AL LLM Ambiguity Lab v1 Interactive context-resolution architecture simulator
Interactive Research & Educational Simulation

What should an AI agent do when it is not sure what you mean?

Natural conversation often relies on shared context. Instructions such as “Continue,” “Do that,” or “Please proceed” may be perfectly clear to a human participant while remaining operationally incomplete for an AI agent.

LLM Ambiguity Lab explores how an agent can detect missing context, identify unresolved parameters, choose whether to proceed or clarify, and make its assumptions visible before execution.

Resolve when sufficiently grounded. Clarify when materially ambiguous. Disclose assumptions when safe to proceed. Stop when the task itself cannot be identified.
Simulation notice: This Lab uses illustrative, deterministic rules to demonstrate an agent-control architecture. The displayed score is not an LLM's internal confidence and does not reveal hidden model reasoning.
Try the Simulator How the Simulation Works
Interactive Demo

1. Prompt Ambiguity Simulator

Enter an instruction or select a preset example. The simulator will estimate how well the task can be resolved from the information available and show the corresponding control action.

Try a preset
Simulation Output PENDING_SPECIFICATION
Context Resolution Score
0.18
Resolution State
PENDING_SPECIFICATION
Target Task
null
Target Object
null
Context Grounding
ungrounded
Ambiguity Type
mixed
Missing Parameters
Plausible Interpretations
Recommended System Action

Do not execute a task yet. Request specification.

I don't have enough context to identify what you want me to continue. What task should I proceed with?

Taxonomy

2. Three Types of Operational Ambiguity

Not all ambiguous instructions fail in the same way. The Lab separates three common sources of uncertainty that can affect agent execution.

Lexical

Lexical Ambiguity

A word or expression permits multiple plausible meanings, and the available context does not sufficiently constrain which meaning is intended.

Operational signal
The same instruction supports materially different interpretations of a word or phrase.
Agent risk
The agent may select the wrong operation, tool, parameter, or interpretation.
Referential

Referential Ambiguity

An instruction depends on a referent such as “it,” “that,” or an omitted object that cannot be uniquely recovered from the available context.

Operational signal
The requested action may be identifiable while the object of that action remains unresolved.
Agent risk
The agent may operate on the wrong document, topic, task, or prior message.
Intent / Task

Intent / Task Ambiguity

The subject may be identifiable while the requested operation, goal, constraints, or expected output remain insufficiently specified.

Operational signal
Multiple materially different actions could satisfy the wording.
Agent risk
The agent may produce a valid response that solves the wrong problem.
Control Loop

3. Detect → Clarify → Resolve → Learn

DCRL is a control loop for deciding what an agent should do before executing an instruction whose operational meaning may be incomplete.

Policy Simulator

4. Context Resolution Policy

Move the score to see how the simulated control policy changes. The thresholds are illustrative policy boundaries, not universal constants and not calibrated probabilities.

Current Context Resolution Score
0.25
0.000.400.700.901.00
Engineering Notes

5. Implementation Guidance

Ambiguity handling is a control-design problem, not merely a prompting problem. The following patterns illustrate common implementation mistakes.

Anti-patternFailurePreferred pattern
Treating grammatical completeness as task completenessA fluent instruction may still lack a usable target or referent.Validate task, object, context grounding, and required parameters separately.
Treating context grounding as BooleanPartially resolved instructions are collapsed into either “known” or “unknown.”Distinguish grounded, partial, and ungrounded.
Treating the score as probabilityAn operational policy value may be misread as calibrated model confidence.Label it explicitly as context_resolution_score and document how it is produced.
Asking broad clarification questionsThe user must reconstruct the whole task instead of resolving one missing parameter.Ask bounded questions that distinguish among the most plausible interpretations.
Executing through critical ambiguityThe agent may act on the wrong target or perform the wrong task.Enter CLARIFICATION_REQUIRED or PENDING_SPECIFICATION before execution.
target_tasktarget_objectcontext_groundingambiguity_typemissing_parametersoutput_formatcontext_resolution_scoreresolution_stateassumption_disclosure
Transparency

6. Method & Transparency

LLM Ambiguity Lab v1 is a deterministic educational simulator. It demonstrates how an agent-control policy can represent missing context and choose among execution, assumption disclosure, clarification, and specification requests.

How the Demo Works

Preset examples return predefined reference scenarios. Free-text input is evaluated using simple, documented heuristics such as explicit task verbs, identifiable targets, referential expressions, requested formats, and vague continuation language.

The v1 simulator does not require an LLM API. Its purpose is reproducibility and architectural explanation rather than natural-language understanding performance.

What the Score Means

The score is an operational simulation value. It is not a probability and is not extracted from an LLM's hidden state.

Hard semantic rules can override the raw score. For example, if the target task cannot be identified, the simulator will not allow the instruction to become READY merely because other features increase the numeric score.

Research claims: No numerical performance improvement is presented as an empirical finding in v1 unless a traceable source or reproducible experiment is provided.
Origin note: The first prototype grew out of a real context-loss incident during a Gemini Deep Research session on September 4, 2026. An approval-style message remained after the working context was lost, and the resulting research unexpectedly focused on how agents should handle context-deficient instructions. The public v1 was subsequently rewritten as a deterministic, model-agnostic educational simulator. Gemini is referenced here only to document development provenance.
Boundaries

7. What This Lab Does Not Do

The Lab is designed to make control decisions around ambiguity easier to inspect. It does not make uncertainty disappear.

The Lab visualizes a control architecture, not a model's mind.

Where v1 Stops

LLM Ambiguity Lab v1 asks whether an agent can identify what the user is asking it to do. It does not evaluate the broader premises carried into that task.

A future SANA OS extension may explore that second layer: not only “What are we doing?” but also “What are we assuming while we do it?”

Related work: SANA OS