Each sensor watches a different part of your AI workflow and warns you when something is going wrong — before the wrong answer reaches your user. Pure measurement over the traces your pipeline already emits: deterministic, reproducible, CPU-only — the opposite of the stochastic model they watch, and cheap enough to run on every request. Open source, MIT-licensed, free forever.
Start with this one. Plugs into Claude Desktop, Cursor, Cline, Windsurf, Claude Code:
npx contrarianai-context-inspector --install-mcp
Watches the bell curve of your AI's working memory in real time. Catches off-topic stuff sneaking in, important details lost when the AI compresses, and slow drift away from the original task — before the output goes bad. One command installs it into any major AI coding tool.
For AI systems that look things up in a knowledge base (RAG). Scores every search result, draws the bell curve, and checks rank ordering, redundancy, score calibration, and whether the results split into two clumps. Catches silent bugs standard accuracy metrics can't see. Tracks the actual quality of search results at r = 0.999 correlation, without ever seeing the ground-truth labels.
For AI workflows where one or more agents call external tools. Scores each call for relevance and aggregates across the whole session. Catches tool calls that quietly returned wrong data, AI fixated on calling the same tool over and over, tools returning bloated responses, mismatches between what the tool sent and what the AI expected, and cascading failures across multiple agents.
Given the history of bell curves over time, predicts what the next bell curve should look like if everything stays healthy. When the actual next bell curve doesn't match the prediction, something has changed — and the gap itself is an early warning. Uses two classic math techniques (Adams-Bashforth-Moulton from the 1800s, or a Kalman filter from the 1960s). On one benchmark this catches drift 17 turns before the simplest output monitor does.
Takes the output from the other four sensors and writes a single audit report you can send to your team or your client — Markdown, HTML, or JSON. Summarizes silent bugs across all four sensors, ranks findings by impact, and lists recommended fixes for each bug type. This is the code that powers the $2,500 Rapid Audit report — the deliverable an engineer would write, generated automatically from the sensor data.
Each tool is an independent command-line tool and MCP server (the new standard for AI tool integration). They share the same data shape so you can chain them without writing glue code: run the sensors against your data, pipe the outputs to the forecaster, send everything to the report writer, get a single PDF.
A hosted version — Bell Tuning Cloud — that aggregates them into one dashboard with time-series, alerting, and downloadable reports is in private beta. Request early access →
The free tools tell you when the bell is wrong. The paid audit tells you why and fixes it. 48-hour turnaround, fixed price, money-back guarantee.
Buy Rapid Audit — $2,500 Free 30-min call first