deedchain — report-fidelity demo

This page runs the same deterministic scoring engine as the Python package — ported to ~250 lines of dependency-free JavaScript (deedchain.js). No server, no build, no LLM judge. It scores an agent’s final report against recorded ground truth, independent of whether the task succeeded.

1. Self-check (frozen cases)

Every case below must match the Python reference engine exactly. Green = engine parity holds in your browser.

Case	Report	Expected quadrant	Expected truthfulness	Result

2. Try it yourself

Report text (the agent’s final answer) Gold metrics (JSON, as recorded by the reference harness) Gold outcome objective success (measured by the harness, not by the report)

(results appear here)

What this is measuring

A success-rate benchmark only measures whether the agent completed the task. deedchain measures the missing half: whether the agent told the truth about what it did. A report that says “Found 9 products” when the page held 5 gets flagged as hallucinated even though the task succeeded.

This is the demo. The full project — the frozen 24-task suite, the open AgentRun format, agent adapters, leaderboard + CI, and the MCP server — lives in the repository.