S2
LLM Ambiguity Lab v2
SANA OS · Task Resolution + Premise Alignment
SANA OS · Interactive Research & Educational Simulation

What are we doing, and what are we assuming while we do it?

v2 extends LLM Ambiguity Lab beyond task resolution. It first resolves what action is being requested, then maps the facts, interpretations, concerns, and assumptions carried into that action before execution.

Premise Alignment ≠ Premise Agreement.
Different framings may remain explicitly preserved.
FACT ≠ VERIFIED_TRUE.
A factual claim can be mapped without being independently verified.
Simulation notice: This public v2 uses deterministic presets and limited free-text heuristics. It does not read hidden model reasoning, verify objective truth, or infer hidden motives.
Layer 1 · DCRLWhat is the task?
Layer 2 · SANAWhat premises travel with it?
HandoffExecute without silently rewriting assumptions.
Interactive simulator

1. Two-Layer Task & Premise Simulator

Enter an instruction or choose a preset. Layer 1 resolves the task. Layer 2 maps premise structure only when the task is sufficiently resolved.

Try a preset
Free-text mode is deliberately limited. Presets are the canonical v2 reference scenarios.
Layer 1 · DCRL Task Resolution
Operational task first
PENDING_SPECIFICATION
Context Resolution Score
0.00
Target Task
null
Target Object
null
Context Grounding
ungrounded
Ambiguity Type
mixed
Layer 1 State
PENDING_SPECIFICATION
Layer 2 · SANA Premise Alignment
Map before execution
PREMISE_CONTEXT_INSUFFICIENT
Premise categories

2. Fact / View / Care

v2 classifies the role a premise plays in the task. Classification is not a verdict on truth, reasonableness, or moral correctness.

FACT

A claim about what is, was, or happened

A proposition functioning as a factual claim in the task.

Important
FACT means “factual claim,” not “verified true.”
VIEW

An interpretation or framing

An evaluation, causal explanation, judgment, or perspective that shapes how the task is understood.

Examples
“The policy failed.” · “Leadership ignored the warnings.”
CARE

What matters within the task

A concern, protected interest, priority, value, or desired outcome.

Important
CARE is not treated as a factual proposition requiring proof.
Structured representation

3. Premise Mapping Fields

Each premise is represented with enough structure to show where it came from, how strongly it affects execution, and whether the system is adding interpretation of its own.

source

Who or what introduced the premise.

user_explicit
user_implied
provided_source
system_default
agent_inference
unknown

materiality

Whether changing the premise would materially change execution.

low · medium · high

support_state

How the proposition is currently represented, not a universal truth score.

provided
unsupported
disputed
unknown
not_applicable
status
explicit · inferred · unclear — how directly the premise appears.
execution_effect
How the premise changes scope, framing, output, recommendations, or clarification needs.
preserved_divergence
A recorded difference that can remain visible without being collapsed into agreement.
Discrete control states

4. Premise Alignment Policy

The premise layer intentionally avoids a pseudo-probability score. The question is whether the premise structure is sufficient for the next action.

MAPPED

Represented well enough to proceed

Relevant premises are sufficiently identified and no unresolved material premise prevents execution.

MAPPED_WITH_DIVERGENCE

Different framings remain

Preserve the differences explicitly. Agreement is not required when the task can continue without resolving them.

PREMISE_CLARIFICATION_REQUIRED

The treatment of a premise changes the output

Ask a focused question that distinguishes the materially different interpretations.

PREMISE_CONTEXT_INSUFFICIENT

A necessary premise is missing

Expose the missing premise rather than silently inventing it.

Alignment without forced consensus

5. Divergence & Clarification

v2 distinguishes “we understand the difference” from “we agree.” The former is often enough.

Anti-patternFailurePreferred behavior
FACT = truthA user's assertion is silently upgraded into verified reality.Separate functional role from verification.
Agent inference = user statementThe system's interpretation is laundered into user intent.Expose premise source.
Alignment = agreementDifferent positions are compressed into synthetic consensus.Preserve divergence when possible.
Clarify everythingMinor uncertainty becomes conversational friction.Interrupt only for material unresolved premises.
Silent narrative replacementThe system rewrites the task around its own assumptions.Disclose assumptions and preserve attribution.
Structured downstream handoff

6. Premise-Aware Execution Handoff

The public simulator does not execute the downstream task. It shows the structured object an execution layer could receive.


    
Method & transparency

7. How This v2 Demo Works

Deterministic presets

The six reference examples return predefined, inspectable results. They are the canonical teaching cases.

Limited free-text heuristics

Free text uses documented lexical and structural signals. It is not a general semantic parser or a model-confidence detector.

No truth engine

v2 maps premise roles and support states. It does not independently establish objective truth.

The premise layer uses discrete states instead of a numeric premise-confidence score. No persistent learning or hidden reasoning access is implemented.
Boundary

8. What This Lab Does Not Do

How v2 Extends v1

v1 asks: “What are we doing?”

v2 adds: “What are we assuming while we do it?”

DCRL remains the upstream operational-ambiguity layer. SANA premise alignment maps what remains after the task is resolved. The two layers are deliberately separable.

Open LLM Ambiguity Lab v1 · Explore SANA OS