SCORESecurity & safety
Jev Decision Model for Real-Time Moderation
Chris@Chrisondesk
0:20Jev is a 'System One' decision model that provides calibrated probabilities and label distributions for tasks like comment moderation, offering zero output cost and zero hallucination by design.
TakeIt returns calibrated probabilities, label distributions, rubric scores.
I moderated 100 comments in real time for $0.002171.
A fifth of a cent. But this isn't a post about the demo. It's about the choice you're making when you add AI to a product: LLM or decision model.
The model is Jev, by TypeSafe AI. Not a chatbot. A "System One" model. You send a state (the thing you're judging) plus typed questions. It returns calibrated probabilities, label distributions, rubric scores. No prompt. No prose.
Same job, both ways.
Comment moderation with a chat LLM:
1. Write a system prompt: "You are a moderation classifier. Respond with ONE JSON object. No markdown. No prose."
2. Hope it complies. It usually appends a sentence anyway
3. Parse the JSON. Handle code fences. Write retry logic for malformed output
4. Read "confidence: 0.8" and pick a threshold anyway
5. Pay for every output token it spent thinking out loud
Comment moderation with Jev:
1. Send the comment as state, plus 6 typed questions
2. Get typed answers: contains_profanity → 0.93, severity → offensive (88%)
3. Apply thresholds in code. Done
The schema is the API. Nothing to parse, nothing to retry.
And two things a chat model structurally cannot give you:
Zero output cost.
Jev doesn't generate anything. It decides. There is no completion to bill, so output tokens are $0. You pay only for input: $0.042 per 1M tokens. An LLM doing this job charges you for the JSON it made you beg for.
Zero hallucination.
An LLM can invent a field, a fact, a citation, a JSON key you never defined. Every output is fresh text, so every output is a new chance to confabulate.
Jev's answer space is closed. It picks from the options you defined. The worst it can do is pick the wrong option, at a probability you can measure. There is nothing to invent.
Which is where RLCD comes in. Calibration.
An LLM's "confidence: 0.8" is a token it generated. Nothing in training ever checked it against outcomes.
Jev is trained with RLCD, TypeSafe's calibration method. Calibrated numbers are frequencies: if Jev scores 100 comments at 0.7 for targeted abuse, about 70 actually are. The number is a measurement, not a mood.
That's what makes thresholds real engineering. Mine: threat ≥ 0.85 blocks, 0.6 to 0.85 goes to human review, below passes. With an LLM you're tuning gut feel against a number with no defined meaning, and finding out months later from your mod queue.
Every answer also ships the full probability distribution, so you can verify calibration on your own data. No benchmark trust required.
The math: ~500 input tokens per comment, 100 comments, $0.002171 total. Two hundredths of a cent per comment. 46,000 comments per dollar.
And it runs between render and readability. Comments arrive clean or already blurred. No queue, no batch job. If the API times out, comments fail safe to human review, never silently approved.
Models for judgment, code for decisions. Calibrated judgment in, deterministic policy out, every removal auditable.
If you're prompt-engineering a chatbot to emit JSON for classification, you're paying more for worse numbers.
@typesafeai #Jev
Related builds
