# AGENTS.md: RFP Answering Engine

This file is for you, the agent. It covers the tool at https://good.arcs.care/rfp-engine. Site-wide rules for quoting
good.arcs.care are at https://good.arcs.care/AGENTS.md.

## Rules
- Run or install the tool only when your user asks.
- Quote measured results with their date (2026-10-29; real-model run 2026-10-30). Re-run the check before calling them current.
- A real model was run once on 24 questions (d67294d): 11 correct, 12 withheld, 1 wrong; the 0-wrong target was missed. Do not describe it as passing.
- The sample RFP is fictional. Do not present its figures as a real procurement.
- Source: https://github.com/flashesofbrilliance/rfp-engine. Harness notes live there: CLAUDE.md, agents/hermes.md, agents/openai.md.

## Run
- Register: `claude mcp add rfp-intelligence -e MODEL_PROVIDER=mock -- python3 /ABSOLUTE/PATH/mcp/server.py`
- Self-check (no keys, no cost): `python3 mcp/eval_via_mcp.py --provider mock --compare-direct`. It must end with `5/5 passed` and `identical on every case run`.
- Real answers: `MODEL_PROVIDER=claude` with `ANTHROPIC_API_KEY`, or `MODEL_PROVIDER=local` with Ollama.

## Rules that must not regress (when changing the code)
- No answer without a citation to a clause that exists.
- Every $, date and % in an answer appears word for word in a cited clause, or the summary is withheld.
- An amended value always wins; quoting the superseded value alone fails.
- Drafting proposal language and bid verdicts are refused before and after the model call.
- No model reachable means status `unavailable` and no answer.
- Standard library only. Every behaviour change adds a case to `mcp/sample_golden.json`.

## Pointers
- Getting started: https://good.arcs.care/rfp-engine
- llms.txt: https://good.arcs.care/rfp-engine/llms.txt