# P3 — Blind Evidence Research Prompt

*Run in a fresh claude.ai chat with web access. Paste this prompt, then up to 10 blind intervention records (the `researchable` ones from `interventions.json`). This stage is BLIND — it sees no candidates, no New Mexico framing, no election. It fills in direction, confidence, summary, detail, and sources for each. Save output as `domains/<domain>/research/batchN.json`.*

*Note: only `researchable` items go to this stage. `posture` and `aspirational` items skip research entirely — leave their ratings null.*

---

## Your task

For each policy intervention given below, research what the best available evidence says about whether it works, and fill in the evidence fields. You are evaluating interventions on their own merits, as general policy questions — you are NOT told who proposes them or where they would be applied, and you should not speculate. Judge the evidence as it stands in the research literature.

## For each intervention, determine

**direction** — does the intervention achieve its intended outcome, per the evidence? One of:
- `Positive` — credible evidence it works, on real outcomes, broadly.
- `Mixed` — works on some outcomes/contexts but not others, or evidence conflicts.
- `No effect` — credible evidence it does not move the outcome.
- `Negative` — credible evidence it backfires or worsens the outcome.

**confidence** — how good is the evidence behind that direction? One of:
- `High` — randomized controlled trials or strong quasi-experimental designs (regression discontinuity, difference-in-differences with credible controls, instrumental variables).
- `Moderate` — credible observational studies or pre/post with valid comparison groups.
- `Low` — correlational, single-case, or descriptive evidence only.

These are independent. A precisely-measured failure is `No effect / High`. A promising but thin result is `Positive / Low`.

## Critical judgment rules

- **Real outcomes, not gameable proxies.** Judge `direction` on real-world outcomes (e.g. actual victimization, actual housing units built), NOT on proxies that can be gamed or that merely reflect reporting/recording (e.g. reported-crime counts, arrests). An effect that appears only on a gameable proxy is a red flag — rate it down, do not call it `Positive`.
- **Broad over narrow.** An intervention that moves the genuine top-line outcome earns `Positive`. One that moves only a narrow sub-measure while the top line is untouched is `Mixed` at best.
- **Net over cherry-picked.** If it improves one real outcome but worsens another, that tradeoff defines the rating — say so in `detail`, do not bury it.
- **Use existing clearinghouses where they cover the intervention** (e.g. CrimeSolutions, What Works Clearinghouse, Results First) as a strong input, then supplement with primary literature.
- **Some interventions resist rigorous study** (e.g. broad enforcement-posture changes are hard to randomize). If so, say that in `detail` and let it lower `confidence` — do NOT mistake "hard to study" for "doesn't work."

## Sourcing

- Every intervention needs at least one source with a working URL. Prefer peer-reviewed research, government evaluations, and nonpartisan clearinghouses.
- Cite what the rating rests on. If evidence is thin or contested, say so in `detail`.

## Hard constraints

- **Stay blind.** Do not guess who proposes this or where it would apply. No New Mexico, no candidates. If the intervention text contains a place or name, ignore it and evaluate the general intervention.
- **No advocacy.** Report what the evidence shows, including null and negative findings, with equal prominence.

## Output format

Return the same JSON array you were given, with the empty fields now filled:

```json
{
  "id": "unchanged",
  "intervention": "unchanged",
  "domain": "unchanged",
  "type": "researchable",
  "direction": "Positive | Mixed | No effect | Negative",
  "confidence": "High | Moderate | Low",
  "summary": "one-line plain-language verdict for the collapsed view",
  "detail": "what was studied, where, what happened, key caveats, the real-vs-proxy note, and any 'hard to study' flag",
  "sources": [{"title": "...", "url": "..."}]
}
```

Do not change `id`, `intervention`, `domain`, or `type`. Fill only the evidence fields.