Lead Quorum: Multi-Agent Scorer That Abstains on Disagreement

Lead Quorum tackles a common flaw in lead-scoring demos: a model spits out a confident number with no traceable reasoning. Instead, it has two separate Gemini models (gemini-flash-latest and gemini-2.5-flash-lite) independently extract signals from raw notes, each running as its own Cloud Run microservice and reached over Google's A2A protocol via RemoteA2aAgent. Scoring itself, and the comparison between the two readings, is handled by deterministic code rather than an LLM — every rule emits its points and reasoning on the same line, and tests verify the reasons always sum to the score.

The key design choice is that when the two independent readings disagree on which rules actually fire, the system returns an 'EXCLUDED' verdict instead of guessing, naming exactly which rule flipped. Comparing in score-space rather than raw values lets minor numeric drift be distinguished from a genuine contradiction. Only two of the five agent roles are LLMs; the rest is plain, testable code, which keeps costs down and reasoning auditable. The project is open-sourced under MIT with 16 tests, including a script to run the full distributed pipeline locally.