Skip to content

Overview

This section documents the public API of credencegraph-extract.

Main Sections

  • MCP server – The agent tools over MCP on stdio, for an agent harness to call.
  • Graph tools – Classifying chunks, adding premises, running checks and finishing a run, through the assembler and the check runners.
  • Agent-tool runs – The run directory an agent's tool calls build: the call log, the lock, crash recovery and replay.
  • Assembler – Applying classify and premise records through the core's edits.
  • Checks – What a check proposal must be, the four check types, check nodes and verdicts.
  • Evaluators – Exact arithmetic, symbolic identity and numeric sampling, over whitelisted expressions.
  • Citation sources – Reading a citation's source, and deciding whether it is tied to the reference.
  • Fetch – Fetching a URL an agent names from public addresses only.
  • Verify – Running the deterministic checks of a run again, and reporting rubric versions reused with different values.
  • Earlier runs – Reading the runs of format 1 that earlier versions made by calling a model, for verify.