Overview

• It's important to mention that this was mostly for my own curiosity and as part of a college project. I plan to keep working on it and addressing its current limitations.
• Two deterministic probabilistic agents exchange messages.
• A statistical analyzer examines the resulting dialogue for evidence that the agents follow different generative models.
• The system uses no language models.
• All responses are produced from fixed templates modulated by internal state variables and seeded random choice.

testvideo.mp4

• It simultaneously measures multiple distributive and sequential signals and maintains an episodic memory of previous analyses, allowing the decision threshold to adapt to previously observed confidence patterns.

How distinction is computed

• Before each analysis the current state (message counts, formality values, total length, window size) is hashed.

• The hash is deterministically embedded into a fixed-dimensional vector. The nearest past episodes are retrieved with cosine similarity weighted by an exponential temporal decay.

• Their labels and confidences are summarized into a short natural-language context string that is later stored with the new result.

• The analyzer then computes seven signals on the (optionally windowed) histories:

• Stylometric divergence of six surface features (word count, average word length, punctuation density, capitalization ratio, presence of question marks and exclamation marks).

  - absolute difference in lexical entropy
  - Kolmogorov Smirnov scores on the same feature distributions
  - Jensen-Shannon divergence of token distributions
  - 2 gram overlap (reported as 1 Jaccard)
  - turn taking length regularity
  - A small prior boost derived from the retrieved episodes
  

• These are combined with fixed weights into a composite score.

• An adaptive threshold is calculated from the confidence distribution of previously stored indistinguishable episodes (or a simple shift of the global mean if too few such episodes exist).

• The final label is distinct when the composite meets or exceeds the threshold; confidence is the normalized distance from the threshold.

• All embeddings and random draws are fully deterministic given the seeds and the hash of the analysis state.