AlexandrAI AGENTS.md
agents-md

LMSSEP Prediction MCP Server Agent Guide

Project-specific AGENTS.md guidance for the LMSSEP Prediction MCP server, super-forecasting flow, 10-outcome probability model, TOON storage, and report generation.

AGENTS.md

Guidance for coding agents working on LMSSEP Prediction MCP Server.

Project Overview

The prediction server implements a structured forecasting workflow inspired by super-forecasting practice. It initializes a forecast, guides the user through analysis steps, requires multiple probabilistic outcomes, finalizes recommendations, and emits session plus report artifacts.

Key paths:

  • predicting/README.md: quick start and workflow.
  • predicting/COMPREHENSIVE_GUIDE.md: deeper reference.
  • predicting/main.py: server entrypoint.
  • predicting/requirements.txt: dependencies.
  • predicting/configs/: local model/provider config.
  • output/predictions/: sessions and reports.

Setup Commands

bashcd <project-root>/predicting
pip install -r requirements.txt
python3 main.py

Use placeholders such as <FORECAST_QUESTION>, <SESSION_ID>, <LLM_PROVIDER>, and <API_KEY> in examples.

Forecast Rules

  • Initialize each forecast with a clear, time-bounded question.
  • Step through the structured analysis sequence rather than jumping directly to final advice.
  • Step 4 requires exactly 10 outcomes when the 10-outcome framework is active.
  • Outcome probabilities must sum to 100 percent.
  • Action recommendations must be specific enough to execute or monitor.
  • Forecast reports should state assumptions, uncertainties, confidence, and scenario distribution.

Storage And Output

  • TOON session files are used to reduce token and file size overhead.
  • Markdown and HTML reports are generated after finalize.
  • Do not hand-edit session files unless repairing a known schema issue.

Testing

  • Test provider config validation separately from forecasting logic.
  • Add checks for probability-sum validation, missing step data, malformed session IDs, and finalize preconditions.
  • Verify report generation for both simple categorical forecasts and trajectory forecasts.

Security

  • Keep LLM API keys in local config or environment only.
  • Do not publish forecast sessions that include private investment, customer, or business strategy data.
  • Redact source citations or raw grounding metadata if it contains private browsing context.