Lucid
Lucid is an open-source epistemic audit tool that points eight published AI-safety research instruments at your personal AI conversation history.
Core trick
Scores exported chat logs with deterministic sycophancy detectors, then checks each memory claim against the conversations that produced it.
Description
Lucid is an open-source epistemic audit tool that points eight published AI-safety research instruments at your personal AI conversation history. It ingests Claude Code session JSONLs and the Claude.ai export, runs deterministic per-turn scoring across SpiralBench (sycophancy), Sharma paired-exchange detection, SycEval, Jain perspective sycophancy, BeliefShift, the Influence Tactics Protocol, time/model attribution, and a novel memory-corpus consistency check, then produces an HTML report whose narrative sections are written by Claude Opus 4.7 with every factual claim cited to a specific finding or turn ID and validated against the database.
The novel contribution is Module H: it extracts atomic claims from your Claude.ai memories.json and verifies each one against the conversations that produced it, surfacing memories the corpus actually supports versus those it doesn't. No other tool audits AI memory features against the conversations that generated them.
The audit pipeline is a three-phase split — deterministic scoring (preserves calibration reproducibility), Opus 4.7 synthesis writing, Sonnet 4.6 structuring + validation — designed so each model does what it's best at.
Video explanation:
https://notebooklm.google.com/notebook/2daeaaeb-cce5-4bd4-8b34-86279576d0a6/artifact/f5808fe7-6f16-45d0-9510-565d2c4ba2c6
README excerpt
# Lucid [](./LICENSE) [](https://www.python.org/downloads/release/python-3130/) [](#testing) **An epistemic audit for your conversations with Claude.** Lucid reads your Claude Code session history (`~/.claude/projects/`) and your Claude.ai conversation export, runs eight published AI-safety research instruments against the corpus, and produces a structured HTML report. Every finding cites the paper that scored it. Narrative sections are written by Claude Opus 4.7 and validated, claim by claim, against the database that produced them. It runs locally. The only network calls are to the Anthropic API (for classification and synthesis) and to Voyage AI (for the embeddings Module H uses). Your conversation content does not leave your machine for any other purpose. --- ## Why this exists Hundreds of millions of people now do substantial cognitive work inside LLM conversations. Published research — Sharma et al. 2023, Spiral-Bench, SycEval, BeliefShift, and others — shows that this work is systematically distorted by sycophancy, capitulation under pressure, and belief drift, and that the rates involved are not edge cases. In 2025, the major AI labs shipped automatic memory features that synthesize "what we know about you" from your conversation history. These memories are n