AI product Open source
Verdict is a local semantic verifier for AI-agent tool calls and command-line runs. It reads the command, exit code, stdout, and stderr, then classifies the observed outcome as success, failure, blocked, or unknown. Its zero-dependency inference model uses hashed character 3–5-grams and word 1–3-grams with multiclass logistic regression; the released weights are quantized int8 and run on the CPU without network calls or telemetry. The repository provides a Python API and CLI, with bounded head-and-tail clipping for long output streams and an optional browser demo using the same quantized weights.
Verdict is an advisory classifier rather than authoritative proof of task completion, code correctness, safety, or truthfulness. The project describes its benchmark as based on synthetic, English-biased examples and recommends deterministic checks and authoritative state verification for consequential actions. The code and model weights are Apache-2.0 licensed, while the synthetic dataset is CC0-1.0.
1 use taken from transcripts — each links to the moment in the video.
A tiny classifier that reads what happened when an AI agent ran a command and labels the run as success, failure, blocked, or unknown. The whole model weighs about 200KB and answers in about half a millisecond with no network calls.
1 in the library.