# Limitations

## Purpose of this document

LLM Ambiguity Lab v1 is intentionally narrow. This document defines what the public tool does **not** claim to measure, infer, or guarantee.

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

## 1. It does not reveal hidden model reasoning

The Lab does not expose chain-of-thought, hidden reasoning traces, attention states, latent representations, or any private internal process of an LLM.

The v1 simulator does not need an LLM API to produce its output.

## 2. It does not measure actual model confidence

`context_resolution_score` is generated by an illustrative deterministic policy.

It is not:

- a calibrated probability;
- an LLM's internal confidence;
- a probability that the interpretation is correct;
- a probability that execution will succeed.

## 3. It cannot know the user's true internal intention

The Lab evaluates the information represented in the input and its simulation context. It cannot inspect a user's private mental state.

`Resolve` means *sufficiently specified under the simulation policy*, not *proven to match the user's objectively true intention*.

## 4. The ambiguity taxonomy is not exhaustive

v1 focuses on:

- lexical ambiguity;
- referential ambiguity;
- intent / task ambiguity.

Real interactions can also involve pragmatic ambiguity, domain ambiguity, conflicting instructions, uncertainty about authority, temporal ambiguity, multimodal ambiguity, and many other forms of underspecification.

## 5. The free-text heuristic is deliberately limited

The free-text analyzer uses a small vocabulary and simple deterministic extraction rules.

It can:

- recognize a limited set of task verbs;
- recognize a limited set of output-format terms;
- detect several referential and continuation signals;
- extract simple explicit target phrases.

It cannot provide general semantic parsing or robust discourse resolution.

A natural-language instruction that is obvious to a human may therefore be classified conservatively or incorrectly by the heuristic.

## 6. Preset results are reference scenarios

The six presets are predefined teaching cases. They are not benchmark measurements and do not establish empirical model performance.

## 7. The thresholds are illustrative

The score bands `0.90`, `0.70`, and `0.40` are architectural demonstration thresholds.

They are not claimed to be universally optimal for:

- production agents;
- high-risk domains;
- particular LLM families;
- particular tool-use systems;
- particular user populations.

A real deployment should define and validate its own policy based on task risk, reversibility, available context, and evaluation data.

## 8. It does not guarantee hallucination prevention

Clarification and explicit assumptions may reduce some classes of execution error, but the Lab does not claim that DCRL eliminates hallucinations or guarantees factual accuracy.

## 9. It does not guarantee agent safety

A correctly resolved task can still be unsafe, unauthorized, undesirable, or incorrect.

Task resolution is only one layer of an agent system. Real systems may additionally require authorization checks, policy enforcement, risk assessment, data validation, tool permissions, human review, and rollback mechanisms.

## 10. It is not a production classifier

The v1 heuristic is an educational implementation. It should not be copied into a production system and treated as validated intent classification without separate testing.

## 11. It is not psychological assessment

The Lab does not infer:

- hidden motives;
- personality;
- mental state;
- psychological diagnosis;
- sincerity or deception.

It analyzes operational completeness of instructions under its own rules.

## 12. It does not evaluate premise truth or premise alignment

v1 asks:

> **What are we doing?**

It does not ask whether the assumptions carried into that task are true, justified, compatible, or aligned.

That broader premise-alignment layer is outside v1 and is reserved for future work associated with SANA OS.

## 13. Language coverage

English is the canonical language for v1. The heuristic vocabulary is primarily English and should not be assumed to behave equivalently in other languages.

## 14. Browser and hosting constraints

The v1 simulator logic executes client-side in JavaScript and does not require an LLM API. Browser behavior, accessibility support, or availability may vary by environment.

The public host may maintain ordinary infrastructure logs independently of the simulator code.

## 15. No empirical performance claim is implied

The public v1 release intentionally contains no unsupported numerical claim that DCRL improves task completion by a specific percentage.

Any future empirical result should be accompanied by a traceable source or reproducible evaluation method.