How this guide was built

The pipeline

1. 1. Extraction. For each policy domain, gather every candidate's stated proposals from campaign sites, debates, voter guides, and reporting — neutral paraphrase, sourced URL, the candidate's own framing kept verbatim.
2. 2. De-identification & dedupe. Strip names off the proposals, merge substantively-identical ones into single interventions with stable IDs, and classify each as researchable (has a mechanism that could be studied), posture (slogan or value statement), or aspirational (real goal but not a discrete intervention).
3. 3. Blind evidence research. Send the de-identified researchable interventions out for evaluation against the empirical literature — no candidates, no New Mexico, no advocacy. Each gets a direction (Positive / Mixed / No effect / Negative) and a confidence (High / Moderate / Low) grounded in cited sources. Posture and aspirational items skip this stage and remain unrated.
4. 4. Assembly. Join the research output back to the (separately-kept) attribution map by ID, validate that no intervention is orphaned and every researchable record has a rating + sources, and emit the final data.json this site renders.

The blinding in stage 3 is the design's load-bearing piece. The model evaluating the evidence does not know who proposed the intervention or where it would apply — it is judging the policy mechanism, not the politician. Names re-join the picture only in stage 4.

What this guide is not

• Not a recommendation. A Positive rating means an intervention's mechanism has worked in studied settings — not that it's the right call for New Mexico, or that any particular candidate would implement it well.
• Not New-Mexico-specific. Evidence reflects whether the intervention worked elsewhere, under studied conditions. Local supply, demographics, geography, and capacity all matter and are not yet accounted for.
• Not jurisdiction-aware. Some proposals (e.g. a U.S. Senator's federal voting-rights bill) operate through federal levers; some (e.g. a county sheriff's policing approach) are local. The site lists proposals as candidates stated them; it does not yet flag which office actually controls each lever.
• Not infallible. See the caveat above. Sources are cited so you can check the work.

The prompts

The five files below are the actual prompts and reference data used to produce this guide. Download them to inspect the methodology, audit a result, or run the same pipeline on another race.

• P1 · Extraction

  Download ↓

  Pull each candidate's stated proposals in one policy domain, with the candidate's own framing and a working source URL. Stenographer mode — no evaluation.

  P1_extraction.md

• P2 · De-identify, dedupe & filter

  Download ↓

  Strip candidate names from the proposals, merge substantively-identical interventions, classify each as researchable / posture / aspirational, and assign stable IDs.

  P2_deid_filter.md

• P3 · Blind evidence research

  Download ↓

  For each researchable intervention, judge what the best available evidence says about whether it works. Run blind — no candidates, no New Mexico, no advocacy. Rates direction (Positive / Mixed / No effect / Negative) and confidence (High / Moderate / Low) with sources.

  P3_research.md

• P4 · Assembly

  Download ↓

  Join the research and the (kept-separate) attribution back together by ID. Validate that every intervention has at least one proposer, every researchable record has a non-null rating, and every source URL is well-formed.

  P4_assembly.md

• Candidate roster (ABQ & Rio Rancho)

  Download ↓

  Reference data: the qualified 2026 NM primary candidates whose ballots include Albuquerque (Bernalillo County) and Rio Rancho (Sandoval County), sourced from the Secretary of State's official export.

  candidate_list_abq_riorancho.md