FD Scenarios — Interactive Demo

Interactive demo for the T × I × R ontology — each cell shows one canonical scenario as a dual-channel waveform synced to its audio. Click ▶ to play; click any waveform to seek.

📐 Three-axis ontology

Each conversational micro-event = (Temporal · Intent · Response) triplet. Failure modes localise to specific cells in this 3-axis space, making them diagnosable rather than vague.

T Temporal — when

T1	Sequential — 200-1500 ms inter-turn gap
T2	Latched — gap < 200 ms
T3	Overlap-onset — one party begins while the other is talking
T4	Concurrent — both ≥ 1 s sustained overlap
T5	Silence — both quiet ≥ 500 ms

I Intent — what for

I1	Information — primary content
I2	Backchannel — "uh-huh", "嗯", no floor claim
I3	Repair — self-correction or clarification
I4	Floor-claim — barge-in to take turn
I5	Floor-yield — release turn to other party
I6	Hesitation — mid-utterance pause, not done
I7	Third-party — non-addressed speech / noise

R Response — system action

R1	Continue — keep talking
R2	Stop & listen — yield within ~200 ms
R3	Wait — stay silent, user not finished
R4	Acknowledge — emit backchannel only
R5	Ignore / filter — drop non-addressed audio
R6	Initiate — start a new turn

Why three axes, not flat categories? The same (Intent, Response) pair has different correctness depending on Temporal context. Example: (I2, R1) = "user backchannel, assistant continues" is correct during T3 (overlap-onset) but wrong during T1 (sequential turn, where the same "uh-huh" might be the user yielding/closing).

🎧 Six canonical scenarios