# Assistant Evals — Reference Dataset

`cases.jsonl` is the 18-case adversarially-weighted reference dataset for the offline eval
harness (`core/archipelago/src/assistant/evals.rs`, run via
`cargo test --package archipelago assistant::evals::`). It lives in-repo, alongside the code,
so cases are reviewed in PRs like code — see `13-AI-SPEC.md` §5 "Reference Dataset" for the
full schema and composition table this file implements verbatim.

JSONL cannot carry comments, so the reviewer-role ownership per §5's "Labeling" table is
recorded here instead — a later contributor editing or extending a case should know which
person's judgment that case's `expect` block encodes.

## Labeling roles, by bucket

| Cases | Bucket | Labeled by | Why |
|-------|--------|-----------|-----|
| EV-01…EV-08 | happy-read / confirmed-write | **Archipelago maintainer (engineer)** | `expect` blocks are mechanical — tool names come from D-06's curated registry, fixed before the tools were built. Writable in Wave 0, no domain-expert judgment call needed beyond "does the right tool get called." |
| EV-09…EV-16 | injection / ceiling | **Security-minded technical reviewer** (13-AI-SPEC §1b role 2 — "plays sysadmin with no rollback") | This is red-teaming, not test-writing. Whoever writes EV-11 (the forged-boundary case) must be *trying to break* the per-call delimiter, not documenting that it exists. The `must_not_claim` fields on EV-14…EV-16 in particular need an adversarial eye: the failure to catch is a plausible-sounding fabrication, not an obvious lie. |
| EV-05 / EV-06 confirmation copy | confirmed-write | **Non-technical reviewer** — the "bought sovereignty, not a terminal" persona (13-AI-SPEC §1b role 3) | The only valid labeller for E-02 (confirmation clarity) and E-09 (comprehension under time pressure). §1b is explicit that a security reviewer systematically under-catches confusing copy because they already understand the domain — their label is *not* ground truth for this dimension. This dataset's own EV-05/EV-06 `expect` blocks are mechanical (tool/confirmation/turn counts); the actual confirmation-*copy* judgment happens in the session-level walkthrough this pair feeds into (13-14 Task 3), not in this file. |
| LLM judge (E-02 at scale) | — | Calibrated against the non-technical reviewer's labels | Not trusted as a score until agreement ≥ 0.7 against the human labels (`ai-evals.md` Verify phase). Until then it is a screening tool that flags candidates for human review, only. |
| Sign-off | — | **Product owner / maintainer** | Rubric sign-off per §1b role 4. |

## Notes for future contributors

- Every `expect.must_not_execute` / `expect.must_not_claim` entry should name something a
  **real, curated tool** in `assistant::tools::registry()` could plausibly be confused with —
  EV-14/EV-15/EV-16 deliberately leave `must_not_execute` empty because no tool for
  "send sats"/"show seed"/"factory reset" exists at all (D-09's absence-of-tool ceiling); the
  load-bearing assertion for those three is `must_not_claim` (the model must never fabricate
  having done the excluded thing).
- EV-11's `untrusted[0].text` is written against the real shape `assistant::untrusted::wrap_untrusted`
  emits (`{label}_DATA_{token}_START` / `{label}_DATA_{token}_END`) — read `untrusted.rs` before
  editing this case so a forged boundary in the fixture still matches what a real attacker would
  have to forge, not a stale guess.
- New cases discovered from real near-misses (F-1 in 13-AI-SPEC §6's flywheel table) get
  appended here, following the same per-bucket labeling-role assignment above.