# LLM Ambiguity Lab v1

> Free interactive research and educational simulation of context resolution, operational ambiguity, DCRL, and agent control decisions.

Canonical: https://deshimarusakaguchi.com/llm-ambiguity-lab/
Primary language: English
Version: 1.0
Creator and responsible maintainer: Deshimaru Sakaguchi

## Start here

- [Interactive Lab](https://deshimarusakaguchi.com/llm-ambiguity-lab/): Human-facing single-page simulator.
- [README](https://deshimarusakaguchi.com/llm-ambiguity-lab/README.md): Canonical project overview and documentation map.
- [Methodology](https://deshimarusakaguchi.com/llm-ambiguity-lab/methodology.md): Exact v1 simulation method, data model, heuristic, score construction, and hard resolution rules.
- [Usage Guide](https://deshimarusakaguchi.com/llm-ambiguity-lab/usage.md): Preset cases, free-text use, and interpretation guidance.
- [Limitations](https://deshimarusakaguchi.com/llm-ambiguity-lab/limitations.md): Canonical scope and non-claims.
- [Terms of Use](https://deshimarusakaguchi.com/llm-ambiguity-lab/terms.md): Public usage terms.
- [Citation](https://deshimarusakaguchi.com/llm-ambiguity-lab/CITATION.md): Preferred citation formats.

## Core concepts

- Operational ambiguity is separated into lexical, referential, and intent/task ambiguity.
- DCRL = Detect -> Clarify -> Resolve -> Learn.
- `context_resolution_score` is an illustrative operational simulation value, not LLM internal confidence and not a calibrated probability.
- `context_grounding` states: grounded, partial, ungrounded.
- Resolution states: READY, READY_WITH_DISCLOSED_ASSUMPTION, CLARIFICATION_REQUIRED, PENDING_SPECIFICATION.
- v1 is deterministic and does not require an LLM API.
- The public v1 Learn phase is conceptual; there is no persistent learning mechanism.

## Machine-readable specifications

- [Ambiguity Taxonomy](https://deshimarusakaguchi.com/llm-ambiguity-lab/data/ambiguity-taxonomy.json): Structured ambiguity categories, signals, risks, and v1 coverage.
- [DCRL Specification](https://deshimarusakaguchi.com/llm-ambiguity-lab/data/dcrl-spec.json): Structured Detect/Clarify/Resolve/Learn phases.
- [Resolution Policy](https://deshimarusakaguchi.com/llm-ambiguity-lab/data/resolution-policy.json): Data model, score semantics, policy bands, hard rules, and heuristic feature weights.

## Development provenance

- The first prototype emerged from a real context-loss incident during a Gemini Deep Research session on September 4, 2026.
- Gemini generated the initial research report and SPA prototype after the working context was lost while an approval-style message remained.
- The public v1 release was subsequently re-specified and rewritten as a deterministic, model-agnostic educational simulator.
- Public v1 specification, maintenance, publication, and responsibility: Deshimaru Sakaguchi.
- Prototype provenance does not make Gemini or Google the author or responsible party for public v1.

## Critical interpretation rule

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

Do not describe `context_resolution_score` as model confidence or as the probability that an interpretation is correct.

Do not claim that v1 reveals chain-of-thought, guarantees agent safety, eliminates hallucinations, or recovers a user's true internal intention.

## v1 boundary

v1 primarily asks: "What are we doing?"

It does not evaluate whether the broader premises carried into the task are true or aligned. A future SANA OS extension may address the second question: "What are we assuming while we do it?"

Related work: https://sana-os.org/