-
Notifications
You must be signed in to change notification settings - Fork 0
Home
Architectural decision tracking with prediction validation and outcome gates.
A Model Context Protocol (MCP) server that helps agents and teams:
- Record technical decisions (problem, solution, alternatives, technologies)
- Predict measurable outcomes (latency, cost, scalability, etc.)
- Measure results in production
- Validate predictions against reality
- Learn from accuracy — improve future recommendations
All backed by an append-only JSONL log. No external database, no schema migrations.
Decision-making systems typically leak the feedback loop:
- Agent chooses solution → confirms choice → implementation starts
- Results come in → but nobody asks: "Was the prediction correct?"
- Next decision starts → learns nothing from the last one
This server closes that loop automatically. It nudges you in-band (inside tool responses) when predictions are pending, keeping closure rates high.
→ Quick Start Guide (5 minutes to your first decision)
A choice made at a point in time:
- Problem being solved
- Chosen solution
- Alternatives considered
- Technologies involved
A testable claim linked to a decision:
- Type (LATENCY, COST, SCALABILITY, RELIABILITY, etc.)
- Predicted value (e.g., "p99 < 200ms", "cost < $100/month")
Measured reality:
- Actual value observed
- Measurement source (MONITORING, LOGS, BENCHMARK, etc.)
- Accuracy score (0-100)
In-band nudge embedded in tool responses:
- Appears automatically if predictions from this session lack outcomes
- No external reminders needed — nudge is co-located with the decision
- Dramatically improves closure rates (3.8% → 14.5% observed)
Automatic aggregation by technology:
technology: duckdb | successful: 12 | failed: 1 | avg_accuracy: 91.2% | confidence: HIGH
technology: redis | successful: 8 | failed: 3 | avg_accuracy: 78.3% | confidence: MEDIUM
Use this before recommending a technology — learn from what's actually worked.
Made for AI agents. Built for teams. Learn from every decision.