Read what the other party wants from you — before you agree to it.
Send the counterpart's own words. Malintent Detector attributes their intent toward you from those signals — whether it diverges from yours, how deep the divergence runs, whether you invited it — and reports how sure it is, and why it isn't surer.
B's target — a 24-month commitment with exit clawbacks and proprietary export — is not co-realizable with A's stated target of a low-risk pilot. The object is A's ability to leave, a condition under which A holds every other target, not the pilot itself.
Your agents negotiate, buy, and settle disputes with counterparts you don't control. The party that costs you the most is usually polite.
Moderation and safety filters score vocabulary, one message at a time. They have nothing to say about a counterpart who breaks no rule in any single turn while steering toward a commitment you cannot exit. That read has to come from the structure of what they say — conditionality, what they deflect, what leaving would cost — not the tone.
Attributed from the counterpart's own signals: statements, a transcript, and optionally their public record.
The headline is their intent toward you. The reverse read is available on request and labeled presumptive.
"Not enough evidence" is a first-class answer, and every response names what limited it.
Send what you know and what they said. Get back a structured read you can act on.
A reading, its provenance, and its limits.
Nothing is a bare number. Each field is grounded in a stated model of intent, and the model is versioned with the service.
Three outcomes. A signal base too thin to decide is reported as such, never rounded to a verdict.
A single target, a class of targets, a condition you hold your targets under, or your capacity to hold them at all.
A hard bargain inside a negotiation you entered is invited. One manufactured to look that way is not.
Overall confidence, whether evidence was corroborated or single-channel, and the factors that limited the read.
Categorical values across relation structure, engagement conduct, signal properties, and temporal trajectory — each with the sentence that supports it.
Plain-language questions that would sharpen the read, and optional web discovery of the counterpart's public record, weighed by source reliability.
Call it wherever your agent decides whether to proceed.
Assess the counterpart's transcript before your agent signs, pays, or grants access. Gate on relation and depth.
Re-assess as the conversation grows. Temporal axes track whether the counterpart's demands are escalating or settling.
Name the counterpart and set a search depth. Discovery folds their public statements into the attribution.
All processing, including the model call, runs inside the service on Google Cloud. Your client holds one API key for one project. It never holds a model provider key, and request content is processed in memory and not retained by default. Security and data handling →
A stated model, applied by a frontier reasoning model, with the work shown.
The assessment applies a written conflict model — a derivation chain from primitives (party, intent, signal, referent) to a single core relation: two intents diverge when no state of the world satisfies both over something they share. Every field in the response is a reading of that model. The reasoning definition is versioned, and every response carries the version and the model that produced it. With include_io, you can see the exact prompt, output, and reasoning summary for a call.
Run your first assessment in minutes.
Sign up with Google, GitHub, or email. Add a card — you're billed only for what you use, starting from the first dollar. No stored balance, no minimum.