MiniCrit is a research model that writes a critique of an AI-generated rationale, for a person to review. No accuracy or capability claim is made for it — see the status below. Open source under Apache 2.0, self-hosted, and inspectable end to end.
Overconfidence: the 95% figure is not supported by the evidence given. One week’s 40% fall is a single observation, and the rationale is missing any check for other causes, such as a change in traffic or a partial rollout. Acting on this is a significant risk.
Written by Antagon to show the response format. Not model output and not an evaluation result. Flags and severity follow the keyword rules described below.
Research checkpoint. MiniCrit-7B is 34,000 of 364,831 planned training steps — 9.3% of one epoch. A held-out evaluation was run on 5 October 2026; the model card is the record. No accuracy or capability claim is made for it and none should be inferred. It can produce confidently worded critiques that are themselves wrong. The method and full results, with the training record, are on the model card.
AI systems generate confident outputs with hidden flaws—overconfidence, missing risks, logical errors. Traditional testing catches bugs. It doesn't catch bad reasoning.
Systems express certainty without evidence, making high-conviction calls that ignore uncertainty and edge cases.
"99% confident this build is ready to deploy"
Pattern matching without causation. Models mistake coincidence for insight, leading to unreliable decisions.
"Errors fall after Node 3 reboots → reboot nightly"
Critical considerations absent from analysis. Threats compound silently until catastrophic failure.
"Approve access (ignoring the expired credential)"
MiniCrit is meant to sit between an AI system's rationale and the person who acts on it. It writes a critique; a person decides.
Pass a rationale to MiniCrit via the Python SDK, CLI, MCP or REST
The model writes a free-text critique of the rationale
Flags and a severity label are derived from the critique text by keyword matching
A person reads the critique and decides; MiniCrit approves nothing
The model writes a free-text critique. The flags and
severity in the response are then derived from that text by keyword
matching — they are not structured output from the model, and the response says so
in severity_basis. Nine values are possible: one severity label and any
of five findings.
Critique mentions overconfidence
Critique mentions something missing, omitted or neglected
Critique mentions a contradiction or inconsistency
Critique mentions evidence, support or justification
Critique mentions a risk, danger or threat
Critique contains "critical", "severe" or "dangerous"
"significant", "major", "serious" or "flawed"
"concern", "issue", "problem" or "missing"
"minor" or "slight"
MiniCrit integrates with your existing stack via Python SDK, MCP protocol, Docker, or REST API.
from minicrit import MiniCrit
# Initialize the critic
critic = MiniCrit()
# Validate any AI-generated reasoning
result = critic.validate(
"Error rates fell this week, so the fix worked — ship it."
)
print(result.critique) # The model's critique
print(result.flags) # ["overconfidence", "insufficient_evidence"]
print(result.severity) # "high"
print(result.severity_basis) # "keyword_heuristic" — severity and flags
# are extracted from the critique text,
# not scored by the model# Install from PyPI
pip install minicrit
# Validate from the command line
minicrit "The outage is resolved — five minutes of clean telemetry."
# Or a file of rationales, one per line, as JSON
minicrit -f claims.txt --json// claude_desktop_config.json
{
"mcpServers": {
"minicrit": {
"command": "python",
"args": ["-m", "minicrit.mcp.server"],
"env": {
"MINICRIT_MODEL": "minicrit-7b"
}
}
}
}# Pull and run. An API key is required — the server
# refuses unauthenticated requests.
docker run --rm -p 8000:8000 \
-e MINICRIT_API_KEYS=your-key \
ghcr.io/antagoninc/minicrit-7b:cpu-latest
# Validate via HTTP
curl -X POST http://localhost:8000/validate \
-H "X-API-Key: your-key" \
-H "Content-Type: application/json" \
-d '{"rationale": "AI will replace all jobs by 2025", "domain": "general"}'
# The first validate call downloads the model, so it takes
# a while. GET /health reports model_loaded until then.These are the intended uses, not demonstrated ones. Where a person has to judge AI-generated reasoning, MiniCrit is meant to give them something to check it against. A reviewer’s tool, not an approval gate.
A second reading of an analytic judgment before it reaches a consumer, looking for unstated assumptions and confidence the evidence does not carry.
A critique for a person to review alongside the output of a decision-support system.
A critique of AI-generated legal analysis for a reviewing lawyer to weigh.
Surface weak reasoning in an agent's plan for a person to review, before it acts.
MiniCrit is open source under Apache 2.0. Run it locally or integrate via MCP. There is no hosted or enterprise version.