0. Outline:

Layer 0 — Natural Language Interface Substrate

0. Front Matter
   0.1 Purpose
   0.2 Scope
   0.3 Core thesis
   0.4 Audience
   0.5 How to read this document
   0.6 Layer 0 invariants

1. Meaning as a Layered Phenomenon
   1.1 Surface form
   1.2 Lexical meaning
   1.3 Constructional meaning
   1.4 Compositional semantics
   1.5 Pragmatic meaning
   1.6 Discourse meaning
   1.7 Social and interactional meaning

2. Indeterminacy and Compression
   2.1 Ambiguity
       2.1.1 Lexical ambiguity
       2.1.2 Syntactic ambiguity
       2.1.3 Scope ambiguity
       2.1.4 Referential ambiguity

   2.2 Vagueness
       2.2.1 Fuzzy boundaries
       2.2.2 Gradable terms
       2.2.3 Context-relative thresholds

   2.3 Underspecification
       2.3.1 Missing parameters
       2.3.2 Recoverable omissions
       2.3.3 Defaults and assumptions
       2.3.4 User-side economy of expression

   2.4 Prototype-Based Categories
       2.4.1 Category resemblance
       2.4.2 Central and peripheral cases
       2.4.3 Domain-specific prototypes

   2.5 Metaphor and Conceptual Transfer
       2.5.1 Technical metaphor
       2.5.2 Evaluative metaphor
       2.5.3 Structural metaphor

   2.6 Summary: Why Natural-Language Meaning Is Partial, Graded, and Compressed

3. Context, Reference, and Grounding
   3.1 Implicit Context
       3.1.1 User context
       3.1.2 Task context
       3.1.3 Domain context
       3.1.4 Organizational context
       3.1.5 Preference context

   3.2 Deixis
       3.2.1 Person deixis
       3.2.2 Time deixis
       3.2.3 Place deixis
       3.2.4 Discourse deixis
       3.2.5 Social deixis

   3.3 Coreference
       3.3.1 Pronoun resolution
       3.3.2 Named-entity resolution
       3.3.3 Repeated descriptions
       3.3.4 Ambiguous referents

   3.4 Ellipsis
       3.4.1 Sentence-level ellipsis
       3.4.2 Turn-level ellipsis
       3.4.3 Task-state ellipsis
       3.4.4 Recoverability conditions

   3.5 Presupposition
       3.5.1 Entity presuppositions
       3.5.2 Event presuppositions
       3.5.3 Authority presuppositions
       3.5.4 State presuppositions

   3.6 Interface Grounding
       3.6.1 Selected object
       3.6.2 Active document
       3.6.3 Cursor position
       3.6.4 Viewport state
       3.6.5 Highlighted text
       3.6.6 Current UI focus

   3.7 Summary: How Meaning Depends on External, Prior, and Situated State

4. Communicative Force and Intent
   4.1 Pragmatics in the Broad Sense
       4.1.1 Meaning as use
       4.1.2 Literal meaning versus intended meaning
       4.1.3 Contextual inference

   4.2 Speech Acts
       4.2.1 Request
       4.2.2 Command
       4.2.3 Question
       4.2.4 Suggestion
       4.2.5 Warning
       4.2.6 Correction
       4.2.7 Confirmation
       4.2.8 Refusal
       4.2.9 Evaluation

   4.3 Indirect Speech Acts
       4.3.1 Ability question as request
       4.3.2 Observation as implied request
       4.3.3 Critique as revision request
       4.3.4 Preference as constraint

   4.4 Implicature
       4.4.1 Relevance implicature
       4.4.2 Scalar implicature
       4.4.3 Consequential implicature
       4.4.4 Domain-specific implicature

   4.5 Relevance Filtering
       4.5.1 Task-relevant interpretation
       4.5.2 Domain-relevant interpretation
       4.5.3 User-goal-relevant interpretation
       4.5.4 Salience and suppression of irrelevant meanings

   4.6 Action Fan-Out
       4.6.1 Single utterance to single action
       4.6.2 Single utterance to parameterized action
       4.6.3 Single utterance to workflow
       4.6.4 Single utterance to multi-step orchestration

   4.7 Summary: How Utterances Become Communicative Acts

5. Discourse State and Conversational Mutation
   5.1 Discourse Structure
       5.1.1 Sequence
       5.1.2 Contrast
       5.1.3 Cause
       5.1.4 Condition
       5.1.5 Exception
       5.1.6 Elaboration
       5.1.7 Summary

   5.2 Discourse Markers
       5.2.1 however
       5.2.2 therefore
       5.2.3 actually
       5.2.4 instead
       5.2.5 unless
       5.2.6 for example
       5.2.7 in short

   5.3 Repair and Correction
       5.3.1 Self-repair
       5.3.2 Other-repair
       5.3.3 Referent correction
       5.3.4 Parameter correction
       5.3.5 Intent correction
       5.3.6 Scope correction

   5.4 Incremental Refinement
       5.4.1 Adding constraints
       5.4.2 Removing constraints
       5.4.3 Narrowing the search space
       5.4.4 Broadening the search space
       5.4.5 Reprioritizing criteria

   5.5 Mixed Initiative
       5.5.1 User-led initiative
       5.5.2 System-led clarification
       5.5.3 Collaborative narrowing
       5.5.4 Reframing and negotiation

   5.6 Redundancy and Emphasis
       5.6.1 Repetition as priority signal
       5.6.2 Repetition as urgency signal
       5.6.3 Repetition as contrast
       5.6.4 Repetition as repair

   5.7 Non-Monotonicity
       5.7.1 Later turns invalidating earlier turns
       5.7.2 Retraction of extracted parameters
       5.7.3 Soft cancellation
       5.7.4 Partial revision
       5.7.5 Conflicting updates

   5.8 Conversational Memory Limits
       5.8.1 Stale references
       5.8.2 Lost salience
       5.8.3 Conflicting constraints
       5.8.4 Overloaded discourse context

   5.9 Summary: How Meaning Evolves Across Turns

6. Social, Epistemic, and Priority Signals
   6.1 Politeness and Face Management
       6.1.1 Softened requests
       6.1.2 Indirect criticism
       6.1.3 Deference
       6.1.4 Mitigation
       6.1.5 Face-saving formulations

   6.2 Epistemic Stance
       6.2.1 Certainty
       6.2.2 Uncertainty
       6.2.3 Inference
       6.2.4 Reported knowledge
       6.2.5 Evidence strength

   6.3 Affective Stance
       6.3.1 Frustration
       6.3.2 Concern
       6.3.3 Skepticism
       6.3.4 Satisfaction
       6.3.5 Reluctance

   6.4 Modality
       6.4.1 Possibility
       6.4.2 Necessity
       6.4.3 Permission
       6.4.4 Obligation
       6.4.5 Prohibition

   6.5 Authority and Role
       6.5.1 Speaker authority
       6.5.2 Addressee authority
       6.5.3 Institutional role
       6.5.4 Ownership
       6.5.5 Approval rights

   6.6 Urgency and Priority
       6.6.1 Explicit deadlines
       6.6.2 Relative deadlines
       6.6.3 Consequence framing
       6.6.4 Repeated emphasis
       6.6.5 Priority tradeoffs

   6.7 Commitment Strength
       6.7.1 Strong assertion
       6.7.2 Tentative assertion
       6.7.3 Preference
       6.7.4 Recommendation
       6.7.5 Hard requirement

   6.8 Summary: How Stance Changes Interpretation

7. Failure, Risk, and Repair Semantics
   7.1 Interpretive Failure Types
       7.1.1 Wrong referent
       7.1.2 Wrong intent
       7.1.3 Missing constraint
       7.1.4 Stale context
       7.1.5 Hallucinated context
       7.1.6 Over-execution
       7.1.7 Under-execution

   7.2 Ambiguity Severity
       7.2.1 Harmless ambiguity
       7.2.2 Recoverable ambiguity
       7.2.3 Execution-blocking ambiguity
       7.2.4 Safety-critical ambiguity

   7.3 Confidence and Commitment
       7.3.1 Infer silently
       7.3.2 Infer with disclosure
       7.3.3 Ask for clarification
       7.3.4 Offer alternatives
       7.3.5 Require confirmation
       7.3.6 Decline execution

   7.4 Repair Protocols
       7.4.1 Clarification question
       7.4.2 Confirmation
       7.4.3 Correction acceptance
       7.4.4 Rollback
       7.4.5 Partial execution
       7.4.6 Safe no-op

   7.5 Summary: When Uncertainty Requires Repair Rather Than Execution

8. Cross-Property Interaction
   8.1 How Properties Co-Occur
   8.2 Ambiguity plus Context Dependence
   8.3 Vagueness plus Urgency
   8.4 Deixis plus Interface Grounding
   8.5 Politeness plus Directive Force
   8.6 Repair plus Non-Monotonicity
   8.7 Modality plus Authority
   8.8 Examples of Multi-Property Utterances

9. Diagnostic Property Index
   9.1 Property Table
   9.2 Core Question Per Property
   9.3 Example Per Property
   9.4 Required Resolution Type
   9.5 Risk Level
   9.6 Downstream Obligation

10. Worked Examples
   10.1 “Move it to tomorrow.”
   10.2 “Make this less aggressive.”
   10.3 “Send that to Sarah before the meeting.”
   10.4 “Actually, not that one — the cheaper option from yesterday.”
   10.5 “We probably should not change the public API before launch.”
   10.6 “Onboard the new contractor.”
   10.7 “Same as before, but only the urgent ones.”

11. Glossary
   11.1 Ambiguity
   11.2 Vagueness
   11.3 Underspecification
   11.4 Deixis
   11.5 Coreference
   11.6 Ellipsis
   11.7 Presupposition
   11.8 Pragmatics
   11.9 Speech act
   11.10 Implicature
   11.11 Discourse repair
   11.12 Mixed initiative
   11.13 Epistemic stance
   11.14 Modality
   11.15 Commitment strength
   11.16 Interface grounding
   11.17 Non-monotonicity

12. Final Summary
   12.1 Layer 0 as an irreducible substrate
   12.2 Meaning is distributed across multiple dimensions
   12.3 Context and pragmatics should remain distinct but adjacent
   12.4 Natural-language interpretation is dynamic, situated, and repairable

0. Front Matter

Layer 0 defines the natural-language substrate that exists before parsing, planning, tool use, validation, execution, or system policy.

Its purpose is not to describe how a system should implement natural-language understanding. Its purpose is to define what natural-language input is like before implementation begins: how meaning is formed, how it depends on context, how intent is inferred, how discourse mutates over time, and how social, epistemic, and priority signals change interpretation.

Layer 0 is therefore the descriptive foundation for any architecture that accepts natural language at its boundary.

0.1 Purpose#

This document defines the foundational properties of natural language as an interface substrate.

Natural language is not simply a less precise way to express formal commands. It is a situated communicative system. Speakers routinely use ambiguity, vagueness, underspecification, deixis, ellipsis, presupposition, indirectness, implicature, repair, and social signaling. These are not peripheral defects. They are normal operating features of natural language.

It answers questions such as:

The document is intended to prevent a recurring architectural mistake: treating natural-language failures as isolated model errors, prompt-design issues, or missing parameters when they are often consequences of deeper linguistic structure.

For example:

"Move it to tomorrow."

This utterance is compact, natural, and ordinary. But it depends on several unresolved properties:

Layer 0 does not treat this as malformed input. It treats it as a normal case of natural-language communication.

The purpose of the document is therefore:

To define the irreducible properties of natural-language input that downstream systems must preserve, resolve, defer, or safely reject.

0.2 Scope#

This document describes natural language at the interface boundary.

It covers the linguistic and semantic properties that shape interpretation before downstream execution occurs. It is concerned with what user language means, how that meaning is distributed, and why interpretation is often non-trivial.

Layer 0 includes:

Layer 0 excludes implementation architecture.

It does not define:

Layer 0 therefore remains descriptive rather than prescriptive.

It defines the input substrate. It does not define the machinery that processes it.

This separation is deliberate. Without it, the document would conflate three different questions:

Layer 0 is restricted to the first question.

0.3 Core Thesis#

Natural language is a situated, context-sensitive, pragmatically inferred, socially mediated communicative system.

Its meaning is not located only in words or syntax. Meaning is distributed across multiple dimensions:

Natural-language meaning
├── surface form
├── conventional semantic content
├── implicit context
├── reference and grounding
├── pragmatic intent
├── discourse history
├── social and epistemic stance
├── priority and urgency
└── repair and revision over time

The core thesis of this document is:

Natural language is not an imprecise command language.
It is a distinct interface substrate whose normal operation depends on indeterminacy, context, inference, discourse mutation, and social stance.

This thesis has several consequences.

First, natural-language interpretation cannot be reduced to sentence decoding.

A sentence may be grammatically complete and still be uninterpretable without context:

"Send it to her tomorrow."

The sentence contains an action, object, recipient, and time expression. Yet each depends on context:

Second, natural-language interpretation cannot be reduced to literal meaning.

A sentence may literally ask one thing while pragmatically doing another:

"Can you send this to Anna?"

Literal meaning:

Are you able to send this to Anna?

Typical intended meaning:

Please send this to Anna.

The ability question functions as a request. The literal form remains relevant, but it does not exhaust the communicative act.

Third, natural-language interpretation cannot be reduced to filling missing parameters.

Underspecification is not merely absence. It is often efficient compression.

A user normally does not say:

Please create a 30-minute calendar event with Alex Chen from the design team,
using the first mutually available slot next week during working hours,
with the title "Roadmap review",
using video conferencing,
and include the current project document as context.

The user says:

"Schedule a roadmap review with Alex next week."

This is not defective language. It is economical language. It assumes context, shared conventions, defaults, and follow-up repair.

Fourth, natural-language interpretation is non-monotonic.

Later utterances can revise earlier ones:

"Make it shorter."
"Actually, keep the detail but shorten the introduction."

The second turn does not simply add information. It changes the interpretation of the first turn. “Shorter” is no longer a general compression instruction. It becomes a localized revision targeting the introduction while preserving detail.

Fifth, natural-language meaning includes stance.

Compare:

The underlying topic is the same. The stance changes the force.

Sixth, natural-language failure is not binary.

An utterance can be:

The appropriate response depends on confidence, consequence, reversibility, and available repair.

Why this structure follows from the thesis#

The document structure is chosen to follow the path by which natural-language meaning becomes actionable interpretation.

It begins with meaning itself, then moves through the forces that make meaning partial, situated, intentional, dynamic, stance-bearing, and risk-sensitive.

The structure is not organized as a traditional linguistics survey. It is organized around the interpretive obligations created by natural-language input.

The sequence is:

1. Meaning as a Layered Phenomenon
2. Indeterminacy and Compression
3. Context, Reference, and Grounding
4. Communicative Force and Intent
5. Discourse State and Conversational Mutation
6. Social, Epistemic, and Priority Signals
7. Failure, Risk, and Repair Semantics

Each section answers a different interpretive question.

This ordering is intentional.

Sections 1-3 establish the semantic substrate:

What does the utterance say?
Why is it incomplete or non-exact?
What context is needed to interpret it?

Sections 4-6 establish the interactional substrate:

What is the speaker trying to do?
How does the conversation change over time?
How does stance alter force, priority, and commitment?

Section 7 establishes the repair substrate:

What can go wrong?
How severe is the uncertainty?
When should interpretation proceed, pause, ask, confirm, roll back, or safely no-op?

This structure reflects the central claim of the document:

Natural-language understanding is not a single operation.
It is the coordinated resolution of layered meaning, indeterminacy, context, intent, discourse state, stance, and risk.

0.4 Audience#

This document is written for people who design, evaluate, or reason about systems that accept natural-language input.

Primary audiences include:

The document assumes the reader is comfortable with technical systems, but it does not require formal training in linguistics.

It introduces linguistic concepts only where they clarify interface behavior.

For example, the document does not discuss deixis as an abstract linguistic category for its own sake. It discusses deixis because expressions such as:

this
that
here
there
now
tomorrow
the other one
same as before

are common in natural-language interfaces and cannot be interpreted without situational grounding.

Similarly, the document does not discuss speech acts for theoretical completeness. It discusses speech acts because systems must distinguish between:

"Can you send this?"
"Did you send this?"
"Do not send this."
"Maybe send this later."
"I’m not comfortable sending this."

These utterances concern similar objects and actions, but they carry different communicative force.

The document is also useful for teams diagnosing natural-language interface failures.

A failure such as:

The system sent the wrong file to the wrong Alex.

may involve several Layer 0 phenomena:

The audience should use Layer 0 as a diagnostic vocabulary before assigning responsibility to a model, prompt, memory component, UI state manager, policy layer, or execution system.

0.5 How to Read This Document#

This document should be read as a layered specification of natural-language input properties.

It is not a linear pipeline in which every utterance passes cleanly through each section. Real utterances often involve many properties at once.

Example:

"Could you maybe send that to Sarah before the meeting?"

This single utterance contains:

No single section fully explains the utterance. The sections are analytical lenses.

The recommended reading order is the document order:

1 → 2 → 3 → 4 → 5 → 6 → 7

This order moves from lower-level meaning to higher-level interactional and repair behavior.

However, the document can also be read diagnostically.

Each section follows the same basic pattern:

Concept
→ Explanation
→ Examples
→ Subtypes
→ Interpretive consequences
→ Summary

The summaries are intended to serve as quick operational references. The examples are intended to show how the phenomenon appears in ordinary user language.

Why the document uses this structure instead of the original A-E grouping#

The original taxonomy grouped the properties as:

A. Indeterminacy of meaning
B. Context dependence
C. Pragmatic meaning and speaker intent
D. Discourse and conversational structure
E. Social and communicative framing

Those groups remain conceptually valid. They are preserved in substance.

The revised document reorganizes them into a more interface-oriented sequence:

Meaning
→ Indeterminacy
→ Context
→ Intent
→ Discourse mutation
→ Stance
→ Failure and repair

This change has three reasons.

First, the revised structure separates introductory linguistic layering from the main property taxonomy.

Syntax, semantics, pragmatics, discourse, and social meaning are foundational layers. They explain that meaning is distributed. But they should not be mixed with every downstream property as if they were mutually exclusive categories.

Second, the revised structure makes the interpretation flow clearer.

A system or analyst must often reason in this order:

What was said?
What is ambiguous, vague, or missing?
What context grounds it?
What act is being performed?
How does this update the conversation?
How does stance affect force?
Is it safe to infer, or is repair needed?

The revised outline follows that reasoning path.

Third, the revised structure gives failure and repair their own section.

The original taxonomy correctly identified repair and conversational structure, but risk-sensitive interpretation deserves explicit treatment. Natural-language uncertainty is not always benign. Some ambiguity is harmless. Some is recoverable. Some blocks execution. Some is safety-critical.

Separating Failure, Risk, and Repair Semantics makes the document more useful for real interface analysis without turning it into an implementation manual.

0.6 Layer 0 Invariants#

Layer 0 is governed by a small set of invariants.

These invariants are the assumptions the rest of the document relies on.

Invariant 1: Natural-language properties are irreducible#

Layer 0 phenomena cannot be engineered away by asking users to be fully explicit, context-free, literal, unambiguous, and parameter-complete.

Users will continue to say things like:

"Move it to tomorrow."
"Make this less aggressive."
"Same as before, but cheaper."
"Actually, not that one."
"Can you send this to Anna?"
"Maybe avoid changing the public API."

These utterances are not malformed. They are normal.

A system may constrain what it accepts, but natural language itself remains indeterminate, contextual, pragmatic, discourse-sensitive, and socially mediated.

Invariant 2: Meaning is distributed, not localized#

Meaning is not contained in a single token, sentence, or syntax tree.

It may be distributed across:

This invariant justifies the document’s layered structure.

Invariant 3: Indeterminacy is functional, not merely defective#

Ambiguity, vagueness, and underspecification are not only sources of error. They also make communication efficient.

For example:

"Find a reasonable hotel nearby."

This utterance omits exact price, distance, quality threshold, dates, room type, and amenities. Yet it is useful because the speaker expects context, defaults, preferences, and follow-up interaction to do part of the work.

Natural-language interfaces must therefore distinguish between:

The goal is not to eliminate indeterminacy. The goal is to handle it according to context and risk.

Invariant 4: Context supplies grounding; pragmatics derives intent#

Context and pragmatics are distinct but adjacent.

Context answers:

What background state is relevant?
What does this refer to?
What prior discourse matters?
What is visible, selected, assumed, or presupposed?

Pragmatics answers:

What is the speaker trying to do?
Is this a request, warning, correction, refusal, suggestion, or evaluation?
What is implied but unstated?
Which interpretation is relevant?

Example:

"This is a little too aggressive."

Context resolves:

Pragmatics infers:

The document keeps context and pragmatics adjacent because interpretation normally requires both.

Invariant 5: Natural-language interpretation is discourse-sensitive#

Utterances are not isolated events. They occur inside a changing discourse state.

Example:

"Find three options."
"The second one looks best."
"Make it cheaper."
"Actually, prioritize reliability."

Each turn depends on the previous ones.

The active meaning changes over time:

This invariant justifies the section on discourse mutation, repair, incremental refinement, and non-monotonicity.

Invariant 6: Later turns can revise earlier meanings#

Natural-language discourse is non-monotonic.

A later utterance can cancel, weaken, or replace an earlier one:

"Send it today."
"Actually, wait for legal approval."

The second utterance changes the status of the first. It does not merely add a new independent instruction.

Possible update types include:

This invariant is essential because natural-language interaction often constructs intent over multiple turns rather than stating it completely at the beginning.

Invariant 7: Social stance is part of meaning#

Politeness, certainty, urgency, affect, role, and commitment strength are not decorative additions to semantic content. They change interpretation.

Compare:

"Send this today."
"Could you send this today?"
"Can you maybe send this today?"
"I need this sent today."
"Do not send this today."
"I’m not comfortable sending this today."

All concern sending and timing. But they differ in force, authority, politeness, urgency, and prohibition.

Social and epistemic signals affect whether an utterance should be interpreted as:

This invariant justifies treating stance as structurally central rather than peripheral.

Invariant 8: Interpretation must be risk-sensitive#

The same degree of ambiguity can require different responses depending on consequence and reversibility.

Example:

"Make this cleaner."

In a draft-editing context, this may be safe to infer.

Example:

"Delete the old records."

In a data-management context, this may require clarification or confirmation.

Layer 0 therefore distinguishes between linguistic ambiguity and operational risk.

This invariant justifies the final section on failure, risk, and repair semantics.

Invariant 9: Repair is part of normal interpretation#

Repair is not an exception path. It is a basic feature of natural-language interaction.

Users naturally correct, revise, clarify, and change their minds:

"No, the other one."
"Actually, make it shorter."
"Sorry, I meant next Friday."
"Let me rephrase."
"Keep the structure, but change the tone."

Repair allows meaning to be negotiated over time.

Layer 0 treats repair as a first-class property because natural-language interpretation often becomes reliable only through interaction.

Invariant 10: Layer 0 defines obligations, not implementations#

Layer 0 identifies the interpretive obligations created by natural-language input.

It does not specify how those obligations must be implemented.

For example:

This invariant preserves the document’s role as a substrate specification.

The document defines the what:

What properties does natural-language input have?
What interpretive obligations do those properties create?

It does not define the how:

How should a particular system represent, validate, execute, or enforce those obligations?

0.7 Front-Matter Summary#

Layer 0 is the natural-language interface substrate.

Its central claim is that natural language is not an imprecise formal command language. It is a situated communicative system whose meaning is distributed across words, syntax, context, intent, discourse, stance, and repair.

The structure of this document follows from that claim.

Meaning is layered.
Meaning is often compressed or indeterminate.
Meaning requires context and grounding.
Meaning expresses communicative force.
Meaning changes across discourse.
Meaning carries social, epistemic, and priority signals.
Meaning can fail and require repair.

The rest of the document expands these claims into a taxonomy of natural-language properties and their interpretive consequences.

1. Meaning as a Layered Phenomenon

Natural-language meaning is not produced by words alone. It is distributed across multiple interacting layers: surface form, lexical meaning, constructional meaning, compositional semantics, pragmatic meaning, discourse state, and social or interactional signals.

These layers are analytically separable, but they are not independent. A single utterance can carry grammatical structure, conventional word meaning, contextual reference, implied intent, discourse function, and social stance at the same time.

Consider:

“Could you make this a little less aggressive?”

This utterance contains several layers of meaning:

Layer 0 treats this layered structure as foundational. A natural-language interface must not assume that the full meaning of an utterance is contained in its literal sentence meaning. Meaning often depends on what the speaker is doing, what has already happened, what is currently salient, what is socially implied, and what can reasonably be inferred.

1.1 Surface Form#

Surface form is the directly observable shape of an utterance: its words, order, punctuation, morphology, and grammatical pattern.

Examples:

“Move the meeting to tomorrow.”

“Can you move the meeting to tomorrow?”

“Maybe move the meeting to tomorrow?”

“The meeting should probably be moved to tomorrow.”

These utterances are closely related, but their surface forms differ. The first is an imperative. The second is an interrogative. The third is a softened suggestion. The fourth is a passive construction with modal and epistemic marking.

Surface form provides important clues, but it does not determine meaning by itself. Interrogative form does not always mean the speaker is asking a literal question. Imperative form does not always mean the speaker has authority. Declarative form does not always mean the speaker is merely stating a fact.

Examples:

Surface form is therefore an entry point into interpretation, not the whole interpretation.

1.2 Lexical Meaning#

Lexical meaning is the conventional meaning contributed by individual words and fixed expressions.

For example:

“Cancel my next meeting.”

The words contribute conventional meanings:

Lexical meaning is necessary but rarely sufficient. The word “meeting” may refer to a calendar event, a recurring sync, a video call, an in-person appointment, or an informal discussion. The word “cancel” may mean delete the event, decline attendance, notify participants, or mark it as no longer happening. The word “next” requires a reference time.

Lexical meaning also includes polysemy: a word may have several related meanings.

Examples:

Natural-language interpretation must therefore distinguish between the conventional possibilities available for a word and the meaning most relevant in context.

1.3 Constructional Meaning#

Constructional meaning is meaning contributed by grammatical patterns or conventional phrasings, not just by individual words.

For example:

“Can you X?”

This construction often functions as a request:

“Can you send the report?”

The literal semantic content concerns ability:

Are you able to send the report?

But the construction commonly carries a request function:

Please send the report.

Other constructions similarly carry conventional pragmatic force.

Constructional meaning matters because natural language does not always build meaning from isolated word meanings upward. Some sentence patterns are conventionalized ways of performing actions, expressing stance, softening criticism, or marking uncertainty.

A natural-language interface that treats every interrogative as a literal information request will misinterpret common requests. A system that treats every declarative as a passive statement will miss warnings, objections, and implied tasks.

1.4 Compositional Semantics#

Compositional semantics concerns how word meanings combine into phrase and sentence meanings.

Example:

“Send the revised contract to Anna before noon.”

The sentence combines several elements:

Compositional semantics gives the utterance its conventional propositional structure. It identifies events, participants, properties, relations, and constraints.

However, even a compositionally clear sentence may remain under-resolved.

In:

“Send the revised contract to Anna before noon.”

open questions may include:

Compositional semantics provides a structured meaning. It does not necessarily provide an executable interpretation.

1.5 Pragmatic Meaning#

Pragmatic meaning is meaning inferred from use in context. It concerns what the speaker is trying to do by saying something.

Example:

“It’s getting late.”

The literal semantic content is a statement about time. Depending on context, the pragmatic meaning may be:

Pragmatic meaning includes requests, warnings, suggestions, confirmations, refusals, objections, corrections, and implied actions.

Examples:

Pragmatic interpretation depends on context, relevance, shared assumptions, and the current task. The same sentence can perform different communicative acts in different situations.

1.6 Discourse Meaning#

Discourse meaning arises across multiple sentences, turns, paragraphs, or interactional moves. It tracks how utterances relate to prior and future utterances.

Example:

“Find three options. The second one looks best. Make it cheaper.”

The third sentence cannot be interpreted in isolation. “It” refers to “the second one,” and “make it cheaper” modifies an ongoing selection or proposal. The instruction depends on prior discourse state.

Discourse meaning includes:

Discourse meaning is dynamic. Later turns can extend, narrow, revise, or invalidate earlier interpretations.

Example:

“Schedule it for Friday. Actually, Monday is better.”

The second sentence does not simply add a new date. It revises the prior date. A natural-language interface must therefore model conversational state as mutable rather than purely additive.

1.7 Social and Interactional Meaning#

Natural language carries social and interactional signals in addition to informational content. These signals include politeness, certainty, authority, urgency, hesitation, frustration, preference strength, and interpersonal stance.

Example:

“Maybe we should avoid changing the public API before launch.”

This utterance does not merely describe an API. It carries several signals:

Social and interactional meaning affects interpretation. A softened request is still often a request. A tentative warning may still mark serious risk. A frustrated statement may require more careful repair than a neutral one.

Examples:

These signals are not decorative. They are part of meaning. They influence how strongly an utterance should constrain interpretation, whether a response should be direct or cautious, and whether clarification or confirmation is appropriate.

1.8 Summary#

Meaning in natural language is layered.

A compact model:

Natural-language meaning
├── surface form
├── lexical meaning
├── constructional meaning
├── compositional semantics
├── pragmatic meaning
├── discourse meaning
└── social and interactional meaning

Each layer contributes different information:

Layer 0 begins with this premise: a natural-language utterance is not merely a string to decode. It is a situated communicative act whose meaning is distributed across form, convention, context, intent, discourse, and stance.

2. Indeterminacy and Compression

Natural language often does not map cleanly to one exact interpretation. Expressions may be ambiguous, vague, underspecified, prototype-based, metaphorical, or intentionally compressed.

This is not accidental. Natural language is optimized for situated communication, not exhaustive specification. Speakers routinely omit information that can be inferred from context, defaults, shared routines, prior discourse, or follow-up interaction. In many cases, saying less is not a failure of precision; it is an efficient communicative strategy.

Indeterminacy and compression include:

Indeterminacy and compression
├── ambiguity
├── vagueness
├── underspecification
├── prototype-based categories
├── metaphor and conceptual transfer
└── economy of explicitness

These properties explain why many natural-language utterances are only partially specified at the moment they are produced. Interpretation often requires resolving, ranking, deferring, or explicitly repairing possible meanings.

2.1 Ambiguity#

Ambiguity occurs when an expression has more than one discrete interpretation.

Example:

“Book a table.”

Possible interpretations include:

Ambiguity differs from vagueness. Ambiguity asks:

Which meaning is intended?

Vagueness asks:

Where is the boundary?

Ambiguity can occur at several levels.

2.1.1 Lexical Ambiguity#

Lexical ambiguity occurs when a word or expression has multiple possible meanings.

Examples:

Lexical ambiguity is usually resolved through context.

Example:

“Run the migration.”

In a software context, “run” likely means execute. In an organizational context, it could mean manage. In a physical event context, it could mean operate or coordinate.

2.1.2 Syntactic Ambiguity#

Syntactic ambiguity occurs when a sentence allows multiple grammatical parses.

Example:

“I saw the man with the telescope.”

Possible interpretations:

1. I used a telescope to see the man.
2. I saw a man who had a telescope.

Another example:

“Review the documents from Anna with comments.”

Possible interpretations:

1. Review the documents from Anna, and include comments.
2. Review the documents from Anna that already contain comments.
3. Review the documents using Anna’s comments.

Syntactic ambiguity can affect task interpretation because the grammatical attachment of a modifier changes what the instruction applies to.

2.1.3 Scope Ambiguity#

Scope ambiguity occurs when an operator, quantifier, negation, or modifier may apply to different parts of an utterance.

Example:

“Email all managers and engineers in Berlin.”

Possible interpretations:

1. Email all managers, plus engineers who are in Berlin.
2. Email all managers and all engineers who are in Berlin.

Another example:

“Do not archive urgent and unresolved tickets.”

Possible interpretations:

1. Do not archive tickets that are both urgent and unresolved.
2. Do not archive urgent tickets or unresolved tickets.
3. Do not archive urgent tickets, and do not archive unresolved tickets.

Scope ambiguity is especially important when an utterance contains quantifiers, negation, conjunction, conditionals, or modifiers.

Common scope-sensitive expressions include:

2.1.4 Referential Ambiguity#

Referential ambiguity occurs when a referring expression could point to more than one entity.

Example:

“Send it to Alex.”

Open questions:

Another example:

“Use the second one.”

Possible referents for “the second one” may include:

• the second search result
• the second paragraph
• the second design
• the second person mentioned
• the second file in a list
• the second option from a prior turn

Referential ambiguity depends heavily on salience. The most recently mentioned or currently visible entity is often preferred, but this is not guaranteed.

2.2 Vagueness#

Vagueness occurs when a term has fuzzy boundaries rather than a fixed exact threshold.

Examples:

Vagueness is not the same as ambiguity.

Vague terms are often gradable. Something can be:

• somewhat cheap
• very cheap
• cheap for Tokyo
• cheap compared with competitors
• cheap for enterprise software
• cheap for a weekend trip
• cheap relative to the user’s usual choices

Vagueness is often context-relative. The same term can imply different thresholds in different domains.

Example:

“Find a fast database.”

“Fast” could mean:

Vague terms often require a threshold. Sometimes the threshold can be inferred; sometimes it must be clarified.

2.3 Underspecification#

Underspecification occurs when an utterance leaves out information that may be needed for a fully resolved interpretation.

Example:

“Schedule a meeting with Alex.”

Missing information may include:

Underspecification is normal. Speakers do not usually provide every parameter in a single utterance because many details can be inferred from prior interaction, task conventions, user preferences, or follow-up dialogue.

Other examples:

Underspecification differs from ambiguity. Ambiguity gives multiple possible meanings. Underspecification gives an incomplete meaning.

Example:

“Schedule a meeting.”

This is not necessarily ambiguous. The action type is reasonably clear. But it is underspecified because required details are missing.

2.3.1 Missing Parameters#

Some underspecification involves missing parameters.

Example:

“Book a flight to Berlin.”

Likely missing parameters:

• origin
• date
• return date
• budget
• airline preferences
• baggage requirements
• passenger identity
• seat preference
• acceptable layovers
• refundability requirements

Parameter omission can be harmless, recoverable, or blocking.

2.3.2 Recoverable Omissions#

An omission is recoverable when the missing material can be inferred with acceptable confidence from context.

Example:

“Same time tomorrow?”

If the current conversation concerns a recurring meeting, “same time” may be recoverable from the prior scheduled event. If there are several recent times under discussion, it may not be recoverable.

Recoverability depends on:

2.3.3 Defaults and Assumptions#

Natural-language interpretation often uses defaults.

Example:

“Set up a call with Maya.”

Possible defaults:

Defaults are useful, but they are not neutral. They encode assumptions. A default can be wrong even when it is plausible.

A natural-language interface should distinguish between:

explicit instruction
inferred value
default value
unresolved value

These should not be treated as equivalent. Explicit instructions carry stronger interpretive force than inferred defaults.

2.3.4 User-Side Economy of Expression#

Underspecification is often an efficiency strategy. Speakers compress meaning because explicit specification has a cost.

Consider the difference between:

“Schedule a meeting with Alex.”

and:

“Create a 30-minute video meeting with Alex Chen from the product team next Tuesday at 10:00 AM Buenos Aires time, title it ‘Roadmap sync,’ invite only Alex and me, use my work calendar, include the usual video link, and avoid conflicts unless the conflict is marked optional.”

The second version is more explicit, but it is also slower, heavier, and cognitively expensive. Natural language typically avoids this burden unless the details matter.

This is the economy of explicitness:

Speakers usually specify only what they believe is necessary for the listener to infer the intended meaning.

This principle explains why underspecification is not simply a defect. It is part of how natural language reduces interaction cost.

A natural-language interface must therefore treat omitted information carefully. Missing details may indicate:

• the user expects the system to infer them
• the user expects a follow-up question
• the user does not care about the omitted detail
• the detail is conventionally defaulted
• the detail is unavailable to the user
• the user has not yet decided
• the user assumes shared context

The same omission can require different handling depending on risk, reversibility, and confidence.

2.4 Prototype-Based Categories#

Many natural-language categories are not defined by strict necessary-and-sufficient conditions. Instead, they are organized around prototypes: central, typical examples of a category.

Example:

“Find startup-like companies.”

There may be no exact definition of “startup-like.” The phrase suggests a cluster of features:

• relatively young company
• growth-oriented
• technology-oriented
• venture-backed or venture-style
• experimental product culture
• high uncertainty
• high upside
• relatively informal operating style
• small or rapidly scaling team

A company may be more or less startup-like. The category is graded rather than binary.

Another example:

“Make the design feel more enterprise.”

This may imply:

Prototype-based categories are common in product, design, hiring, strategy, and technical evaluation.

Examples:

Prototype categories are powerful because they allow speakers to communicate complex clusters of expectations compactly. But they also create interpretive uncertainty because different speakers may have different prototypes.

2.5 Metaphor and Conceptual Transfer#

Metaphor uses one conceptual domain to structure another.

Examples:

Metaphor is not merely decorative. It is a normal way of reasoning. Many technical and organizational terms are metaphorical:

• threads
• locks
• queues
• pipelines
• memory
• garbage collection
• branches
• forks
• roots
• leaves
• deadlocks
• bottlenecks
• ownership
• handoff
• rollback
• health checks

Metaphors often compress complex judgments. Saying:

“This interface feels heavy.”

may imply several possible issues:

• too many visible controls
• excessive visual weight
• slow perceived responsiveness
• too much text
• too much cognitive load
• unclear hierarchy
• too many required decisions

Because metaphor transfers structure from one domain to another, interpretation requires identifying which aspects of the source domain are relevant.

Example:

“This proposal needs a stronger spine.”

Possible intended meanings:

• clearer central argument
• stronger logical structure
• more explicit thesis
• better sequencing
• more support for conclusions

Metaphor therefore often requires pragmatic and domain-sensitive interpretation.

2.6 Economy of Explicitness#

The economy of explicitness is the principle that natural-language speakers usually avoid specifying everything they could specify.

This economy is not laziness. It is part of efficient communication.

Natural-language utterances are shaped by tradeoffs:

Example:

“Same as last time, but cheaper.”

This short utterance compresses a large amount of information:

The speaker avoids restating everything because prior context is assumed to be available.

This creates a central Layer 0 constraint:

Natural-language input is often a compressed representation of intent.

Interpreting it requires decompression:

• recover omitted entities
• infer relevant defaults
• bind references
• identify implied constraints
• determine which prior state still applies
• distinguish hard requirements from preferences
• identify when clarification is necessary

Compression is useful, but it creates risk. A listener or system can decompress incorrectly.

Example:

“Use the usual format.”

Possible interpretations:

• the format used in the last document
• the team’s standard template
• the user’s personal preference
• the organization’s style guide
• the format used for this document type
• the format previously approved by a stakeholder

The more compressed the utterance, the more interpretation depends on context and shared assumptions.

2.7 Summary#

Indeterminacy and compression explain why natural-language meaning is often partial, graded, or multiply interpretable.

A compact model:

Natural-language utterances often compress intent.

Compression appears as:
├── multiple possible meanings
├── fuzzy thresholds
├── omitted parameters
├── graded categories
├── metaphorical transfer
└── reliance on shared context

The interpretive task is not always to eliminate indeterminacy immediately. Sometimes the correct response is to infer. Sometimes it is to ask. Sometimes it is to preserve alternatives. Sometimes it is to delay resolution until more context is available. Sometimes it is to decline action until the ambiguity is resolved.

3. Context, Reference, and Grounding

Natural-language meaning often depends on context. An utterance may not be interpretable from the sentence alone because it relies on the speaker, addressee, time, location, prior discourse, shared knowledge, visible objects, task state, or interface state.

Example:

“Move this to tomorrow.”

This utterance requires several contextual resolutions:

Context dependence includes:

Context, reference, and grounding
├── implicit context
├── deixis
├── coreference
├── ellipsis
├── presupposition
├── temporal and spatial grounding
├── discourse grounding
└── interface grounding

This section focuses on how expressions attach to the world, the conversation, the task, and the active interface.

3.1 Implicit Context#

Implicit context is unspoken background information that speakers assume is available.

Example:

“Do it like last time.”

This short utterance depends on several unstated facts:

Implicit context may come from:

• previous conversation
• shared work history
• current task state
• visible interface state
• active document or selected object
• user preferences
• organizational conventions
• domain norms
• physical environment
• cultural assumptions
• prior decisions
• recurring routines

Implicit context is not an edge case. It is a normal operating condition of natural language.

Examples:

Implicit context is closely related to pragmatics, but the two are distinct.

Implicit context supplies background information.
Pragmatics uses that background to infer intended meaning.

For example:

“It’s cold in here.”

Context may identify the location, temperature, open window, thermostat, car, room, or user preference. Pragmatics determines whether the speaker is merely observing, complaining, asking for the heat to be increased, or requesting that a window be closed.

3.2 Deixis#

Deixis refers to expressions whose meaning depends on the situation of utterance. Deictic expressions point to people, times, places, discourse segments, social roles, or visible objects.

Common deictic categories:

Example:

“I need this by tomorrow.”

Interpretation requires:

Deixis makes natural language compact. Speakers can say “this” instead of restating a full object description, and “tomorrow” instead of giving an absolute date.

But deixis also creates risk when the anchor is unclear.

Example:

“Move that over there.”

This utterance may be perfectly clear in a shared visual workspace. It may be impossible to interpret in a text-only context.

3.2.1 Person Deixis#

Person deixis identifies participants relative to the speech situation.

Examples:

The pronoun “we” is especially context-sensitive.

Example:

“We should delay the launch.”

Possible meanings of “we”:

• the speaker and addressee
• the product team
• the engineering organization
• the company
• a leadership group
• everyone involved in launch planning

The scope of “we” can affect authority, responsibility, and actionability.

3.2.2 Time Deixis#

Time deixis anchors expressions to a reference time.

Examples:

Example:

“Send this tomorrow morning.”

Open questions:

• What date is tomorrow?
• Which timezone applies?
• What counts as morning?
• Is the instruction tied to the user’s location, recipient’s location, or organization’s working hours?

Time deixis is especially sensitive when participants, systems, or events span timezones.

3.2.3 Place Deixis#

Place deixis anchors expressions to a spatial frame.

Examples:

Example:

“Find a hotel nearby.”

“Nearby” may depend on:

• the user’s current location
• a destination under discussion
• an event venue
• a map viewport
• a selected address
• a prior search result

Place deixis can be physical, digital, or discourse-based.

3.2.4 Discourse Deixis#

Discourse deixis refers to parts of the surrounding discourse.

Examples:

Example:

“This is too long. Cut it by half.”

“This” and “it” likely refer to a prior generated text, selected text, or visible artifact. The utterance cannot be interpreted without discourse or interface grounding.

3.2.5 Social Deixis#

Social deixis encodes social relationships, roles, status, or institutional position.

Examples:

Example:

“Send this to legal before sharing it with the client.”

This requires grounding both “legal” and “the client.” These are not merely nouns; they invoke organizational roles and process expectations.

3.3 Coreference#

Coreference occurs when multiple expressions refer to the same entity.

Example:

“Sarah sent the contract yesterday. She said it was ready. Forward it to legal.”

Coreference links:

Coreference allows speakers to avoid repeating full descriptions. It also creates dependency across clauses and turns.

Example with ambiguity:

“Anna sent Maria the report after she revised it.”

Open questions:

Coreference resolution depends on grammar, salience, recency, world knowledge, discourse structure, and plausibility.

3.3.1 Pronoun Resolution#

Pronoun resolution identifies what pronouns refer to.

Examples:

Pronouns are efficient but risky when multiple candidate referents are available.

3.3.2 Named-Entity Resolution#

Named-entity resolution identifies which real or discourse entity a name denotes.

Example:

“Send it to Alex.”

Possible issues:

• multiple contacts named Alex
• Alex as person vs. team alias
• Alex in different organizations
• Alex from recent conversation vs. Alex from user’s address book
• incomplete or outdated contact information

Named entities may look explicit while still being ambiguous.

Examples:

3.3.3 Repeated Descriptions#

Coreference is not limited to pronouns. Repeated or varied descriptions may refer to the same entity.

Example:

“The vendor sent over the agreement. The contract still needs legal review.”

“The agreement” and “the contract” may refer to the same document.

Another example:

“The migration plan is risky. This proposal needs another review.”

“The migration plan” and “this proposal” may refer to the same artifact, depending on context.

Natural language often uses different descriptions for the same thing to emphasize different properties.

3.3.4 Ambiguous Referents#

A referent is ambiguous when more than one candidate fits.

Example:

“Compare the two proposals and send the stronger one to Maya.”

Open questions:

• stronger by what criterion?
• which proposal is stronger?
• who is Maya?
• should the weaker one be retained, ignored, or summarized?

Referential ambiguity is especially common when an utterance includes:

• pronouns
• demonstratives
• names shared by multiple entities
• ordinal expressions
• comparative descriptions
• role labels
• recent lists or options
• visible interface objects

3.4 Ellipsis#

Ellipsis occurs when part of an expression is omitted because it can be recovered from context.

Example:

“Find flights to Berlin. Cheapest, not fastest.”

The second sentence omits much of the full structure:

“Find the cheapest flights to Berlin, not the fastest flights to Berlin.”

Ellipsis is common in multi-turn interaction.

Examples:

Ellipsis is efficient because speakers do not need to repeat shared or recently mentioned material. It is also highly context-dependent.

3.4.1 Sentence-Level Ellipsis#

Sentence-level ellipsis omits material recoverable from the same sentence or adjacent sentence.

Example:

“Maya wants the PDF, and Alex the spreadsheet.”

Recovered meaning:

“Maya wants the PDF, and Alex wants the spreadsheet.”

Another example:

“I’ll handle backend; you frontend.”

Recovered meaning:

“I’ll handle backend; you handle frontend.”

3.4.2 Turn-Level Ellipsis#

Turn-level ellipsis depends on prior conversational turns.

Example:

“Find three candidates.” “Only senior ones.” “With Rails experience.” “Not contractors.”

The later turns are fragments. They modify the original search criteria.

Recovered interpretation:

Find three candidates who are senior, have Rails experience, and are not contractors.

But even this recovered interpretation may need further refinement. “Senior” and “Rails experience” are vague. “Not contractors” may mean exclude independent contractors, staffing-agency candidates, temporary workers, or non-employees.

3.4.3 Task-State Ellipsis#

Task-state ellipsis depends on the current task rather than only prior text.

Example:

“Run it again.”

To interpret this, one must know the current task state:

• What was run?
• With what parameters?
• Should the same inputs be reused?
• Should failed steps be retried only, or the whole process?
• Should outputs be overwritten or versioned?

Another example:

“Same settings, different file.”

This requires identifying both prior settings and the new file.

Task-state ellipsis is common in workflows because users often operate over a current object or active process.

3.4.4 Recoverability Conditions#

Ellipsis is interpretable only when omitted material is recoverable.

Recoverability depends on:

When recoverability is weak, ellipsis becomes a source of interpretive failure.

Example:

“Do the same thing with the other one.”

This is easy if there are exactly two visible files and one prior operation. It is difficult if the conversation contains several prior operations and several candidate objects.

3.5 Presupposition#

A presupposition is background information that an utterance treats as already true.

Example:

“Cancel my next meeting with Anna.”

Presuppositions:

1. The speaker has meetings.
2. There is at least one upcoming meeting with Anna.
3. One of those meetings is the next relevant one.
4. The speaker has some authority or ability to cancel it.

Presuppositions differ from direct assertions. They are not the main point of the utterance, but the utterance depends on them.

Another example:

“Stop sending weekly reports to the client.”

Presuppositions:

• weekly reports are currently being sent
• there is a relevant client
• the reports are sent on a recurring basis
• the sending can be stopped
• the speaker has standing to request the change

Presuppositions can fail.

Example:

“Reschedule my call with Jordan.”

Possible presupposition failures:

• there is no call with Jordan
• there are multiple calls with Jordan
• Jordan cannot be identified
• the call is already canceled
• the speaker does not have permission to reschedule it

Presupposition failure often requires repair:

Presuppositions are important because they reveal assumed state. An utterance often asks for an action while also implying that certain background conditions already hold.

3.6 Temporal and Spatial Grounding#

Temporal and spatial grounding resolve expressions to time and place.

These are related to deixis but deserve separate attention because time and place often determine whether an utterance can be acted on.

3.6.1 Temporal Grounding#

Temporal grounding resolves expressions such as:

• now
• today
• tomorrow
• yesterday
• next week
• this Friday
• end of day
• before lunch
• after the meeting
• later
• soon
• in a bit
• by launch
• before the deadline

Example:

“Remind me before the meeting.”

Open questions:

Temporal expressions may be absolute, relative, vague, recurring, or event-based.

Temporal grounding must often account for timezone, locale, working hours, holidays, and event calendars.

3.6.2 Spatial Grounding#

Spatial grounding resolves expressions such as:

• here
• there
• nearby
• around me
• in this room
• on the left
• above
• below
• in the next section
• at the top
• in the sidebar
• on the map

Example:

“Move this above the chart.”

This may refer to a document layout, dashboard, slide, web page, or design canvas. The spatial frame must be known.

Spatial grounding can be:

Spatial expressions are often easy for humans in a shared visual environment and difficult without the relevant visual or interface state.

3.7 Discourse Grounding#

Discourse grounding binds expressions to prior or surrounding discourse.

Example:

“Use the second option, but make it cheaper.”

This requires identifying:

• the relevant option set
• which item is second
• what “it” refers to
• what “cheaper” means in relation to that option
• which properties should remain unchanged

Discourse grounding is required for expressions such as:

Discourse grounding must track salience. The most recent entity is not always the intended referent. A speaker may refer back to something earlier if it remains topically important.

Example:

“The migration plan is risky. The API change is manageable. The bigger issue is that it affects billing.”

“It” might refer to the migration plan, the API change, or a specific subcomponent depending on discourse focus and domain knowledge.

3.8 Interface Grounding#

Interface grounding occurs when natural-language meaning depends on the active software environment.

Examples:

“Delete this row.”

“Make this chart larger.”

“Move that section above the summary.”

“Use the selected text.”

“Change this column to currency.”

These utterances cannot be resolved from language alone. They require access to interface state.

Interface grounding may include:

Interface grounding extends deixis into digital environments. In a physical conversation, “this” may be resolved by pointing. In a software environment, “this” may be resolved by selection, cursor position, viewport, or UI focus.

Example:

“Make this shorter.”

Possible interface-grounded referents:

• selected paragraph
• current draft
• highlighted sentence
• previous assistant response
• active document
• text box with focus
• visible email draft
• current slide

Without interface grounding, the utterance remains ambiguous.

3.8.1 Selected Object#

The selected object is often the strongest grounding signal.

Example:

“Summarize this.”

If a paragraph is highlighted, “this” likely refers to the highlighted paragraph. If a file is selected, “this” may refer to the file. If no object is selected, the referent may be the current document or previous message.

Selection can bind:

• text
• rows
• columns
• cells
• images
• shapes
• slides
• files
• tickets
• emails
• calendar events
• map regions
• search results

3.8.2 Active Document#

The active document supplies context for references such as:

• “this draft”
• “the title”
• “the intro”
• “the conclusion”
• “the table”
• “the current section”
• “the second paragraph”
• “the chart on page 3”

Example:

“Make the introduction less formal.”

The active document tells the system which introduction is meant.

If multiple documents are open, active-document grounding may be insufficient. The system may need to distinguish between the focused document, most recently edited document, visible document, or explicitly named document.

3.8.3 Cursor Position#

Cursor position can ground insertion and local revision.

Examples:

Cursor grounding is useful but can be fragile. If the cursor was accidentally left in an unrelated place, the interpretation may be wrong.

3.8.4 Viewport State#

Viewport state is the visible region of an interface.

Example:

“Move the chart above the table.”

If only one chart and one table are visible, viewport state may resolve the referents. If the full document contains many charts and tables, visibility matters.

Viewport grounding applies to:

• documents
• spreadsheets
• dashboards
• maps
• design canvases
• code editors
• slide decks
• search results
• ticket boards

Viewport state can also influence salience. Visible objects are more likely to be intended by “this,” “that,” or “the one on the left.”

3.8.5 Highlighted Text#

Highlighted text is a common grounding anchor for revision requests.

Examples:

Highlighted text can narrow scope. Without highlighting, “make this clearer” might refer to the whole document, current paragraph, previous response, or active section.

3.8.6 Current UI Focus#

UI focus identifies the active element or current interaction mode.

Examples:

UI focus is not linguistic content, but it is part of the interpretive context for interface-mediated language.

3.9 Relationship Among Context Mechanisms#

The mechanisms in this section are related but distinct.

Context dependence
├── implicit context
│   └── broad unstated background assumptions
│
├── deixis
│   └── expressions anchored to speaker, time, place, discourse, or interface
│
├── coreference
│   └── multiple expressions referring to the same entity
│
├── ellipsis
│   └── omitted structure recovered from context
│
├── presupposition
│   └── background facts treated as already true
│
├── temporal/spatial grounding
│   └── time and place resolution
│
├── discourse grounding
│   └── binding to prior conversational state
│
└── interface grounding
    └── binding to active digital environment

They often co-occur.

Example:

“Actually, move that to next Friday instead.”

This includes:

Another example:

“Use the same format, but only for the selected rows.”

This includes:

3.10 Summary#

Context, reference, and grounding explain how natural-language expressions attach to the world, the task, the conversation, and the interface.

A compact model:

Natural-language meaning is situated.

It may depend on:
├── who is speaking
├── who is being addressed
├── when the utterance occurs
├── where the utterance occurs
├── what has already been said
├── what task is active
├── what object is visible or selected
├── what assumptions are shared
└── what prior state remains salient

The central point of Section 3 is that natural language is not self-contained. Meaning is often distributed across the utterance, the surrounding situation, prior discourse, task state, and the active interface.

Below is a document-ready draft for Sections 4-7. It keeps the structure you specified, assumes Sections 1-3 already cover layered meaning, indeterminacy/compression, and context/reference/grounding, and excludes boundaries/comparisons with formal systems. This draft extends the uploaded taxonomy and incorporates the critique additions around action fan-out, non-monotonicity, and repair semantics.

4. Communicative Force and Intent

Natural-language utterances do not merely describe states of affairs. They perform communicative actions.

A user may ask, request, command, warn, suggest, correct, confirm, evaluate, refuse, or imply a desired course of action without explicitly naming the action type. The communicative force of an utterance is therefore not always recoverable from syntax alone. It depends on surface form, context, speaker goals, discourse history, social stance, and the current task.

This section describes how utterances become communicative acts: how a sentence, fragment, or conversational move comes to count as something the speaker is trying to do.

4.1 Pragmatics in the Broad Sense#

Pragmatics concerns meaning in use. It asks how people infer intended meaning from context, shared assumptions, goals, conventions, and prior discourse.

The same sentence can perform different functions in different contexts. Conversely, the same communicative function can be expressed through many different surface forms.

Example:

“Can you send this to Anna?”

Literal content:

Are you able to send this to Anna?

Typical intended meaning:

Please send this to Anna.

The literal meaning is not false or irrelevant. Rather, it is only one layer of the utterance. The communicative force is usually a request, not a neutral inquiry about ability.

4.1.1 Meaning as use#

Meaning is not limited to what words conventionally denote. In natural language, meaning also includes what an utterance accomplishes in a situation.

Example:

“The deadline is tomorrow.”

Possible uses:

The sentence has stable semantic content: there is a deadline, and it is tomorrow relative to the relevant reference time. But its use varies. It may function as a reminder, warning, justification, criticism, or scheduling constraint.

This is why natural-language interpretation must account for communicative purpose, not only sentence meaning.

4.1.2 Literal meaning versus intended meaning#

Literal meaning is the conventional semantic content of an expression. Intended meaning is what the speaker is trying to communicate or accomplish with that expression in context.

Example:

“It’s cold in here.”

Literal meaning:

The temperature in the current place is low.

Possible intended meanings:

The utterance does not explicitly request an action. However, depending on the situation, it may strongly imply one.

Literal meaning and intended meaning can diverge in several ways:

The intended meaning does not replace the literal meaning. It is inferred from the literal meaning plus context.

4.1.3 Contextual inference#

Contextual inference is the process by which a listener uses surrounding information to infer what the speaker means.

Relevant context may include:

Example:

“This is a little too aggressive.”

Possible interpretation:

The utterance is grammatically a statement. In context, it functions as a revision request.

4.2 Speech Acts#

A speech act is an action performed through language.

Natural language does not only encode propositions. It allows speakers to perform acts such as requesting, commanding, warning, confirming, refusing, correcting, and evaluating. These acts are often inferred rather than explicitly labeled.

Example:

“No, use the other file.”

This utterance performs several acts at once:

Speech-act recognition is central to natural-language interpretation because the same propositional content can have different force.

Compare:

4.2.1 Request#

A request asks another participant to do something.

Requests may be direct:

“Send me the document.”

They may also be indirect:

“Could you send me the document?”

Or highly softened:

“When you have a chance, could you send me the document?”

Requests often contain politeness markers, optionality, or timing softeners, but these do not necessarily make the request weak. A polite request can still be operationally important.

Common request forms:

The last example is not grammatically a request, but in many contexts it functions as one.

4.2.2 Command#

A command directs another participant to perform an action, usually with stronger force than a request.

Examples:

“Delete the duplicate rows.” “Restart the service.” “Do not publish this version.”

Commands may be marked by imperative syntax, urgency, authority, or explicit prohibition.

Commands differ from requests in force, not always in grammatical form. The same imperative can function as a command, instruction, suggestion, or routine step depending on context.

Example:

“Try the second option.”

Possible force:

Commands require attention to authority, obligation, and risk. Some commands are routine; others require confirmation, especially when they affect important objects or irreversible outcomes.

4.2.3 Question#

A question seeks information, confirmation, clarification, or a decision.

Examples:

“When is the meeting?” “Is this the final version?” “Which Sarah do you mean?” “Are we comfortable with this risk?”

Questions can perform several functions:

Questions are not always neutral requests for information. They can imply criticism, warning, skepticism, or recommendation.

4.2.4 Suggestion#

A suggestion proposes a possible course of action without necessarily requiring it.

Examples:

“We could simplify this section.” “Maybe use the second chart.” “It might be better to wait until Monday.”

Suggestions often use modal expressions:

Suggestions can become stronger when combined with role, expertise, urgency, or repeated emphasis.

Example:

“I really think we should not change the API before launch.”

Although phrased as a recommendation, this may function as a strong warning or constraint.

4.2.5 Warning#

A warning alerts the listener to possible risk, harm, failure, or undesirable consequence.

Examples:

“This might break production.” “The client will see this if we publish it.” “That deadline is not realistic.” “We do not have approval yet.”

Warnings may be explicit or implicit.

Explicit warning:

“Be careful – this includes private data.”

Implicit warning:

“This includes private data.”

The second sentence does not contain the words “be careful,” but in many contexts it functions as a warning.

Warning force often depends on consequence framing:

A warning can override otherwise plausible interpretations. If an utterance marks high risk, later interpretation should treat the situation as more constrained.

4.2.6 Correction#

A correction revises a previous statement, assumption, action, or interpretation.

Examples:

“No, I meant the other file.” “Actually, use Monday, not Tuesday.” “That’s not the final version.” “I meant Sarah from legal.”

Corrections often target one of several dimensions:

Corrections are discourse-sensitive. Their meaning depends on what was previously said, inferred, selected, or done.

4.2.7 Confirmation#

A confirmation accepts, validates, or affirms a prior interpretation.

Examples:

“Yes.” “Exactly.” “That one.” “Correct.” “Yes, use that version.”

Confirmations may confirm different things:

Minimal confirmations are often underspecified. A bare “yes” may confirm the immediately prior question, the visible option, the proposed action, or the system’s interpretation. The target must be recovered from discourse state.

4.2.8 Refusal#

A refusal rejects a proposed action, claim, request, or assumption.

Examples:

“No.” “I can’t approve that.” “Let’s not do that.” “That won’t work.” “I’m not comfortable with this.”

Refusals may be direct or softened.

Refusals often imply an alternative path, even when unstated.

Example:

“Let’s not send this today.”

Possible implications:

Refusal can also function as a safety or boundary signal when a proposed interpretation would be inappropriate, risky, unauthorized, or unwanted.

4.2.9 Evaluation#

An evaluation expresses a judgment about quality, suitability, correctness, desirability, risk, or priority.

Examples:

“This is too long.” “Option B is safer.” “The argument is weak.” “That looks good.” “This feels risky.”

Evaluations often imply action.

Evaluations can carry different levels of commitment:

Evaluation is often where semantics, pragmatics, social stance, and intended action converge.

4.3 Indirect Speech Acts#

An indirect speech act occurs when the literal form of an utterance differs from its intended communicative function.

Indirectness is not a rare exception. It is a normal feature of natural language. Speakers often use indirect forms to preserve politeness, reduce imposition, signal uncertainty, avoid confrontation, or leave room for negotiation.

Example:

“Could you make this shorter?”

Literal form:

Question about ability.

Likely speech act:

Request to shorten the text.

4.3.1 Ability question as request#

Ability questions frequently function as requests.

Examples:

The ability question creates a polite frame. It allows refusal or negotiation while still communicating a desired action.

However, not every ability question is a request.

Example:

“Can you export this system to PDF?”

Possible interpretations:

The distinction depends on context and discourse state.

4.3.2 Observation as implied request#

An observation can imply a request when it identifies a problem, gap, or undesirable state.

Examples:

Observation-as-request is common in editing, design, planning, and task management contexts.

The implied request is stronger when:

• the speaker is reviewing an object;
• the listener is expected to modify it;
• the observation identifies a fixable defect;
• the current activity is revision-oriented;
• the statement follows a prior request for feedback or action.

4.3.3 Critique as revision request#

Critique often functions as an implicit instruction to revise.

Example:

“The opening is too vague.”

Possible intended meaning:

Revise the opening to be more specific.

Critique can target many dimensions:

Critique may be softened for social reasons:

“This might be a little too direct.”

The softening does not eliminate the revision request. It modulates the social force of the critique.

4.3.4 Preference as constraint#

A preference can function as a constraint on future interpretation or action.

Examples:

Preferences vary in strength. Some are weak tendencies; others are hard constraints.

Compare:

A natural-language interface must distinguish preference from requirement where possible, but it must also recognize that speakers often express requirements indirectly.

4.4 Implicature#

Implicature is meaning that is implied but not explicitly stated.

Implicature relies on assumptions about relevance, informativeness, cooperation, and shared context. Speakers often communicate more than they literally say, and listeners infer the unstated content.

Example:

“Some of the tests passed.”

Literal meaning:

At least some tests passed.

Common implicature:

Not all tests passed.

The implicature arises because, if all tests had passed, the speaker would normally be expected to say so.

4.4.1 Relevance implicature#

A relevance implicature arises when the listener assumes that the speaker’s utterance is relevant to the current situation.

Example:

“The client is on the call.”

Possible implicatures:

The utterance does not explicitly command caution, but its relevance in context may make that implication clear.

Other examples:

Relevance implicature is highly context-sensitive. The same utterance can imply different actions in different settings.

4.4.2 Scalar implicature#

A scalar implicature occurs when a weaker expression implies that a stronger one does not apply.

Example:

“Some users are affected.”

Common implication:

Not all users are affected.

The implication comes from a scale:

some < many < most < all

If the speaker knew that all users were affected, they would normally say “all.”

Common scalar terms:

Examples:

Scalar implicatures are defeasible. They can be cancelled.

Example:

“Some of the tests passed – in fact, all of them did.”

The second clause cancels the usual implication.

4.4.3 Consequential implicature#

A consequential implicature arises when an utterance implies practical consequences, risks, or next steps.

Example:

“The contract expires Friday.”

Possible implications:

The utterance itself reports a fact. But the fact is relevant because it changes what should happen next.

Other examples:

Consequential implicature is especially important when the speaker states a condition rather than a direct instruction.

4.4.4 Domain-specific implicature#

Domain-specific implicature depends on conventions, norms, or expectations within a particular field or activity.

Examples:

Domain-specific implicature can be opaque to outsiders. The words may be ordinary, but their significance depends on the domain.

Example:

“The API is public.”

In a software context, this may imply compatibility constraints, documentation requirements, versioning concerns, and higher risk for changes.

4.5 Relevance Filtering#

Natural language usually permits many possible interpretations. Relevance filtering is the process by which the listener selects the interpretation that best fits the current task, context, domain, and user goal.

Example:

“Is Jordan available tomorrow?”

Possible meanings of “available” include:

• free for a meeting;
• available for work;
• available by phone;
• available as a contractor;
• emotionally available;
• visible in a scheduling system;
• not blocked by permissions.

In a scheduling context, the relevant interpretation is usually:

Is Jordan free for a meeting tomorrow?

Natural-language interpretation therefore requires selecting among plausible meanings, not merely listing them.

4.5.1 Task-relevant interpretation#

Task-relevant interpretation selects the meaning that best fits the active task.

Example:

“Make it lighter.”

Possible meanings:

The same phrase changes meaning depending on the task frame.

Task relevance is often established by prior turns:

“Revise this email.” “Make it lighter.”

Here, “lighter” likely refers to tone, not color, weight, or file size.

4.5.2 Domain-relevant interpretation#

Domain-relevant interpretation selects the meaning that fits the specialized vocabulary and norms of the domain.

Example:

“The branch is stale.”

Possible meanings:

In a code-review context, the software interpretation is dominant.

Domain relevance can also resolve metaphor.

Example:

“The pipeline is blocked.”

In a data engineering context, this likely concerns a processing pipeline. In a construction context, it may refer to a literal pipe or project pipeline.

4.5.3 User-goal-relevant interpretation#

User-goal relevance selects the interpretation that best serves the user’s likely objective.

Example:

“I need something more polished.”

Possible meanings:

The phrase “more polished” is vague, but the user’s goal narrows the expected interpretation.

User-goal relevance may come from:

• explicit stated goals;
• prior preferences;
• current project;
• selected artifact;
• constraints already introduced;
• audience or use case.

4.5.4 Salience and suppression of irrelevant meanings#

Salience determines which meanings are prominent enough to consider. Relevance filtering also suppresses technically possible but contextually irrelevant meanings.

Example:

“Book a table.”

In a restaurant-planning context, “book” means reserve and “table” means restaurant seating. The possible interpretation “create a database table” is suppressed.

Suppression is not deletion. If later context changes, a suppressed interpretation can become relevant.

Example:

“Book a table in the reservations database.”

Now “table” may refer to a database object, and “book” may be interpreted as recording or creating an entry.

Salience can be affected by:

Relevance filtering is therefore both enabling and dangerous. It makes efficient communication possible, but it can also produce wrong assumptions when context is incomplete or misleading.

4.6 Action Fan-Out#

Action fan-out occurs when a single utterance implies more than one underlying action.

Natural language often compresses complex activity into a short phrase. A user may describe an outcome, process, or goal without listing every required step.

Example:

“Onboard the new contractor.”

Possible implied actions:

The utterance is a single speech act at the conversational level, but it may correspond to a multi-step activity at the task level.

Action fan-out is not implementation detail by itself. It is a semantic property of natural-language intent: speakers often name a high-level activity and expect its conventional sub-actions to be understood.

4.6.1 Single utterance to single action#

Some utterances map naturally to one primary action.

Examples:

Even simple action mappings may still require reference resolution, permissions, or clarification. But the intended action type is relatively narrow.

4.6.2 Single utterance to parameterized action#

Some utterances imply one action type but require parameters to be filled.

Example:

“Schedule a meeting with Alex.”

Primary action:

Schedule meeting.

Missing or inferable parameters:

The utterance names the action but leaves many arguments underspecified. This is normal in natural language.

4.6.3 Single utterance to workflow#

Some utterances name an outcome that conventionally requires a sequence of actions.

Examples:

Workflow-level utterances often combine intent, quality criteria, and domain expectations.

Example:

“Make this investor-ready.”

This may imply:

• sharpen the narrative;
• improve credibility;
• add metrics;
• reduce internal jargon;
• improve visual polish;
• anticipate objections;
• check consistency.

The phrase compresses a broad standard into a compact request.

4.6.4 Single utterance to multi-step orchestration#

Some utterances imply coordinated actions across multiple objects, people, times, or systems.

Examples:

At this level, the utterance does not only specify an action. It invokes a socially or institutionally recognized activity pattern.

Action fan-out creates several interpretation questions:

Action fan-out is one reason high-level natural-language intent cannot be treated as a single flat instruction.

4.7 Summary: How Utterances Become Communicative Acts#

Communicative force and intent explain how utterances become actions in context.

A sentence may look like a question but function as a request. A statement may function as a warning. A critique may function as a revision instruction. A preference may function as a constraint. A short phrase may imply a multi-step activity.

Key properties:

The central principle of this section is:

Natural-language interpretation requires identifying not only what an utterance says, but what the speaker is doing by saying it.

5. Discourse State and Conversational Mutation

Natural-language meaning evolves across turns.

Users rarely express all constraints, corrections, preferences, and clarifications in a single complete utterance. They refine meaning incrementally. They correct themselves. They revise earlier instructions. They add constraints, retract assumptions, shift priorities, and rely on prior conversational state.

Discourse is therefore not just a sequence of sentences. It is a changing state of commitments, referents, constraints, goals, assumptions, and salience.

This section describes how meaning is organized and mutated across conversational context.

5.1 Discourse Structure#

Discourse structure concerns how clauses, sentences, turns, and larger units relate to one another.

Natural language uses explicit markers, ordering, and contextual inference to express relations such as sequence, contrast, cause, condition, exception, elaboration, and summary.

Example:

“First summarize the report, then extract the risks, but skip the appendix unless it contains financial data.”

This utterance contains multiple discourse relations:

Discourse structure affects how instructions combine. It determines whether clauses are additive, contrastive, conditional, corrective, or subordinate.

5.1.1 Sequence#

Sequence indicates ordering among events, actions, or ideas.

Markers include:

first
then
next
after that
before
finally
once
after

Example:

“First clean the data, then generate the chart.”

Meaning:

Sequence can be temporal, procedural, narrative, or argumentative.

Examples:

Sequential interpretation matters because later actions may depend on earlier ones.

5.1.2 Contrast#

Contrast indicates difference, opposition, or correction between discourse units.

Markers include:

but
however
although
though
instead
rather
whereas
not X but Y

Example:

“The draft is clear, but the tone is too casual.”

Meaning:

Contrast often shifts focus to the second element.

Example:

“Use the shorter version, not the more detailed one.”

The contrast identifies the intended choice by excluding an alternative.

Contrast can also indicate preference:

“Cost matters, but reliability matters more.”

The second criterion has higher priority.

5.1.3 Cause#

Cause indicates that one event, state, or fact explains another.

Markers include:

because
since
as
due to
therefore
so
that means
as a result

Example:

“The import failed because the file is malformed.”

Meaning:

Causal relations support diagnosis, explanation, and prioritization.

Examples:

Cause can be explicit or inferred.

Example:

“The file is locked. I can’t edit it.”

The causal connection is not marked by “because,” but it is strongly implied.

5.1.4 Condition#

Condition indicates that one action, claim, or outcome depends on another.

Markers include:

if
unless
provided that
as long as
only if
when
in case

Example:

“If Alex approves it, send it to the client.”

Meaning:

Conditional structure can affect permission, timing, and action eligibility.

Examples:

Conditional interpretation is often critical because it determines whether an action should occur at all.

5.1.5 Exception#

Exception indicates that a general rule, instruction, or description does not apply in a specific case.

Markers include:

except
except for
unless
other than
apart from
but not
excluding

Example:

“Invite everyone except contractors.”

Meaning:

Exceptions can apply to people, objects, time ranges, actions, or conditions.

Examples:

Exceptions are easy to miss because they may appear after a broad instruction. The later exception narrows or overrides the earlier generalization.

5.1.6 Elaboration#

Elaboration adds detail, specification, examples, or clarification to a prior statement.

Markers include:

specifically
in particular
for example
that is
meaning
namely
especially

Example:

“The issue is access control. Specifically, external users can view internal comments.”

Meaning:

Elaboration narrows or enriches interpretation.

Examples:

Elaboration is often used to resolve vagueness or provide operational criteria.

5.1.7 Summary#

Summary compresses prior discourse into a shorter formulation.

Markers include:

in short
overall
basically
the main point is
to summarize
bottom line

Example:

“In short, the migration is risky but necessary.”

A summary may preserve only the most salient claims, dropping details that were previously discussed.

Summaries can function as:

A summary can also establish the current discourse state. Once a speaker summarizes, later references may attach to the summary rather than to every prior detail.

5.2 Discourse Markers#

Discourse markers are words or phrases that guide how one part of discourse relates to another.

They do not always contribute ordinary propositional content. Instead, they signal structure, stance, correction, contrast, inference, exception, example, or compression.

Examples:

however
therefore
actually
instead
unless
for example
in short

A discourse marker can change how an utterance updates conversational state.

Compare:

5.2.1 however#

“However” marks contrast, qualification, or partial opposition.

Example:

“The proposal is strong. However, the timeline is unrealistic.”

Function:

“However” often shifts interpretive weight to the second statement. The second statement may identify the main caveat.

Other examples:

5.2.2 therefore#

“Therefore” marks inference or conclusion.

Example:

“The client has not approved the draft. Therefore, do not publish it yet.”

Function:

“Therefore” signals that the second clause follows from the first. It often converts facts into action-relevant conclusions.

Other examples:

5.2.3 actually#

“Actually” often marks correction, revision, contrast with expectation, or change of mind.

Examples:

“Actually, use the other file.” “Actually, make it more concise.” “Actually, I meant next Thursday.”

Functions:

“Actually” is a strong signal that prior discourse state may need to be updated or partially invalidated.

5.2.4 instead#

“Instead” marks replacement.

Example:

“Don’t send the email. Instead, save it as a draft.”

Meaning:

“Instead” differs from simple addition. It usually means the new action supersedes an earlier action, option, or plan.

Examples:

5.2.5 unless#

“Unless” marks a negative condition or exception.

Example:

“Send it tomorrow unless legal objects.”

Meaning:

“Unless” can make an action conditional even when the main clause appears directive.

Examples:

“Unless” requires tracking both default behavior and exception conditions.

5.2.6 for example#

“For example” introduces an instance, illustration, or non-exhaustive member of a broader category.

Example:

“Add stronger evidence, for example customer quotes or usage data.”

The examples illustrate possible evidence types. They may not be the only acceptable options.

Interpretation question:

Are the examples mandatory, preferred, illustrative, or exhaustive?

In many contexts, “for example” indicates non-exhaustiveness. But in task settings, examples may still guide expected output.

Other examples:

5.2.7 in short#

“In short” introduces a compressed summary or conclusion.

Example:

“In short, the plan is viable but risky.”

Function:

“In short” often signals that the speaker is establishing the high-level takeaway. Later interpretation may treat this summary as the current salient state.

Other examples:

5.3 Repair and Correction#

Repair occurs when a speaker fixes, revises, clarifies, or replaces part of a prior utterance or interpretation.

Natural language supports repair as a normal part of communication. Speakers often do not produce final, fully specified meaning in one pass. They rely on later turns to adjust reference, constraints, scope, priority, and intent.

Examples:

“No, I meant the other file.” “Actually, make it shorter, not more formal.” “Sorry, I meant Sarah from legal.” “Let me rephrase that.”

Repair may target:

5.3.1 Self-repair#

Self-repair occurs when the speaker corrects or revises their own utterance.

Examples:

“Send it to Alex – sorry, I mean Jordan.” “Schedule it for Friday. Actually, Monday is better.” “Make it more formal – no, more concise.”

Self-repair may occur immediately or after intervening discourse.

Immediate self-repair:

“Use the blue version – sorry, the green one.”

Delayed self-repair:

“Earlier I said to use the green version. Actually, use the blue one.”

Self-repair requires updating prior state. The corrected material may replace, narrow, or cancel the original material.

5.3.2 Other-repair#

Other-repair occurs when another participant identifies or corrects a problem in the speaker’s utterance or interpretation.

Examples:

“Do you mean Alex Chen or Alex Rivera?” “That meeting is already cancelled.” “There are two files with that name.” “I think you meant next Tuesday, not this Tuesday.”

Other-repair may take the form of:

Other-repair is not merely error handling. It is part of collaborative meaning construction.

5.3.3 Referent correction#

Referent correction changes which entity an expression refers to.

Examples:

“Not that file – the one from yesterday.” “I meant Sarah from legal, not Sarah from finance.” “Use the second chart, not the first.” “No, the other Alex.”

Referent correction often relies on contrastive language:

not X, Y
the other one
the one from yesterday
the second one
that one, not this one

Referent correction can also affect previous actions or planned actions. If the wrong referent was selected, subsequent instructions may need to be reattached to the corrected referent.

5.3.4 Parameter correction#

Parameter correction changes a value associated with an action, object, or constraint.

Examples:

“Use Monday, not Friday.” “Make it 30 minutes instead of an hour.” “Send it to the internal list, not the full list.” “Actually, use the latest version.”

Parameters may include:

Parameter correction is often local: it changes one part of an otherwise stable intent.

5.3.5 Intent correction#

Intent correction changes the action or communicative force itself.

Examples:

“I’m not asking you to send it, just summarize it.” “Don’t rewrite it – explain what’s wrong.” “I only wanted a recommendation, not a full plan.” “I meant compare them, not choose one.”

Intent correction is deeper than parameter correction. The object or topic may remain the same, but the requested act changes.

Compare:

Intent correction can invalidate an entire action path.

5.3.6 Scope correction#

Scope correction changes the range of objects, people, cases, or conditions to which an instruction applies.

Examples:

“Only the urgent ones.” “Not all customers – just enterprise accounts.” “Apply that to the summary, not the whole document.” “Use it for internal emails only.”

Scope correction can narrow, broaden, or shift the target set.

Scope correction is especially important because users often express broad instructions first and then constrain them later.

5.4 Incremental Refinement#

Incremental refinement occurs when a speaker gradually modifies an earlier request, search, plan, or interpretation.

Example:

“Find hotels in Madrid.” “Under $200.” “Near the conference venue.” “Actually, prioritize quiet ones.”

Each turn modifies the active task. Later utterances may be fragmentary because they rely on the prior state.

Incremental refinement is normal because users often discover their requirements through interaction. Natural language allows the task representation to be built progressively.

5.4.1 Adding constraints#

Adding constraints narrows or conditions the current task.

Examples:

“Only include open-source options.” “Make it under 500 words.” “Use examples from healthcare.” “Avoid anything too technical.” “Include the risks.”

Added constraints attach to an existing task or object.

Example:

“Draft a message to the team.” “Make it more direct.” “Also mention the Friday deadline.”

The later utterances do not start new tasks. They refine the original drafting task.

5.4.2 Removing constraints#

Removing constraints relaxes a prior condition.

Examples:

“Actually, budget does not matter.” “Ignore the word limit.” “It does not have to be in table format.” “You can include vendors after all.” “No need to keep it formal.”

Constraint removal can broaden the solution space or change the interpretation of earlier instructions.

Example:

“Find options under $100.” “Actually, remove the price limit.”

The price constraint no longer applies.

5.4.3 Narrowing the search space#

Narrowing reduces the set of acceptable options.

Examples:

“Only show remote roles.” “Just the last quarter.” “Focus on enterprise customers.” “Use the top three.” “Exclude anything deprecated.”

Narrowing often appears in elliptical fragments:

“Only 14-inch ones.” “Just internal users.” “No vendors.” “Last month only.”

These fragments depend on the current task state for interpretation.

5.4.4 Broadening the search space#

Broadening expands the set of acceptable options.

Examples:

“Include contractors too.” “Show options outside New York.” “It can be longer if needed.” “Consider paid tools as well.” “Include last year’s data too.”

Broadening may cancel an earlier narrowing constraint.

Example:

“Only open-source tools.” “Actually, include commercial tools if they are clearly better.”

The new instruction broadens the candidate set while adding a condition.

5.4.5 Reprioritizing criteria#

Reprioritizing changes the relative importance of criteria.

Examples:

“Actually, speed matters more than cost.” “Prioritize reliability over elegance.” “Make accuracy the main thing.” “Tone matters more than length here.” “Focus less on features and more on risk.”

Reprioritization may not add or remove constraints. Instead, it changes how tradeoffs should be resolved.

Example:

“Find a cheap laptop with strong battery life.” “Actually, battery life matters more than price.”

The criteria remain, but their ordering changes.

5.5 Mixed Initiative#

Mixed initiative means that more than one participant can guide the interaction.

In natural conversation, the user does not always provide a complete instruction, and the other participant does not merely execute. Either side may ask clarifying questions, propose interpretations, identify constraints, suggest next steps, or reframe the problem.

Example:

User: “Help me compare these options.” Response: “Do you care more about cost, risk, or speed?” User: “Risk first.”

The interaction jointly constructs the task.

5.5.1 User-led initiative#

User-led initiative occurs when the user controls the direction of the interaction.

Examples:

“Start with the risks.” “Now compare the pricing.” “Ignore that and focus on implementation.” “Ask me one question at a time.” “Use the second option.”

User-led initiative may introduce:

5.5.2 System-led clarification#

System-led clarification occurs when the other participant asks for information needed to interpret or proceed.

Examples:

“Which Alex do you mean?” “Should this be formal or casual?” “Do you want a summary or a full rewrite?” “Should I include external recipients?” “Is the deadline today or tomorrow?”

Clarification is appropriate when the missing information materially affects interpretation, risk, or outcome quality.

Clarification questions can target:

5.5.3 Collaborative narrowing#

Collaborative narrowing occurs when participants jointly reduce uncertainty or narrow a broad request.

Example:

User: “Find a good venue.” Response: “Should I prioritize price, capacity, or location?” User: “Capacity first, then location.”

The task becomes clearer through interaction.

Collaborative narrowing is useful when:

• the initial request is broad;
• multiple reasonable interpretations exist;
• user preferences are unknown;
• tradeoffs are unavoidable;
• proceeding without clarification would require unstable assumptions.

5.5.4 Reframing and negotiation#

Reframing changes the way the task or problem is represented. Negotiation adjusts goals, constraints, or acceptable outcomes.

Examples:

“Instead of rewriting the whole memo, maybe revise the executive summary first.” “The issue may not be tone; it may be missing evidence.” “Rather than scheduling a meeting, we could send a decision memo.” “If speed matters most, we should drop the custom design.”

Reframing can be initiated by either participant.

Negotiation may involve:

Reframing is a discourse-level operation because it changes the interpretation of the entire task, not just a local parameter.

5.6 Redundancy and Emphasis#

Natural language uses repetition and redundancy for more than error tolerance. Redundancy can signal importance, urgency, contrast, emotional force, correction, or priority.

Example:

“I need this today, before end of day, ideally by 5.”

The deadline is repeated in several forms. The repetition reinforces urgency and reduces the chance of misinterpretation.

Redundancy is not automatically noise. It may be semantically meaningful.

5.6.1 Repetition as priority signal#

Repetition can indicate that something matters more than other criteria.

Examples:

“Keep it simple. Very simple.” “The main issue is trust – trust is the thing to solve.” “Accuracy matters here. Accuracy over speed.” “Do not change the API. No API changes.”

Priority repetition often identifies a dominant constraint.

Example:

“Reliability, reliability, reliability.”

This may mean reliability should override competing goals.

5.6.2 Repetition as urgency signal#

Repetition can intensify time pressure.

Examples:

“Today, not tomorrow.” “I need this now – really now.” “Before the meeting. Before 3 PM.” “This has to go out today.”

Urgency repetition may combine with deadlines, consequences, or emotional stance.

Example:

“We need this before the client call, otherwise we lose the window.”

The repetition plus consequence framing signals high priority.

5.6.3 Repetition as contrast#

Repetition can mark contrast between accepted and rejected alternatives.

Examples:

“The internal version, not the external version.” “Monday, not Tuesday.” “The summary, not the full report.” “Shorter, not less detailed.”

Contrastive repetition clarifies which dimension is being changed.

Example:

“Make it simpler, not shorter.”

This indicates that simplification, not length reduction, is the intended revision direction.

5.6.4 Repetition as repair#

Repetition can repair a misheard, misunderstood, or wrongly inferred element.

Examples:

“No, Alex Chen. Alex Chen from legal.” “The Q4 report – Q4, not Q3.” “I said internal users, internal only.” “Use the draft version. The draft version.”

Repair repetition increases salience and reduces ambiguity.

Repetition may also indicate frustration if prior corrections were missed.

Example:

“No, not that one. The other one. The other one from yesterday.”

Here repetition functions as both repair and affective signal.

5.7 Non-Monotonicity#

Non-monotonicity means that later discourse can invalidate, revise, weaken, or cancel earlier discourse.

In natural conversation, adding new information does not always simply add to the previous state. It may change what counts as the current instruction, current referent, current constraint, or current belief.

Example:

“Make it larger. Actually, leave it as-is.”

The second sentence cancels the first. The conversational state does not contain both instructions as equally active.

Non-monotonicity is central to repair, correction, and refinement.

5.7.1 Later turns invalidating earlier turns#

A later turn may invalidate an earlier instruction or assumption.

Examples:

“Send it to Alex. Actually, don’t send it yet.” “Use the first version. No, use the second.” “Schedule it for Friday. Wait, Monday is better.” “Summarize the whole report. Actually, just the risks.”

The later turn changes the active state.

Invalidation may be explicit:

actually
no
wait
scratch that
instead
I changed my mind

Or implicit:

“Use the second version.”

If a first version had previously been selected, this may replace it.

5.7.2 Retraction of extracted parameters#

A speaker may retract a parameter that was previously introduced.

Examples:

“Make it under 500 words. Actually, length does not matter.” “Schedule it with Maria. No, leave Maria out.” “Use the New York data. Actually, exclude New York.” “Send it tomorrow. Wait, do not send it yet.”

Retraction differs from overwriting. The earlier parameter may become inactive rather than replaced by a new specific value.

Example:

“Under $100.” “Actually, no price limit.”

The price parameter is not changed from $100 to another number. It is removed.

5.7.3 Soft cancellation#

Soft cancellation weakens or partially cancels a prior instruction without an explicit contradiction.

Examples:

“Maybe don’t worry about the formatting.” “Actually, the tone is more important than the length.” “Leave the details if they help.” “It does not have to be that short.”

Soft cancellation often uses hedges or reframing rather than direct negation.

Example:

“Make it concise.” “But don’t cut the examples.”

The second utterance partially cancels one likely interpretation of “concise”: removing examples.

5.7.4 Partial revision#

Partial revision changes one component while preserving others.

Examples:

“Keep the structure, but make the tone warmer.” “Use the same content, just shorten the intro.” “Leave the chart, but update the labels.” “Keep option B, but change the rollout date.”

Partial revision requires distinguishing stable elements from revised elements.

Partial revision is one of the most common conversational operations in editing, planning, and decision support.

5.7.5 Conflicting updates#

Conflicting updates occur when later information is incompatible with earlier information, but the discourse does not clearly resolve which should dominate.

Example:

“Keep it under 300 words.” “Include all technical details.”

These constraints may conflict.

Other examples:

Conflicting updates require resolution through context, priority, clarification, or repair.

5.8 Conversational Memory Limits#

Conversational meaning depends on memory, salience, and state tracking. But discourse context is limited. References can become stale, priorities can be forgotten, constraints can conflict, and earlier assumptions can become unreliable.

Memory limits are not only technical. Human conversations also suffer from drift, overload, and loss of salience.

5.8.1 Stale references#

A stale reference points to an object or state that is no longer current, salient, available, or valid.

Examples:

“Use that one.” “Same as before.” “The previous version.” “Move it to tomorrow.”

These references may become stale if:

• many turns have passed;
• multiple candidates have been introduced;
• the visible object has changed;
• a previous selection was replaced;
• the referenced item no longer exists;
• the task frame has shifted.

Stale references require caution because the speaker may assume a referent is still obvious when it is no longer uniquely recoverable.

5.8.2 Lost salience#

Lost salience occurs when a previously important entity, constraint, or goal is no longer prominent in the discourse.

Example:

“Make it shorter.” Several turns later: “Also include the full methodology.”

The earlier brevity constraint may still matter, but it may no longer be salient.

Lost salience can affect:

Salience must be maintained, refreshed, or checked when needed.

5.8.3 Conflicting constraints#

Conflicting constraints arise when active discourse state contains requirements that cannot all be satisfied.

Examples:

“Make it comprehensive but under 100 words.” “Use simple language but keep all legal precision.” “Do it today, but wait for next week’s data.” “Send it to everyone, but keep it confidential.”

Some conflicts are hard contradictions. Others are tradeoffs.

Conflicts often require clarification, prioritization, or restatement of the current active constraints.

5.8.4 Overloaded discourse context#

Overloaded discourse context occurs when too many entities, constraints, revisions, or topics are active at once.

Example:

“Use the second option, but only for internal users, except contractors, unless they are in the pilot group, and keep the old version for Europe until legal approves the new terms.”

This utterance may be interpretable, but it places heavy demands on state tracking.

Overload signs include:

• many candidate referents;
• nested exceptions;
• multiple unresolved constraints;
• repeated corrections;
• long-distance pronouns;
• topic shifts;
• changing priorities;
• ambiguous “this,” “that,” or “it.”

When discourse state is overloaded, interpretation becomes fragile. Summarizing the current state can become necessary.

5.9 Summary: How Meaning Evolves Across Turns#

Discourse state explains how meaning persists, changes, and breaks across conversation.

Key properties:

The central principle of this section is:

Natural-language meaning is dynamic: later turns can extend, narrow, repair, replace, or invalidate earlier meanings.

6. Social, Epistemic, and Priority Signals

Natural-language utterances carry more than propositional content and task intent. They also encode stance.

A speaker may signal politeness, hesitation, certainty, doubt, urgency, authority, reluctance, preference, emotional attitude, or strength of commitment. These signals affect interpretation. They help determine whether an utterance is a hard requirement, weak suggestion, tentative inference, urgent warning, softened criticism, or socially managed refusal.

This section describes the signals that shape how strongly, urgently, authoritatively, or cautiously an utterance should be interpreted.

6.1 Politeness and Face Management#

Politeness and face management concern how speakers preserve social relations, reduce imposition, soften criticism, and allow room for disagreement or refusal.

Politeness does not mean the utterance is unimportant. A softened request may still be a real request. A mitigated criticism may still require revision.

Example:

“Could you maybe make this a bit shorter?”

This is polite and softened, but the intended action is still likely:

Make this shorter.

6.1.1 Softened requests#

Softened requests reduce directness or imposition.

Examples:

Softeners include:

could
would
maybe
possibly
when you have a chance
if you do not mind
I was wondering if

Softening affects social tone, not necessarily task importance.

6.1.2 Indirect criticism#

Indirect criticism expresses a negative evaluation without stating it bluntly.

Examples:

Indirect criticism often implies a revision request.

Example:

“This may be a little too casual for the client.”

Likely interpretation:

Make it more formal before sending to the client.

6.1.3 Deference#

Deference signals respect for authority, expertise, role, seniority, or ownership.

Examples:

“You may already have considered this, but…” “I defer to legal on this.” “If you think it makes sense…” “With your approval, we can proceed.” “I may be missing context.”

Deference can reduce apparent commitment strength even when the underlying concern is important.

Example:

“I may be missing something, but this looks risky.”

Despite the hedge, the utterance may function as a serious warning.

6.1.4 Mitigation#

Mitigation weakens the apparent force of an utterance.

Common mitigators:

a little
somewhat
kind of
maybe
I think
it seems
possibly
not ideal
could be better

Examples:

Mitigation is often used to preserve face, avoid overclaiming, or reduce conflict.

6.1.5 Face-saving formulations#

Face-saving formulations allow a speaker to express disagreement, correction, or criticism without directly threatening the listener’s competence or status.

Examples:

“We may be looking at different versions.” “I think there may be a mismatch here.” “Maybe I explained that poorly.” “Let’s try a different framing.” “This might be clearer if we reorganize it.”

Face-saving formulations often distribute responsibility.

Compare:

These formulations alter social interpretation while preserving the underlying repair or revision function.

6.2 Epistemic Stance#

Epistemic stance indicates how certain the speaker is, what evidence they have, and how strongly they commit to a claim.

Examples:

“This is broken.” “This seems broken.” “This might be broken.” “Apparently, this is broken.” “I know this is broken.”

Each version expresses a different relationship between the speaker and the claim.

6.2.1 Certainty#

Certainty signals strong commitment to a proposition.

Examples:

definitely
clearly
certainly
I know
there is no doubt
this is

Examples in context:

Certainty can increase the weight of an utterance, but it does not guarantee correctness. It indicates the speaker’s stance toward the claim.

6.2.2 Uncertainty#

Uncertainty signals weak or incomplete commitment.

Examples:

maybe
possibly
might
could
I am not sure
it seems
it looks like
I think

Examples:

Uncertainty can indicate that further verification is needed or that the speaker is leaving room for correction.

6.2.3 Inference#

Inference signals that the speaker is drawing a conclusion from evidence rather than reporting direct knowledge.

Examples:

it looks like
it seems
I assume
I infer
that suggests
probably
given that

Example:

“The invite is gone, so I assume the meeting was cancelled.”

Structure:

Inference is defeasible. New information can overturn it.

Example:

“The invite is gone, so I assumed the meeting was cancelled – but it was actually moved to another calendar.”

6.2.4 Reported knowledge#

Reported knowledge indicates that the speaker’s claim comes from another source.

Examples:

apparently
I heard
according to
they said
the report says
legal mentioned

Examples:

Reported knowledge carries source-dependent reliability. “Legal said” may carry more institutional force than “I heard.”

6.2.5 Evidence strength#

Evidence strength indicates how well-supported a claim is.

Examples:

Evidence strength affects whether a claim should be treated as fact, hypothesis, warning, or candidate explanation.

6.3 Affective Stance#

Affective stance expresses emotional or evaluative attitude.

It can be explicit:

“I’m frustrated with this process.”

Or implicit:

“Great, another last-minute change.”

Affective stance affects how an utterance should be interpreted interpersonally. It may signal urgency, dissatisfaction, risk sensitivity, reluctance, approval, or escalation.

6.3.1 Frustration#

Frustration signals dissatisfaction, friction, repeated failure, or unmet expectations.

Examples:

“This is frustrating.” “We keep going in circles.” “I already explained this.” “This still is not what I asked for.”

Frustration may indicate:

Frustration often co-occurs with repair and repetition.

6.3.2 Concern#

Concern signals perceived risk, caution, or unease.

Examples:

“I’m worried about this migration.” “This feels risky.” “I’m concerned the client will misunderstand this.” “This may create problems later.”

Concern may function as warning, objection, or request for mitigation.

Example:

“I’m concerned this sounds too definitive.”

Likely implied action:

Qualify or soften the claim.

6.3.3 Skepticism#

Skepticism signals doubt, challenge, or weak acceptance.

Examples:

“Are we sure about that?” “I’m not convinced.” “That seems optimistic.” “Do we have evidence for this?” “I would not assume that.”

Skepticism may target:

Skepticism often invites justification, evidence, revision, or caution.

6.3.4 Satisfaction#

Satisfaction signals acceptance, approval, or positive evaluation.

Examples:

“This looks good.” “That works.” “Much better.” “I like this version.” “This is close.”

Satisfaction can imply different next steps:

Satisfaction is not always final approval. Phrases such as “close” or “better” may indicate progress without completion.

6.3.5 Reluctance#

Reluctance signals hesitation, discomfort, or unwillingness.

Examples:

“I’d rather not do that.” “I’m not comfortable with this.” “I hesitate to send this.” “I do not love that option.” “If we have to, but it is not ideal.”

Reluctance can signal a weak refusal, risk concern, preference, or need for approval.

Example:

“I’m not comfortable sending this externally yet.”

Likely meaning:

Do not send externally until the concern is resolved.

6.4 Modality#

Modality expresses possibility, necessity, permission, obligation, prohibition, likelihood, or ability.

Modal expressions shape how strongly an utterance constrains interpretation.

Examples:

can
could
may
might
must
should
need to
have to
allowed to
not allowed to

Modality interacts with authority, certainty, politeness, and commitment strength.

6.4.1 Possibility#

Possibility indicates that something may occur, may be true, or may be an available option.

Examples:

“This could work.” “We might use the second option.” “It may be delayed.” “There is a chance the numbers change.”

Possibility is weaker than commitment.

Compare:

Possibility may invite exploration rather than execution.

6.4.2 Necessity#

Necessity indicates that something is required or unavoidable.

Examples:

“We need to update the timeline.” “This has to be approved first.” “The report must include the risk section.” “You need to remove the private data.”

Necessity can be practical, logical, institutional, social, or safety-related.

6.4.3 Permission#

Permission indicates that an action is allowed.

Examples:

“You can share this internally.” “You may use the shorter version.” “It is okay to skip the appendix.” “Feel free to include examples.”

Permission does not always imply recommendation.

Compare:

Permission can also be conditional:

“You can send it if Maria approves.”

6.4.4 Obligation#

Obligation indicates that an action should or must be performed.

Examples:

“You should update the owner.” “We have to notify the client.” “The draft must include the disclaimer.” “Please make sure legal reviews this.”

Obligation strength varies:

Obligation can come from speaker authority, policy, task logic, social expectation, or risk.

6.4.5 Prohibition#

Prohibition indicates that an action should not or must not occur.

Examples:

“Do not share this externally.” “You must not delete those records.” “We should avoid changing the API.” “Do not include private data.” “This cannot go to the client yet.”

Prohibition strength varies:

Prohibitions often override competing requests unless explicitly revised.

6.5 Authority and Role#

Authority and role affect how utterances should be weighted.

A statement from a decision-maker, owner, reviewer, legal approver, manager, customer, or affected party may carry different force from the same statement by someone without that role.

Example:

“This must be approved before launch.”

The force of this utterance depends partly on who says it and what authority they have.

6.5.1 Speaker authority#

Speaker authority concerns the speaker’s right or standing to make claims, issue instructions, grant permission, or impose constraints.

Examples:

Speaker authority can be explicit or inferred from context.

6.5.2 Addressee authority#

Addressee authority concerns whether the listener has the ability, permission, or role required to act.

Examples:

“Can you approve this?” “Please publish the update.” “Could you grant Alex access?” “Can you make the final call?”

These utterances presuppose or ask about the addressee’s authority.

If the addressee lacks authority, the appropriate interpretation may be referral, refusal, clarification, or escalation.

6.5.3 Institutional role#

Institutional roles shape rights, obligations, and expected interpretation.

Examples:

Role-based authority can make indirect language stronger.

Example:

Legal reviewer: “I’m not comfortable with this wording.”

Likely interpretation:

Do not use this wording without revision or approval.

6.5.4 Ownership#

Ownership concerns who is responsible for an object, decision, task, or domain.

Examples:

“This is Maria’s doc.” “Alex owns the rollout.” “The design team owns the final call.” “Finance owns the budget assumptions.”

Ownership affects who can approve, revise, delegate, or object.

Example:

“Ask Priya before changing that section – she owns the policy language.”

The utterance implies that Priya’s role constrains permissible changes.

6.5.5 Approval rights#

Approval rights concern who can authorize an action or mark something as final.

Examples:

“This needs legal approval.” “Only the owner can approve this.” “Wait for client sign-off.” “Do not send until Maria confirms.” “Finance has to approve the budget.”

Approval rights often create conditional discourse states.

Example:

“Send it after legal approves.”

The action is not simply “send it.” It is conditional on approval.

6.6 Urgency and Priority#

Urgency and priority indicate how soon something matters and how it should be ranked against competing goals.

Urgency concerns time sensitivity. Priority concerns relative importance. They often overlap but are not identical.

Example:

“This is important, but not urgent.”

Meaning:

6.6.1 Explicit deadlines#

Explicit deadlines specify a time, date, or event boundary.

Examples:

“By 5 PM.” “Before the meeting.” “No later than Friday.” “Before launch.” “By the end of the quarter.”

Explicit deadlines still require grounding. “Friday” depends on the relevant week, and “before the meeting” depends on which meeting.

Deadlines may be absolute or relative:

6.6.2 Relative deadlines#

Relative deadlines express urgency without exact time.

Examples:

soon
ASAP
when you can
as soon as practical
before too long
later
eventually

Relative deadlines depend on context.

Example:

“Send this soon.”

Possible interpretations:

Relative urgency is inherently vague. It may require inference or clarification when timing matters.

6.6.3 Consequence framing#

Consequence framing signals urgency or priority by describing what happens if action is delayed, omitted, or mishandled.

Examples:

“Otherwise we miss the launch.” “The client is waiting.” “If this goes out wrong, legal will object.” “We need this before procurement closes.” “Without this, the team is blocked.”

Consequence framing can make an otherwise ordinary request urgent.

Example:

“Can you review this today? The client call is tomorrow morning.”

The second sentence increases urgency.

6.6.4 Repeated emphasis#

Repeated emphasis reinforces importance or urgency.

Examples:

“Today, not tomorrow.” “This is really important.” “No changes to the public API – none.” “Please keep this very simple.” “We need a final answer, not a draft.”

Repetition may indicate priority, frustration, or risk sensitivity.

Repeated emphasis should not be automatically collapsed as duplicate content. It can change the weight of an instruction.

6.6.5 Priority tradeoffs#

Priority tradeoffs indicate how to resolve competing goals.

Examples:

“Speed matters more than polish.” “Accuracy over brevity.” “Keep it simple, even if it is less complete.” “Cost matters, but reliability matters more.” “Do not optimize for elegance; optimize for maintainability.”

Tradeoffs are central when multiple criteria cannot be maximized at once.

Common tradeoff pairs:

Priority signals help determine which interpretation is preferred when constraints conflict.

6.7 Commitment Strength#

Commitment strength indicates how strongly the speaker endorses a claim, preference, judgment, or instruction.

Two utterances may have similar content but different force.

Compare:

Commitment strength is closely related to epistemic stance, modality, and authority.

6.7.1 Strong assertion#

A strong assertion expresses high confidence or firm commitment.

Examples:

“This is incorrect.” “The launch date is Friday.” “Option B is the safest choice.” “The customer will reject this.” “This must be fixed before release.”

Strong assertions often use direct declarative forms, certainty markers, or strong modals.

Strong assertions may still be wrong. The point is not truth, but speaker commitment.

6.7.2 Tentative assertion#

A tentative assertion expresses a claim while leaving room for revision.

Examples:

“This seems incorrect.” “I think option B is safer.” “It looks like the deadline moved.” “This may be the latest version.” “I suspect the issue is authentication.”

Tentative assertions are useful when evidence is incomplete.

They may invite verification, clarification, or alternative interpretations.

6.7.3 Preference#

A preference expresses desire or ranking without necessarily creating a requirement.

Examples:

“I prefer option A.” “I’d rather keep this short.” “I like the second version better.” “I would avoid a formal tone.” “I prefer not to use vendors.”

Preferences can become constraints when they are strong, repeated, role-backed, or tied to consequences.

Example:

“I really prefer not to use vendors because procurement will delay us.”

This is stronger than a simple aesthetic preference.

6.7.4 Recommendation#

A recommendation proposes what should be done.

Examples:

“We should use option B.” “I recommend delaying launch.” “It would be better to revise this first.” “You may want to add a disclaimer.” “I’d suggest confirming with legal.”

Recommendations vary in strength.

Recommendations often sit between preference and obligation.

6.7.5 Hard requirement#

A hard requirement expresses a non-negotiable or strongly binding constraint.

Examples:

“Do not share this externally.” “This must include the disclaimer.” “Only use approved vendors.” “The final answer has to be under 500 words.” “We need legal approval before sending.”

Hard requirements often use:

must
need to
have to
only
do not
required
not allowed

Hard requirements should be treated as active constraints unless later revised or invalidated.

6.8 Summary: How Stance Changes Interpretation#

Social, epistemic, and priority signals shape how utterances should be interpreted beyond their literal content.

Key properties:

The central principle of this section is:

Natural-language meaning includes stance: how strongly, urgently, confidently, authoritatively, or socially the speaker presents the utterance.

7. Failure, Risk, and Repair Semantics

Natural-language interpretation can fail.

Failure does not always mean the utterance is malformed. Natural language often leaves information implicit, compresses detail, relies on context, and permits multiple plausible interpretations. In many cases, this works well. In other cases, the same properties create risk.

This section describes common interpretive failures, how severe ambiguity can be, how confidence affects commitment, and when repair is required before proceeding.

7.1 Interpretive Failure Types#

Interpretive failure occurs when the meaning inferred from an utterance does not match the speaker’s intended meaning or the situation’s requirements.

Failure may involve the wrong referent, wrong intent, missing constraint, stale context, hallucinated context, over-execution, or under-execution.

Example:

“Send it to Alex.”

Possible failure points:

Natural language often allows these details to remain unstated. The risk depends on whether the missing or ambiguous information matters.

7.1.1 Wrong referent#

Wrong referent failure occurs when an expression is attached to the wrong entity.

Examples:

“Send that to Maria.” “Open the other file.” “Use the second version.” “Move it to tomorrow.”

Possible wrong referents:

Wrong referent failure is common when multiple candidate entities are salient.

7.1.2 Wrong intent#

Wrong intent failure occurs when the listener infers the wrong speech act or action.

Examples:

Wrong intent failure is especially likely with indirect speech acts, softened criticism, and implied requests.

7.1.3 Missing constraint#

Missing constraint failure occurs when an important limitation is not included, not inferred, or not preserved.

Examples:

“Summarize this.” Missing constraint: for executives, under 200 words, no technical detail.

“Find a hotel.” Missing constraint: budget, location, dates, accessibility, cancellation policy.

“Send the update.” Missing constraint: only after approval, internal recipients only, use final version.

Constraints may be unstated because the speaker assumes they are obvious from context.

Missing constraints can produce outputs or actions that are technically responsive but situationally wrong.

7.1.4 Stale context#

Stale context failure occurs when interpretation relies on context that is no longer current.

Examples:

“Use that version.” “Same as before.” “Move it to tomorrow.” “Apply the same settings.”

The relevant prior object or setting may have changed.

Stale context can result from:

• long conversations;
• topic shifts;
• changed selected object;
• revised priorities;
• updated files;
• cancelled meetings;
• replaced options;
• lost salience.

A stale reference can be more dangerous than an explicit ambiguity because it may appear resolved while actually pointing to an outdated state.

7.1.5 Hallucinated context#

Hallucinated context failure occurs when missing information is filled with an unsupported assumption.

Examples:

“Send it to Alex.”

Unsupported assumptions might include:

Some assumptions may be reasonable; others may be unsafe. The problem is not inference itself, but ungrounded inference treated as fact.

Hallucinated context is especially likely when the utterance is underspecified and the surrounding context is weak, stale, or conflicting.

7.1.6 Over-execution#

Over-execution occurs when the response or action goes beyond what the user intended.

Examples:

Over-execution often results from inferring a stronger intent than the speaker expressed.

It is particularly likely with high-level action phrases:

“Handle the onboarding.” “Take care of the client follow-up.” “Fix the spreadsheet.”

These may imply broad workflows, but the permissible scope may still be limited.

7.1.7 Under-execution#

Under-execution occurs when the response or action does less than the user intended.

Examples:

Under-execution often results from interpreting a high-level activity too narrowly.

The challenge is that natural language often names outcomes rather than every required sub-action. Correct interpretation requires recognizing conventional activity structure and user expectations.

7.2 Ambiguity Severity#

Not all ambiguity has the same consequence.

Some ambiguity is harmless. Some is recoverable. Some blocks meaningful action. Some creates unacceptable risk.

Severity depends on:

• consequence of error;
• reversibility;
• number of plausible interpretations;
• confidence in context;
• user expectation;
• affected parties;
• sensitivity of the object;
• whether action is external, destructive, or irreversible.

7.2.1 Harmless ambiguity#

Harmless ambiguity has little or no material consequence.

Example:

“Make this a bit cleaner.”

Possible interpretations:

• improve wording;
• improve formatting;
• reduce clutter;
• simplify structure.

If the context is low-risk and revision is reversible, the ambiguity may be acceptable.

Other examples:

Harmless ambiguity can often be resolved through best-effort interpretation.

7.2.2 Recoverable ambiguity#

Recoverable ambiguity may produce a wrong first attempt, but the error is easy to detect and repair.

Example:

“Make the intro shorter.”

If the wrong amount is cut, the user can say:

“A little less – keep the second example.”

Recoverable ambiguity is common in drafting, brainstorming, search, summarization, and low-risk planning.

Examples:

Recoverability depends on whether the action can be inspected, adjusted, or undone.

7.2.3 Execution-blocking ambiguity#

Execution-blocking ambiguity prevents responsible interpretation without additional information.

Examples:

“Delete the old one.” “Send it to Alex.” “Cancel my meeting.” “Approve the request.” “Move the deadline.”

These utterances may be clear in intent but ambiguous in referent, scope, authority, or consequence.

Execution-blocking ambiguity usually requires clarification when:

• multiple plausible referents exist;
• the action affects others;
• the action cannot be easily undone;
• the missing parameter determines the action;
• context does not strongly favor one interpretation.

Example clarification:

“Which meeting should be cancelled?”

7.2.4 Safety-critical ambiguity#

Safety-critical ambiguity occurs when a wrong interpretation could cause serious harm, rights violation, exposure, loss, legal risk, financial risk, security risk, or irreversible damage.

Examples:

“Share the file externally.” “Delete all inactive accounts.” “Approve the transfer.” “Publish the report.” “Remove the safety note.” “Give them access.”

Safety-critical ambiguity requires conservative interpretation.

Potential risk dimensions:

In safety-critical contexts, plausible inference is not always enough. Explicit confirmation or refusal may be required.

7.3 Confidence and Commitment#

Confidence is the estimated reliability of an interpretation. Commitment is the degree to which one proceeds as though the interpretation is correct.

The two should remain distinct.

A listener may have moderate confidence in an interpretation but low commitment because the consequences are high. Conversely, a listener may proceed with low certainty when the stakes are trivial and recovery is easy.

Example:

“Make this friendlier.”

Confidence may be moderate, and commitment may be acceptable because the action is reversible.

Example:

“Send this to the client.”

Confidence may be high, but commitment may still require confirmation if the object, recipient, or approval status is uncertain.

7.3.1 Infer silently#

Silent inference is appropriate when the missing information is low-risk, highly conventional, and easily recoverable.

Example:

“Make this more concise.”

Reasonable silent inferences:

Silent inference is less appropriate when the action is external, destructive, irreversible, permission-sensitive, or high-impact.

7.3.2 Infer with disclosure#

Inference with disclosure is appropriate when an interpretation is plausible but not certain, and the user benefits from seeing the assumption.

Example:

“I’ll treat ‘more polished’ as improving structure, tone, and wording while preserving the substance.”

This allows progress while making assumptions visible.

Useful disclosure patterns:

Disclosure is especially useful for vague quality terms, soft preferences, and underspecified revision requests.

7.3.3 Ask for clarification#

Clarification is appropriate when missing information materially changes the interpretation.

Examples:

“Which Alex do you mean?” “Should this go to the client or stay internal?” “Do you want a full rewrite or comments only?” “Should I preserve the examples?” “Which meeting should be moved?”

Clarification should target the smallest missing distinction that matters.

Good clarification:

“Which Alex should receive it: Alex Chen or Alex Rivera?”

Less useful clarification:

“Can you clarify?”

Clarification is especially important for ambiguous referents, destructive actions, external communication, permissions, and conflicting constraints.

7.3.4 Offer alternatives#

Offering alternatives is appropriate when multiple interpretations are plausible and the user may choose among them.

Example:

“I can interpret ‘make it lighter’ in three ways: lighter tone, shorter content, or less visual density. Which one do you mean?”

Alternative presentation is useful when:

• ambiguity is real;
• there are a small number of plausible interpretations;
• the user’s preference cannot be safely inferred;
• choosing incorrectly would waste effort or create risk.

Examples:

7.3.5 Require confirmation#

Confirmation is appropriate when an interpretation is plausible but the action is consequential.

Example:

“Confirm before I send this externally to the client.”

Confirmation is often required for:

Confirmation should restate the interpreted action clearly.

Example:

“Please confirm: send the final Q4 report to Alex Chen by email today.”

7.3.6 Decline execution#

Declining execution is appropriate when the requested action is unsafe, unauthorized, impossible, contradictory, or too ambiguous to proceed responsibly.

Examples:

“I cannot determine which file you mean.” “I cannot proceed without knowing whether external sharing is approved.” “These constraints conflict: the answer cannot be both under 50 words and include every section in detail.” “I do not have enough information to identify the correct recipient.”

Declining execution can still preserve helpfulness by explaining the blocking condition and identifying the minimum information needed.

Decline does not always mean refusal of the overall task. It may mean refusal to perform one unsafe or under-specified step.

7.4 Repair Protocols#

Repair protocols are interaction patterns for resolving uncertainty, error, conflict, or failed interpretation.

Repair can occur before action, during interpretation, after a mistaken output, or after the user revises the task.

Repair protocols include clarification questions, confirmation, correction acceptance, rollback, partial execution, and safe no-op.

7.4.1 Clarification question#

A clarification question asks for missing information needed to interpret the utterance.

Examples:

“Which file should I use?” “Do you mean Sarah Kim or Sarah Patel?” “Should the tone be formal or conversational?” “What counts as ’nearby’ here?” “Should this include external users?”

Clarification questions should be:

Poor clarification:

“Please clarify your request.”

Better clarification:

“Which meeting should I move: the 10 AM product sync or the 2 PM client review?”

7.4.2 Confirmation#

Confirmation verifies an inferred interpretation before treating it as final.

Examples:

“Confirming: you want the shorter version sent to Maria today.” “Just to confirm, should I exclude contractors?” “Before proceeding: do you want the public version or internal version?” “Please confirm that external sharing is approved.”

Confirmation differs from clarification. Clarification asks for missing information. Confirmation checks whether an inferred complete interpretation is correct.