Methodology

This page summarises the extraction methodology. Full version in METHODOLOGY.md.

Pipeline

1. Triage—Claude Haiku 4.5 reads each post's caption and assigns a bucket (substantive / contextual / skip) with a stated reason.
2. Extraction—Claude Opus 4.7 reads substantive posts (with audio transcripts via Deepgram) and emits structured records (positions, pledges, actions, endorsements, appearances, quotes).
3. Verification—Action records are cross-referenced against the City of Toronto's public council voting record.

Equal-billing rules

• Candidates listed alphabetically by surname.
• Identical fields shown for every candidate.
• No ranking, no editorial weighting.
• "Awaiting declaration" is a tracked state.

Synthesis (per topic, per candidate)

Each candidate's Said vs. Done tab includes an LLM-generated synthesis paragraph per policy topic, plus a consistency label and cited evidence. The system prompt that generates these is reproduced in full so anyone can audit how syntheses are derived:

You are synthesizing a Toronto mayoral candidate's positions on a single
public-policy topic, based on their public Instagram content.

RULES:
1. Synthesize POSITIONS only. Do NOT characterize the candidate's intent,
   motivation, character, sincerity, or political identity.
2. Every claim about a stance, position, or change must cite at least one
   shortcode from the input records (in `supporting_records`).
3. If you detect a stance change, classify it as "shifted" only when the new
   stance directly contradicts the prior stance. Refinement or specificity
   is "evolving", not "shifted". The `changes` array is required for
   "shifted" and each entry must cite at least 2 supporting records.
4. If fewer than 5 substantive records exist on this topic, return
   synthesis_skipped_reason="insufficient_data" and null fields for
   summary and consistency.
5. Use the candidate's name (not pronouns) in the first sentence of summary.
6. The summary is 80–150 words, plain prose. No headers, no lists.
7. NEVER speculate about future actions, party affiliation, or electoral
   strategy.

OUTPUT: emit a single tool call (emit_synthesis) with the structured
fields. No prose outside the tool call.

• Model: Claude Opus 4.7
• Insufficient-data threshold: 5 records (kind = position/pledge/action) for a topic. Below this, no synthesis is shown.
• Stance-change bar: a "shifted" label requires at least 2 supporting records that demonstrate the contradiction.
• Cache rule: synthesis regenerates only when the underlying records change, the system prompt changes, or the model changes. The cache key is sha256(records) + sha256(prompt) + model.
• Bradford alias: records from @beybradford (his councillor archive account) are merged into Bradford's main dossier and participate in synthesis input.
• Always verify against sources: synthesis is generated by an LLM; cited shortcodes are the primary source. If a synthesis claim doesn't match its cited record, the record wins.

Landing-card consistency dot

Each candidate's card on the landing page shows a small dot encoding overall consistency:

• Green — all topics with synthesis data are consistent.
• Yellow — at least one topic is evolving; none are shifted.
• Red — at least one topic shows a contradicted prior stance.
• Gray — fewer than 3 topics have synthesis data.

Corrections

Open an issue on GitHub if you spot an error.