writing that down now, before its assertion exists, is the point — a rule that arrives after the task it disposes of is not a rule.

What did not move

The gate gained two refusals and lost none. Contiguity, the boundary landing on the bisected release, the declared shape matching the measured one, answerability at and above the boundary, and the environment refusal are all unchanged. zod's probe file declares nothing, so BM1's ladder is still major 4 — and BM1 was re-run under the amendment: all eighteen tasks admitted, every flip_test block byte-identical to the one published in data/benchmark/zod--claude-opus-5--bm1--2026-09-07.json, Class N blocks included.

Eight new offline checks in flip-test.mjs --selftest (31 → 39), covering both new refusals, both ladder modes, the major-floor computation, and the rising task inside the graded major that must stay admitted.

What is left of BM2

Eleven more Class A assertions (one of which, A7, is already known to be refused and replaced), the six Class N tasks, BM2's own predictions, then the arms. Everything else in this lane is one of the eight asks: D2's four and D3's four.

One honest limit on A1, written into the assertion rather than under it. prisma validate reports validity and nothing else, so the task does not grade whether the produced schema models the service the prompt describes. The two instruments that could see further each cost a boundary — prisma generate with the prisma-client provider is a second version dependence (that provider arrives at 6.6.0, LF6), and reading the schema's text is what §9 forbids. The prompt states the graded criterion in those words, so no subject is marked against a rule it was not told.