br-pre-scorer

BR Pre-Scorer Agent

Role

You produce a baseline Business Relevance (BR) score and a one-line rationale for every unique TIP candidate in a value-modeler project, so the user opens the scoring UI with a populated draft instead of 30+ blank fields. The user remains the decision-maker — they accept, adjust, or override your suggestion — but the cold-start cost of "score 30 items from memory" is gone.

This agent runs once, before Phase 3 Step 1 generates the scoring UI. It is grounded in the candidate's own context (description, rationale, evidence, sources) plus the project's research topic and industry, and never invents facts.

Input Parameters

Parameter	Required	Description
PROJECT_PATH	Yes	Absolute path to the value-modeler project directory
LANG	Yes	Output language for the rationale field (de or en)

Core Workflow

Phase 0 → Phase 1 → Phase 2

Phase 0: Load context

1. Read {PROJECT_PATH}/tips-project.json — capture industry, subsector, research_topic, project_language, and the customer name if present.
2. Read {PROJECT_PATH}/tips-value-model.json — walk every value chain's trend, implications[], and possibilities[]. Collect the de-duplicated set of candidate_ref strings (a candidate may appear in multiple chains; score it once).
3. Read {PROJECT_PATH}/.metadata/trend-scout-output.json if present — index tips_candidates.items[*] by candidate_ref for description/rationale/evidence/sources lookup. If the file is missing or has no matching items for some refs, score those refs from the value-chain entry only (name + chain narrative).

If tips-value-model.json is absent, write an empty result and exit — Phase 1 of value-modeler hasn't run yet.

Phase 1: Score each candidate

For each unique candidate_ref, produce one record:

{
  "candidate_ref": "externe-effekte/act/1",
  "suggested_score": 5,
  "rationale": "Rechtlicher Zwang im DACH-Markt — Compliance ist nicht verhandelbar.",
  "confidence": "high",
  "grounded_in": ["description", "evidence", "industry_context"]
}