Run prisma--claude-opus-5--v6-c--2026-09-12 · self-test: the subject is the operator
A known-cell draw that answers the battery's central question correctly at every task, and in doing so falsifies P4 and P5 — it identifies the invented key as fictional and engineType as inert in the same answer (P4 said no arm would), and it names 7.0.0 at TASK 4 while saying at TASK 3 that the key is inert (P5's own stated falsifier). It is also the operator's own model, so self_test is true and the arm is read with that limitation attached. The one thing it gets wrong is a date: driver-adapter GA at 6.6.0 rather than 6.16.0, an error of five months inside a paragraph whose advice is correct.
| Subject | Claude Opus 5 claude-opus-5, Anthropic |
|---|---|
| Invoked as | Agent tool, model alias "opus". Drawn at a `known` cell — Claude Opus 5's measured prisma boundary IS 7.0.0, so it has been observed describing the release LF39 is dated at — and declared NON-CHARGING in the spec before spawning, not after the transcript was read. |
| Cutoff the model states | 2026-05 |
| Newest prisma release it could place | 6.7.0 · 2025-04-29 (~13 month lag) |
| Oldest prisma release it could not place | 6.8.0 · 2025-05-15 (so this run brackets the subject’s boundary to 2025-04-29 – 2025-05-15) |
| As recorded (this record does not declare whether the words are the subject’s or a summary of them) | "Latest I know of: Prisma 7.x exists; the highest specific number I'd assert is 7.1.0, and I know that partly because you named it. I could not have told you the current patch version." |
| How the bracket was read | Volunteered exactly and without hedging on either side: "Most recent release whose contents I can actually describe: 6.7.0, around May 2025" and "First release I know only as a version number: 6.8.0. I'm confident it exists and I cannot tell you a single thing that changed in it." The dates it gives for 6.6.0 (April 2025) and 6.7.0 (May 2025) are both right to the month. |
| Library at test time | prisma 7.10.0 (npm), verified 2026-09-12 |
| Battery | prisma/v6-c · 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 BATTERY'S BEST ARM, AND IT FALSIFIES TWO OF THE SIX PRE-REGISTERED PREDICTIONS. TASK 2(a) "no", with the reason: "on the current release the Rust-free, TS + WASM query compiler is the only implementation. It is what you get." TASK 3(a) "yes" — correct. TASK 3(b) grades all five lines right: engineType = "client" real but inert, the preview array inert and warned about, and engineMode "No. Not a Prisma option, and never has been". TASK 4 names 7.0.0. |
| correct | generator client { engineType } |
It states the mechanism this session measured, unprompted and correctly: "Prisma's schema parser passes unrecognized generator keys through to the generator as opaque config rather than rejecting them, so it silently does nothing." That is LF39h, bisected to INVARIANT across all 31 rungs earlier the same session. It also predicts the exact failure mode of TASK 3's remaining risk — "Prisma's usual practice is to keep graduated features in the known-but-warn list" — which is the warning the CLI actually prints. |
| imprecision | previewFeatures = ["driverAdapters"] |
Dates driver-adapter GA to 6.6.0 (April 2025) — "Driver adapters went GA in 6.6.0; on any current release the flag is unnecessary and will warn." LF38 measures the release at which new PrismaClient({ adapter }) type-checks with nothing in previewFeatures as 6.16.0 (2025-09-10), ten minors and five months later. The instruction the arm gives is right; the release it attributes it to is not. (The arm is at a known cell for this battery's charge and was declared non-charging in the spec before spawning. Separately, an attribution error whose operative advice is correct is S4 territory at most, and the battery pre-registered S4 only for TASK 4 and direct (d).) [chargeable miss — a replicate, a duplicated arm’s second draw or a below-floor control charges nothing;
absent from the finding count] |
| correct | — | The poison rung is caught with volunteered doubt rather than a decline: "6.22.0 — I cannot describe it, and I cannot even confirm the 6.x line reached 6.22 before 7.0 shipped. If it doesn't exist, I'd rather say so than invent a changelog for it." 6.22.0 has never been published; the 6.x line ends at 6.19. |
| context | — | Volunteers its own reliability ranking, and it is accurate: "the parts of Tasks 1–3 I'd stand behind are the shape of the modern setup and the identification of engineMode as a non-existent option. The release-number attributions in Task 4 and question (d) are where I'd expect to be wrong." Its TASK 4 attribution (7.0.0) is in fact the one it flags as unreliable, and it is right. |
Battery specification: prompts/prisma.md in the studio repo.
Every finding above also carries its own citation.