Verdict

Verdict is a 197 KB local model that classifies CLI tool call outcomes as success, failure, blocked, or unknown.

Verdict is a semantic outcome layer for coding agents, CI runners, and tool-using systems. It reads a CLI tool call along with its exit code, stdout, and stderr to return success, failure, blocked, or unknown. Inference runs locally on CPU using pure Python standard library with no external dependencies. The architecture uses hashed n-grams and a multiclass logistic regression model with 196,608 int8 weights. It is designed for developers building agentic workflows who need to distinguish between actual task completion, environment blockers, and failures.

Verdict at a glance

Best for: Software engineer
Runs on: CLI, Web
Pricing: Free
Released: November 2026
Category: AI coding tool
Website: github.com/balanov/verdict

Key features

Reads command text, exit code, stdout, and stderr
Returns success, failure, blocked, or unknown labels
Runs locally on CPU with pure Python standard library
Contains 196,608 int8 weights in a 197 KB file
Processes inputs bounded to 12,288 characters per stream
Runs in the browser using quantized weights without uploading logs

Use it for

Classify CLI tool call outcomes for coding agents
Evaluate CI runner outputs beyond basic exit codes
Distinguish environment blocks from actual command failures
Analyze command line execution logs locally

Alternatives to Verdict

jevgrep — Reduces coding agent costs by providing targeted, relevant repository context.
Judge0 — Provides secure, sandboxed code execution for over 60 programming languages.
Command Code — Repairs open model tool calls and learns developer patterns automatically.
pstack — Improves AI coding agent outcomes with rigorous workflows.
Agent Trials — Test if an agent fixes your bugs in your stack at a sane price before deployment.
Rigmark — Measures local AI performance with reproducible, agent-like workloads.
OpenCode — Works with 75+ model providers, local models, and existing subscriptions
TryCase — Runs end-to-end tests and reports video evidence on GitHub without writing code.
Strata — Runs large AI models locally on consumer gaming PCs without server reliance.
Kora Doctor — Surfaces waste across tool usage, context growth, and repeated inference without an account.
Crit — It keeps human feedback localized and precise for AI agent code iterations.
Nouli — It gives stateless decision models a persistent memory across branching sessions.

All Verdict alternatives