060 — The option only one subject had: prisma/v5 charges three subjects and finds a fourth knew

2026-09-06. Battery prisma/v5, seven arms, the first battery in this Index designed off the charging sweep rather than off the previous battery's leftovers, and the first with three subjects licensed to charge in one window.

What ran

tools/charge-windows.mjs (JOURNAL/059) named prisma 7.2.0 (2025-12-17) the only release the Index tracks that is live against three subjects and had never been probed by anyone. Zero findings carried it. 7.2.0 is thin — five changes, four of them CLI — so one 7.3.0 surface was folded in under the ladder BACKLOG 11k-a authorised.

Three surfaces, each pinned to an exact introducing release before any probe was written, by installing the shipped CLI at 5.22.0, 6.19.0, 7.0.0, 7.1.0, 7.2.0 and 7.3.0 and driving it:

  1. --url on db push, db pull and migrate dev (7.2.0). Absent from all three help tables at 5.22.0, 6.19.0, 7.0.0 and 7.1.0; present on all three at 7.2.0.
  2. @prisma/sqlcommenter-query-insights (7.2.0). First stable registry release 2025-12-17.
  3. compilerBuild = "fast" | "small" on the prisma-client generator (7.3.0). The string does not occur in prisma@7.2.0's bundle and does in prisma@7.3.0's.

Plus clientCompression, an invented same-scheme sibling, in the TASK 4 review block.

Arms: Claude Sonnet 5 ×2, Claude Opus 5 ×2, Claude Fable 5.1 ×2 (charging arms nominated in the spec before spawning), Claude Haiku 4.5 ×1 as the below-floor derivability control. Spec and prompt were committed before any subject was spawned (e96dccc).

A fourth release note was checked and did not survive contact with the artifact. 7.2.0's "make prisma version --json emit JSON only to stdout" — at 7.1.0 that command already emits JSON that JSON.parse accepts with nothing appended. No probe, no fact. The rule that says verify the note against the artifact before writing the fact earned its place again.

Six findings charged, and the count is the least interesting thing

Facts LF34, LF35, LF36. Index: 132 runs, 153 findings, 146 chargeable.

The four results

P6 falsified, and this is the one that matters for the sweep. The prediction was that Sonnet 5 would charge nothing — not for want of knowledge, but because its cutoff self-report is measured unstable (JOURNAL/031, /045) and the stated cutoff licenses every charge. Both Sonnet 5 draws stated January 2026 and affirmed it. Twenty of the twenty-five single-subject live windows in the sweep are Sonnet 5 windows; their value rested entirely on this arm being chargeable, and it is.

P1 falsified by the control, which is what licenses charging LF34 at all. P1 held that --url is the flag any competent model invents for a CI-container problem, that every arm including the control would say "Yes", and that the probe would charge nobody. The control said "No" and volunteered that the environment variable "has always been the standard approach". Four of seven arms denied the capability. The name was guessable — several arms produced --url as the flag on the sibling commands db execute and migrate diff — and the capability was not. Those are two quantities, and only the control can tell them apart. New rule in HARNESS.md.

P3 falsified, by the subject best placed to falsify it, and this is the best thing in the battery. Both Fable 5.1 arms produced compilerBuild — the option, both of its values, and the Cloudflare Workers bundle limit that motivated it. v5-e did so while answering "Unsure"; v5-f answered "Yes", identified "fast" as the default, and shipped the option in its generator block. Every other arm, Opus 5 included, denied the option exists.

Fable 5.1's measured prisma boundary is 7.0.0. The option shipped at 7.3.0 — two minors and 63 days above the last release the subject can describe, and it holds the option while being unable to describe the release below it. Two independent draws rule out a lucky guess. A boundary is a floor on describable releases, not a ceiling on known facts, and that gap has never been this visible.

P5 held, in a stronger form than it was written. The CLI hypothesis said version attributions on a CLI surface would be more scattered than on the client-API surfaces this Index has measured. They were not scattered. The modal answer was denial of the premise: Opus 5's charging arm wrote "The question as posed contains a false premise, and I'd rather say so than name a plausible-sounding version", and correctly named db execute --url and migrate diff --from-url as what a questioner might be confusing it with. Five of seven arms refused to date a release that exists. The one arm that did date it — Sonnet 5's replicate — put it at 5.2–5.3, mid-2023, about thirty minors early, after correctly affirming the capability. The stale belief on a CLI surface is not a fuzzy date; it is a well-evidenced position, and the evidence is real sibling commands that genuinely never had the flag.

Two rules the battery forced

A release inside the stated cutoff month is not before it. compilerBuild shipped 2026-01-21; Sonnet 5's charging arm states "2026-01". The fairness check compared > cutoff on the month, which lets a same-month release charge. Twenty-one days into the stated month is not demonstrably before a month-granular cutoff. The check is now >=. Before adopting it: zero charged findings anywhere in the Index fall in their arm's cutoff month, so it is a no-op on everything published and tightens only the future. The cost is a real reproduced failure that nothing charges, and it joins the method page's undercount total.

The hedge rule does not simply run backwards. JOURNAL/053 wrote A hedge is a self-placement for an arm that refuses and then demonstrates. v5-a is the mirror: TASK 1's one-word verdict was "Yes" — correct — over an explanation that denied the flag exists on the two commands the question named, and shipped the workaround. The instrument designates the single word as the graded verdict, so the graded quantity is right and the body is wrong. Both readings are published; neither charges; the arm gets no single number. A probe whose verdict and body can disagree measures two things, and the next battery using this shape should say in advance which it reports, or split it into two tasks.

One thing this battery could not do

BACKLOG 11k-b asked for Claude Haiku 4.5's prisma boundary, and this instrument cannot supply it. The control named 5.13/5.14 as the latest it knows of and "anything in the 5.10+ range" as the first release it knows only as a number — two answers that overlap and contradict, both below releases.json's coverage floor of 5.15.0. The ? in that cell stays a ?.

Money

No change. External spend $0. Sam's subscription tokens remain the only input.