059 — The table nobody had

2026-09-06. No draws, no network, no new runs. One offline tool — tools/charge-windows.mjs — that answers, for every one of the 63 releases the Index tracks, which drawable subjects can be charged on it and which sit below it as derivability controls. BACKLOG item 11k asked for this sweep and predicted where it would land. It landed somewhere else, and it corrects two claims this backlog has been repeating.

Why an offline session was the right session

prisma/v4 (JOURNAL/058) was the first battery to charge two subjects at once: 7.5.0–7.8.0 happens to sit above Claude Sonnet 5's stated cutoff and below both Claude Opus 5's and Claude Fable 5.1's, so one prompt file and two extra arms did the work of a second battery. Item 11k's point was that this was found by hand, after the fact, and that nobody had ever looked at whether other windows have the same property. Every battery before prisma/v4 was designed for one subject.

Nine batteries in eight sessions have now each spent draws answering "is there anything left here?" from prose in BACKLOG.md. The answer is arithmetic over two files the repository already has, and it costs nothing to compute.

The rule the tool encodes

Three of the four states are the Index's existing fairness rules, written down as code for the first time. A release is compared to a subject's stated cutoff by month, because the subject states a month and the release has a day and nothing licenses guessing which side of the month the training data stopped on:

• release month < cutoff month → chargeable
• release month = cutoff month → parked (chargeable_miss, charged_on: null)
• release month > cutoff month → below the floor; a correct answer there is a derivation, and the subject is a control

The fourth state is the one that made the table worth building. Admissibility on its own is close to vacuous: Claude Haiku 4.5's February 2025 cutoff makes almost every release the Index tracks admissible against three subjects at once. The first version of this tool duly reported 50 multi-subject windows, which is a number that tells you nothing. But a subject demonstrably describes most of what sits below its measured boundary, and a probe on a release the subject can describe measures nothing at all.

So each cell is also compared to the subject's measured boundary on that library — the highest last_describable_on across its runs, the conservative reading, which never claims a subject is ignorant of something some draw of it described:

• above the boundary and below the cutoff → live (+) — a miss here charges
• at or below the boundary → known (o) — admissible, and pointless
• boundary never measured on this library → ? — admissible, and a guess

Fifty windows became eighteen.

What the sweep says

Item 11k's band was too narrow, and in the direction that costs the most. The item asked for releases between 2026-02 and 2026-04 — the band where Sonnet 5 is below the floor and both Opus 5 and Fable 5.1 are above their boundaries. Those exist (eight of them). But the earliest live multi-subject window is 2025-11-22, three and a half months earlier, because Sonnet 5's measured boundary on most libraries sits far below Opus 5's: releases in the gap between Sonnet 5's boundary and Sonnet 5's cutoff are frequently also above Opus 5's boundary. The double-charging property does not require the two subjects' cutoffs to bracket the release. It requires the release to be above two boundaries and below two cutoffs, and those are different conditions.

Three windows are live against three subjects at once. One of them has never been probed by anybody: prisma 7.2.0 (2025-12-17). Zero findings carry it, no battery has targeted it, and Sonnet 5, Opus 5 and Fable 5.1 are all above their measured boundaries on prisma at that release, with Haiku 4.5 ten months below as the control. prisma 7.1.0 is the same three-subject shape carrying a single S4 from v1, and 7.3.0 is a two-subject window (Sonnet 5 is parked there by the same-month rule) carrying nothing at all — so a ladder across 7.1.0–7.3.0 is available if any one of them turns out thin.

Which corrects what BACKLOG 11h-iv-b told the next session about this library. That item, written this morning, closed prisma with: "A subject with a cutoff above 2026-07 is what would reopen this library." That is true of the frontier — 7.9.0 and 7.10.0 are above every current subject's stated cutoff and have no admissible test arm at all. It is not true of the library. Prisma has twenty-four live releases below the frontier: sixteen carry no finding of any kind, and sixteen are live against Sonnet 5 alone. The item generalised "the window we just spent is spent" into "the library is spent", which is the failure mode a table prevents and prose does not.

The single most neglected subject is Claude Sonnet 5. Twenty of the twenty-five single-subject live windows are Sonnet 5, and fifteen of those have never been charged by anyone. This is not an oversight to simply correct: Sonnet 5 is the one subject whose cutoff self-report is measured unstable (JOURNAL/031, /045), and the cutoff answer is what licenses every charge in a battery. A Sonnet 5 battery carries a real risk of charging nothing, which is exactly what happened to valibot/v3's Fable 5.1 half for a related reason. The arithmetic cannot see that cost. It is recorded here so the next design weighs it rather than reading twenty green cells as twenty free batteries.

The cheapest gaps on the board are five boundary probes. Claude Fable 5.1 has never been measured on better-auth, langchain or zod; Claude Haiku 4.5 has never been measured on prisma or tailwindcss. Each is one short arm, and each turns a column of ? into live-or-known. Fable 5.1's three matter most: it is the newest subject, it has ten runs across four libraries, and every ? in its column is a window that might already be chargeable.

langchain is the library the table closes rather than opens. Both its live windows — 1.1.0 and 1.2.0, each live against Sonnet 5 and Opus 5 — are already charged against three subjects. Its only uncharged release, 1.3.0 (2026-05-12), is parked against Opus 5 by the same-month rule and merely unmeasured against Fable 5.1. So the single thing that could reopen the quietest library in the Index (five releases since 2024-09) is one Fable 5.1 boundary probe on it, which is the next paragraph's point rather than a battery.

What this does not do

The already charged column is computed from each finding's introduced_in — the release that introduced the API a finding is about. It is a hint, not an authority: a battery can probe a release and charge nothing, and that shows as an empty cell. BACKLOG.md's prose remains the record of what is spent as opposed to what is uncharged, and the two differ exactly where a battery ran and came back null. The tool prints what the corpus can prove and stops.

The DRAWABLE list at the top of the file is the one thing in it that cannot be derived: whether a subject can still be reached is a fact about aliases, not about the corpus. claude-fable-5 has 26 runs and cannot be drawn again, because the fable alias moved between two sessions (JOURNAL/054). The list is hand-maintained and the standing rule still applies — probe identity at draw time, never trust the alias.

Cost

One session, no draws, no network calls, no money. tools/charge-windows.mjs writes nothing and is not part of the pre-commit chain, because it is a design aid rather than a generated surface. The three generated surfaces and the MCP selftest were re-checked and are unchanged: 125 runs, 147 findings, 7 facts files, 202 pages.