Pragmatic Inference
TL;DR
Pragmatic inference is Grice’s subject: what is meant beyond what is said. On his account of implicature, speakers convey more than their words literally carry and hearers routinely recover it — and regulated work requires that recovery. A customer who writes “I never authorized this and I’ve called three times” asserts an error claim and a service failure without naming either, and a system reading only the literal text will mis-route it. The same faculty is also where systems invent things, because inferred content presented as stated content is how an assumption becomes a fact of record. So the graph runs both sides: the literal content is quoted verbatim, candidate readings are generated with the cues that licensed them, and context decides which hold. The separation gate keeps every inference labelled and out of the quoted record; the confirmation gate sends an outcome-changing inference back to the person or to a reviewer.
Four lines in the servicing queue
The message is four lines long and contains none of the words the routing rules look for. No “dispute.” No “error.” No “complaint.” What it says is that the customer never authorized the charge and has called three times. Read literally — which is what a keyword router does — it is an inquiry about a charge, and it goes to general servicing, to wait its turn behind requests for statement copies.
A seasoned representative reads the same four lines and sees two things the customer did not write: an assertion of an unauthorized transaction, which is an error claim and starts a clock; and three unremedied prior contacts, which is a service failure with its own reporting consequences. Neither reading is a stretch, and neither is optional — reading only the literal text is the error here, not the inference. But the same faculty, unmarked, is how the file later reports a call on the 3rd when what the customer wrote was “three times” — a reasonable reading hardened into a quoted fact.
An unguided model does both at once and marks neither: it infers fluently, then reports the inference in the same voice as the transcript. The pragmatic graph must hold the said and the meant in separate columns, name the cue behind every reading, and route the inference that changes the outcome to a person.
How the graph works, step by step
Retrieve the literal content. The said is captured first and quoted exactly, with locations: this sentence, in this channel, on this date. It is the column nothing downstream may rewrite, and establishing it before any reading happens is what makes the separation checkable later — a verbatim record is the only thing an inference can be held apart from.
Generate candidate implicatures. Plural, deliberately. “I never authorized this” supports a reading as an assertion of error; “three times” supports a reading of prior unremedied contact, and, in some contexts, of a deadline already running. Each candidate is emitted with the cue that licensed it — the span, the channel, the sequence — and competing readings of one cue are carried forward together rather than collapsed into the most fluent.
Evaluate support from context. Account history, prior contacts, the channel and its conventions, what else the message does and does not say: the node scores each candidate as supported, unsupported, or genuinely ambiguous. Ambiguity is a permitted output; resolving it because one reading writes better is the failure this node exists to interrupt.
Verify the marking and gate it. Every admitted inference is labelled, attached to its cue, and kept out of the quoted record — the separation gate’s check. The confirmation gate then looks at consequence: a reading that would start a clock, reclassify the matter, or trigger a right is confirmed with the person or escalated. The unstated claim is still acted on — that is the point of reading pragmatically — but as a question put back to the customer, not as a fact supplied on their behalf.
The gates, operationally
| Gate | What it checks | Fails when |
|---|---|---|
| Cue gate | Each candidate reading names the span or contextual feature that licensed it, and rival readings of the same cue are carried forward | A reading appears with nothing behind it — fluency supplying content the message never carried |
| Separation gate | Inferred elements are labelled as inference, bound to their cue, and held apart from the quoted record, which stays verbatim | The paraphrase becomes the record — the file reports what the person meant as though it were what the person wrote |
| Confirmation gate | An inference that changes the outcome — starts a clock, reclassifies the matter, triggers a right — is confirmed with the person or escalated, never assumed | The system acts on its own reading without checking it, and by the next review the assumption reads as a fact of record |
Where it fits — three use cases
1. Complaint and dispute intake
The unstated claim and the unstated deadline are what intake most often misses, and both are pragmatic rather than lexical: people describe what happened to them, not the category it belongs to. A graph that generates candidate readings with cues attached routes on what was meant while leaving a trace of which words were taken to mean it — which is how a wrong routing gets found rather than repeated.
2. Reading regulatory text and guidance
Obligations are often implied by structure rather than stated outright: an enumerated set of exceptions implies a general rule, a required sequence implies timing, an illustration implies the class it illustrates. Practitioners make these readings constantly and correctly. Marking them as readings, with the structural cue named, keeps an interpretive step from being cited as though it were the text — which matters in the review that asks where a requirement came from.
3. Outbound correspondence
The same strategy runs in the other direction, on your own letters. An adverse-action notice, a dispute outcome, a collections message: what does it imply beyond what it says — about the customer’s options, about finality, about whether writing back is worth it? Checking a draft for its implicatures is unglamorous, and it is the territory where fair-treatment findings are made.
When to reach for it
Reach for pragmatic inference wherever the input is language from a person under no obligation to use your vocabulary — intake channels, transcripts, free-text fields — and wherever your own outbound language carries consequences. It sits close to steelmanning, which asks for the strongest version of what someone argued; this asks the prior question of what they were doing with the words.
The honest limit is built into the phenomenon. Implicature is defeasible by nature: the same cue supports different readings across contexts, registers, languages, and populations, and a reading obviously right for one customer is an imposition on another. That is exactly why the strategy marks rather than resolves — it does not claim to recover what the person meant, only to make each reading, its cue, and its rivals inspectable and challengeable, and to route outcome-changing inferences to a human. It reduces and interrupts the drift from inference to record; it does not remove the ambiguity that made inference necessary.
- Implicature
- What a speaker conveys beyond the literal content of their words — licensed by a cue, defeasible by context, and routinely load-bearing in regulated correspondence.
- Said / meant separation
- Two columns that never merge: the quoted record stays verbatim, and every inferred element travels labelled, with the cue that licensed it.
- Outcome-changing inference
- A reading that starts a clock, reclassifies a matter, or triggers a right — the confirmation gate’s trigger, put back to the person or escalated instead of assumed.