Intent is attributed from signals. The service reads the signals, states the attribution, and grades its own confidence.
There is no hidden score. The assessment applies a written model of intent and divergence, and every field in the response is a reading of that model. This page walks through the model, the procedure, and what comes back.
You are Party A. They are Party B.
You tell the service as much or as little about yourself as you like — your situation, your intent, your characteristics. The counterpart is characterized only from evidence: what they said, a transcript of the exchange, and optionally who they are, so their public record can be discovered. The question the service answers is one-directional: what does B intend toward A?
a.intent
a.characteristics
— any or all optional —
b.conversation
b.identity → discovery
The reverse reading, A's intent toward B, is returned only when you ask for it, and is marked presumptive whenever your own intent was not directly stated. The two directions are evaluated independently.
Divergence, defined.
Two intents diverge when a target of one is not co-realizable with a target of the other over a referent they share — when no single state of the world satisfies both.
Everything else is built from that. An intent has targets (states a party acts to bring about) and conditions (the states under which it holds those targets — including the engagement it entered, and its own capacity to form intents at all). Where a divergence lands in that structure is what the response reports.
Shallowest: one target instance. Deeper: a class of your targets. Deepest: a condition you hold your targets under, up to your capacity to hold any. Depth is read from the structure, not from tone or stakes.
Entering a negotiation admits certain divergences: a hard bargain is invited. A divergence outside what you admitted is uninvited — and one engineered to appear invited is named as such.
Three stages, and the first one can stop.
Does the evidence cover the things both parties' intents touch? If not, the answer is insufficiency — with the components that would need eliciting. This is an output, not an error.
Over the shared referents, is there a target of B that no state of the world satisfies together with a target of A? divergence or no_divergence.
For each divergence: the object, its depth, whether it was invited — and the provenance of the whole reading, including how deeply the evidence was instrumented.
The full reading, one categorical value at a time.
Each axis takes one value from its catalog — or insufficient or not_applicable — with a confidence and the evidence sentence that supports it. The list is flat; the groups are for reading.
directional_configuration
object_depth
referent_locus
referent_temporality
target_generality
frame_integrity
revocability
response_policy
demand_conditionality
response_loop_valence
resolution_mode
intent_structure_provenance
source_identity_concordance
signal_availability
signal_cost_structure
signal_materiality
signal_routing
signal_tempo
reception_scope
referent_concentration
attribution_veridicality
attribution_congruence
attribution_regime
instrumentation_depth
exploration_state
pair_resolution_state
enactment_progression
condition_trajectory_provenance
trajectory_aperture
temporal_extension
manifestness
engagement_history_density
party_scope
pair_structure
Every response grades itself, and tells you how to improve it.
Overall confidence. Whether B's evidence was corroborated or single-channel. How deep the instrumentation went. And a list of limiting factors in plain language — a sparse A-side lowers confidence and says so.
On request: plain-language questions, in your terms, that would sharpen the read — and why each one matters. Your agent can put them to the counterpart, or to you.
Name the counterpart and set search_depth. Web discovery gathers their public statements, rates each source's reliability, and folds what corroborates into the attribution.
A frontier model, a versioned reasoning definition, and the work shown.
The assessor assembles the conflict model, the axis catalog, and the response contract into a system prompt and runs it on Claude (currently Opus 5) with extended reasoning. It parses and validates the result against the catalog before returning it: an axis value outside its vocabulary never reaches you.
What this is not: a verdict. The reading is decision support for your agent or your reviewer. It recommends nothing and acts on nothing.