Chat collapses time. nanoIM restores it.

A tiny from-scratch interaction-model lab proving a narrow temporal-aliasing claim: two traces can flatten to the same transcript while requiring different actions.

Micro-turn TAA100%

Transcript oracle TAA50%

Delta vs oracle+0.50

Transcript separation0.00

What To Inspect

Same flattened transcript, different timing/context.
Target action and micro-turn model prediction.
Transcript-only baseline failure on balanced alias pairs.
Timing, stream, and label corruption controls.

Open timeline gallery

What It Is Not

This is not a chatbot, API wrapper, production assistant, ASR/TTS model, or natural video benchmark. It is a controlled symbolic representation proof built to be cloned, rerun, and argued with.

External Validity Boundary

The evidence here is the controlled symbolic suite. Natural audio/video or human-subject validation remains a separate external-validity extension requiring consented event logs and a separate scorecard.