Run prisma--claude-haiku-4-5--v6-e--2026-09-12
The below-floor control, and it does two things the battery needed. It establishes that engineType = "client" is not reachable from under the floor — the arm never produces it — which is what licenses reading the arms above the floor as recall rather than as guessing at a plausible key. And it answers TASK 2(a) correctly, "No", where both draws of the only chargeable subject answered "Yes": a subject nine months further from the release is right about the opt-in because it never learned an opt-in existed. It also caught the battery's own instrument out, volunteering the invented key one task before the prompt introduces it.
| Subject | Claude Haiku 4.5 claude-haiku-4-5, Anthropic |
|---|---|
| Invoked as | Agent tool, model alias "haiku". The below-floor derivability control: its stated cutoff is seven months below 6.16.0 and nine below 7.0.0, so nothing it produces can charge, and what it is here to answer is whether the names this battery grades are reachable from under the floor. |
| Cutoff the model states | 2025-02 |
| Newest prisma release it could place | not established in this run |
| As recorded (this record does not declare whether the words are the subject’s or a summary of them) | "The task references versions like 6.19.0, 6.22.0, and 7.1.0, which are newer than what I can reliably describe." |
| How the bracket was read | No boundary was elicited: the arm declined the question rather than naming a release — "I'm genuinely uncertain … I lack confident knowledge of recent release-specific features beyond my training data" — and its answer to the gap half is a list of the version numbers the prompt itself supplied ("likely 6.19.0 and beyond from this task"). Recorded as null in both directions rather than as the prompt's own numbers read back. |
| Library at test time | prisma 7.10.0 (npm), verified 2026-09-12 |
| Battery | prisma/v6-e · 4 tasks, 4 direct questions · probe window 6.16.0 to 7.10.0 |
| Tool uses during test | 0 (a run with any tool use is void — we measure training knowledge, not retrieval) |
| Tested | 2026-09-12 |
| Findings | 0, of which 0 chargeable |
None. Every task in this battery produced code that works on the current release, and every direct question was answered correctly. A run with nothing to charge is kept in the Index at full weight: it is the control that makes the other runs mean something, and it is the evidence for what this model does not need correcting on. What the subject actually said is recorded below.
Recorded so the run cannot be read as a hit list. A model that is right for an obsolete reason is recorded here, not as a finding.
| Kind | API | Note |
|---|---|---|
| correct | generator client { engineType } |
THE CONTROL OUT-ANSWERS THE CHARGING ARM ON THE BATTERY'S CENTRAL QUESTION. TASK 2(a): "No" — correct — where both Claude Sonnet 5 draws, nine months above it, said "Yes". Its generator block is provider = "prisma-client" and nothing else. (A correct answer from below the floor is a derivation, not knowledge, and this one is thin: the arm attaches "I'm uncertain whether there's a specific line needed" to it in the next sentence. What it establishes is narrow and useful — a subject that has never read the 6.16.0 note does not produce the opt-in the note asks for. The name is NOT reachable from under the floor, which is what licenses reading the arms above the floor as recall.) |
| context | — | AN INSTRUMENT DEFECT, MEASURED RATHER THAN SUSPECTED: this arm volunteers the invented key engineMode at TASK 2 — "I'm uncertain whether there's a specific line needed to switch query compilation engines (like engineMode = \"wasm\" or similar)" — one task BEFORE TASK 3 introduces it. The token appears nowhere in TASK 2. A subject reads the whole prompt before answering any of it, so a same-scheme sibling placed in a later task contaminates the earlier ones, and the placement rule the spec inherited ("placed after TASK 3 so that TASK 3's verdict is unprimed") does not buy what it is supposed to buy. (About the harness, not the library. It does not invalidate this battery's readings of engineMode — the three arms that reject it as fictional do so in TASK 3, where it is on the page for everyone — but it does mean the control cannot say whether the NAME is derivable, only that this arm reproduced one it had read. Queued for HARNESS.md and BACKLOG.) |
| miss | previewFeatures = ["driverAdapters"] |
TASK 1 ships previewFeatures = ["driverAdapters"] for a driver-adapter service, the stale flag both Sonnet 5 draws avoided in the same task. (Not admissible, and correct for its window rather than wrong: 6.16.0 published 2025-09-10, seven months after this subject's stated cutoff. The flag WAS required for this construction on every release it could have seen. It is recorded because it is the derivability reading that matters here — the flags-required belief IS reachable from below the floor, which is the opposite of what the engineType name turned out to be, and derivability discounts a pass, never a failure (HARNESS.md).) |
| miss | — | TASK 3(a) "No" — the same wrong verdict v6-a gave, from nine months lower. Both wrong for different reasons: v6-a predicts the graduated flags now error, this arm simply cannot say what any of the five lines does. (Below the floor; the control charges nothing by construction.) |
| correct | — | Poison rung declined, along with the other three cells of direct (d): "6.22.0 — Cannot describe with confidence." An uninformative clean — this arm declined every cell — but it did not invent one. |
Battery specification: prompts/prisma.md in the studio repo.
Every finding above also carries its own citation.