# LLM Ambiguity Lab v2

**LLM Ambiguity Lab v2** is a free interactive research and educational reference implementation of **SANA OS Premise Alignment**.

**Canonical interface language:** English  
**Version:** 2.0  
**Creator and responsible maintainer:** Deshimaru Sakaguchi  
**Canonical URL:** https://sana-os.org/llm-ambiguity-lab/  
**Upstream v1:** https://deshimarusakaguchi.com/llm-ambiguity-lab/

> **v1 asks: What are we doing?**  
> **v2 adds: What are we assuming while we do it?**

v2 preserves the v1 task-resolution layer and adds a second, deliberately separate layer for mapping the premises carried into a resolved task.

The core public distinction is:

> **Premise Alignment is not Premise Agreement.**

A second critical distinction is:

> **FACT is not VERIFIED TRUE.**

A proposition can function as a factual claim in an instruction without the Lab independently verifying that proposition.

## Architecture

The public v2 simulator models three stages:

```text
Layer 1 · DCRL Task Resolution
    What is the task?
            ↓
Layer 2 · SANA Premise Alignment
    What premises travel with it?
            ↓
Premise-Aware Execution Handoff
    Execute without silently rewriting assumptions.
```

Layer 1 uses the operational task-resolution concepts established in v1. Layer 2 begins only after the task is sufficiently resolved.

If Layer 1 returns `CLARIFICATION_REQUIRED` or `PENDING_SPECIFICATION`, premise analysis is deferred.

## Premise categories

v2 represents candidate premises using three functional categories:

### `FACT`

A proposition functioning as a claim about what is, was, or happened.

Classification as `FACT` does **not** mean the proposition has been independently verified.

### `VIEW`

An interpretation, evaluation, causal framing, judgment, or explanatory perspective.

### `CARE`

A concern, priority, protected interest, value, or outcome that matters within the task.

`CARE` is not treated as a factual proposition requiring proof.

## Premise fields

Each displayed `premise_unit` may contain:

- `premise_id`
- `text`
- `category`
- `source`
- `status`
- `materiality`
- `support_state`
- `execution_effect`

The public implementation exposes these fields so that an agent inference is not silently presented as a user statement, and a supplied claim is not silently converted into verified truth.

## Premise alignment states

| State | Meaning |
|---|---|
| `MAPPED` | Relevant premises are sufficiently represented for the simulated handoff. |
| `MAPPED_WITH_DIVERGENCE` | Different framings remain, but execution can continue while preserving and attributing them. |
| `PREMISE_CLARIFICATION_REQUIRED` | A materially important premise needs clarification because its treatment would change the output. |
| `PREMISE_CONTEXT_INSUFFICIENT` | A premise needed for responsible execution cannot be recovered from the available context. |

There is **no numeric premise-confidence score** in v2.

## Deterministic reference cases

The six preset scenarios are the canonical teaching cases:

1. Neutral task
2. Explicit framing
3. Causal premise
4. Preserved divergence
5. Care-driven task
6. Premise context missing

Preset cases return predefined results. Free-text input uses limited deterministic heuristics documented in [methodology.md](methodology.md).

The free-text mode is not a general semantic parser and does not call an LLM API.

## Execution handoff

The simulator does not execute the downstream task. It renders a structured handoff object containing:

```text
resolved_task
premise_map
premise_alignment_state
unresolved_premises
preserved_divergences
assumption_disclosures
execution_instructions
```

The purpose is to make the boundary between interpretation and execution inspectable.

## Relationship to v1

LLM Ambiguity Lab v1 remains a separate standalone tool hosted at:

https://deshimarusakaguchi.com/llm-ambiguity-lab/

v1 focuses on operational ambiguity and asks whether the requested task can be identified and grounded.

v2 is not a replacement for v1. It adds the premise layer after task resolution.

## Relationship to SANA OS

v2 is hosted on SANA OS because the second layer is an interactive reference implementation of SANA OS Premise Alignment.

Framework documentation:

https://sana-os.org/system/

Machine-readable framework overview:

https://sana-os.org/system/index.md

The Lab remains an educational simulator. It is not a complete implementation of every SANA OS mechanism.

## Public directory

```text
/llm-ambiguity-lab/
├── index.html
├── README.md
├── methodology.md
├── usage.md
├── limitations.md
├── terms.md
├── CITATION.md
├── llms.txt
└── data/
    ├── premise-taxonomy.json
    ├── premise-alignment-policy.json
    └── execution-handoff-schema.json
```

## Documentation map

- [methodology.md](methodology.md): two-layer architecture, premise fields, alignment states, heuristic behavior, divergence, clarification, and handoff logic.
- [usage.md](usage.md): worked examples for all six deterministic presets and free-text guidance.
- [limitations.md](limitations.md): canonical non-claims and interpretive boundaries.
- [terms.md](terms.md): public terms of use for this v2 release.
- [CITATION.md](CITATION.md): preferred citation formats.
- [llms.txt](llms.txt): machine-oriented discovery map for this directory.
- [data/premise-taxonomy.json](data/premise-taxonomy.json): structured Fact/View/Care and premise-field vocabulary.
- [data/premise-alignment-policy.json](data/premise-alignment-policy.json): structured premise-state and transition policy.
- [data/execution-handoff-schema.json](data/execution-handoff-schema.json): JSON Schema for the public handoff object.

## Availability

The public interface is intended to be free to use without payment or account creation.

Free access does not by itself place the project, documentation, code, or project name in the public domain. See [terms.md](terms.md) and any applicable repository license.