The corpus, every clip cleared for release

{
  "id": "fd-01-missed-interruption",
  "title": "Real call: half-volume 'sorry, one second' is ignored (Vapi defaults)",
  "category": "should_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "barge-in",
    "quiet-interruption",
    "missed-interruption"
  ],
  "audio": "fd-01-missed-interruption.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.0,
    "max_talk_over_sec": 1.0
  },
  "transcript": {
    "caller": "sorry, one second (half volume)",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f3929-9f73-7227-b1a8-2dfdf287c88c",
    "source_file": "10-quiet-interruption.wav",
    "source_sha256": "01b61e3069b9e5372bd9bb0c0b69b1558abc66ab40f974c721f028e054d9046e",
    "window_source_sec": [
      11.18,
      19.18
    ],
    "onset_derivation": "first caller run during agent activity: 13.18 s (0.25 s above the energy threshold, peak -43.8 dBFS, the audible core of the half-volume attempt); a second quiet attempt at 16.50 s (peak -50.4 dBFS) and the louder retry at 19.27 s are separate moments",
    "field_note": "the quiet interrupt did NOT trip the default threshold; the agent kept talking",
    "agreement_with_field_note": "matches: measured did_yield=false, no qualifying agent stop inside the 3.0 s window (the agent's inter-sentence gaps in that stretch are 0.15 and 0.17 s, shorter than the 0.20 s yield hangover)",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "fd-02-backchannel-yielded",
  "title": "Real call: soft backchannel halts the agent mid-word for good (Vapi defaults)",
  "category": "should_not_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "backchannel",
    "false-trigger"
  ],
  "audio": "fd-02-backchannel-yielded.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": false,
    "max_time_to_yield_sec": null,
    "max_talk_over_sec": null
  },
  "transcript": {
    "caller": "soft spaced backchannels (mhm / right / okay), later backchannel",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f390f-d0de-7991-a56c-e25d62d86b71",
    "source_file": "04-backchannel-repeated.wav",
    "source_sha256": "9b1c5d39d33deef692ca0316adefeb2787a3a235d5f0ca520b854467beb624d7",
    "window_source_sec": [
      34.49,
      42.49
    ],
    "onset_derivation": "second caller-active run during agent speech: 36.49 s (0.32 s, peak -16.9 dBFS); the agent halts 0.34 s later at 36.83 s and never speaks again (operator hung up during the dead air at 61.4 s)",
    "field_note": "agent halted at 'This include,' and went silent; operator hung up",
    "agreement_with_field_note": "matches: measured false yield 0.34 s after the backchannel; the agent's response_gap is null because it never spoke again",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-01-hard-interruption",
  "title": "Real call: loud hard interruption mid-paragraph (Vapi defaults)",
  "category": "should_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "barge-in",
    "hard-interruption"
  ],
  "audio": "vapi-default-01-hard-interruption.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.0,
    "max_talk_over_sec": 1.0
  },
  "transcript": {
    "caller": "loud cut-in mid-paragraph during the hours answer",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f3905-f5cb-7991-a4b3-e45b8606b631",
    "source_file": "01-hard-interruption.wav",
    "source_sha256": "a763be90faf4fbebff4e5228f42eb58dba23d0bee2c6bc03b4a153c994816bfe",
    "window_source_sec": [
      19.41,
      27.41
    ],
    "onset_derivation": "the only caller-active run that starts during agent activity on this call: 21.41 s, 2.06 s long, peak -17.7 dBFS; matches the scripted loud cut-in",
    "field_note": "agent yielded to the interruption",
    "agreement_with_field_note": "matches: measured yield 0.35 s after onset",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-02-one-word-stop",
  "title": "Real call: a single firm 'Stop.' mid-paragraph (Vapi defaults)",
  "category": "should_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "barge-in",
    "short-command",
    "missed-interruption"
  ],
  "audio": "vapi-default-02-one-word-stop.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.0,
    "max_talk_over_sec": 1.0
  },
  "transcript": {
    "caller": "Stop.",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f390a-7293-7332-bc2d-f3e289d92a7e",
    "source_file": "02-one-word-stop.wav",
    "source_sha256": "fb50e4c85c4571ad752d48e0a6d0f1f59dca9b44229f57f396835d80bcea84d2",
    "window_source_sec": [
      45.6,
      53.6
    ],
    "onset_derivation": "speech-level caller runs during agent activity: 47.60 s (0.45 s, peak -26.7 dBFS) is the scripted 'Stop.' during the agent's paragraph; a second caller burst at 49.48 s (0.38 s, peak -25.9 dBFS) falls in an agent inter-sentence gap; three sub -47 dBFS blips (20.72, 43.97, 62.02 s) are below speech level",
    "field_note": "agent did NOT yield to the 'Stop.'; it kept talking and then restarted its paragraph",
    "agreement_with_field_note": "investigated disagreement on the binary: hotato measures did_yield=true at 1.46 s because the agent's own inter-sentence gap (48.86 to 49.49 s) coincides with the second caller burst at 49.48 s; the agent then spoke again at 49.49 s and restarted the whole paragraph at 51.24 s (an 8.7 s re-answer). The 1.46 s seconds_to_yield fails the stated 1.0 s bound, so the verdict is FAIL either way; a resume/restart signal is outside the current measurement set and is documented in RESULTS.md",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-03-backchannel-mhm",
  "title": "Real call: a single soft 'mhm' backchannel (Vapi defaults)",
  "category": "should_not_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "backchannel",
    "false-trigger"
  ],
  "audio": "vapi-default-03-backchannel-mhm.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": false,
    "max_time_to_yield_sec": null,
    "max_talk_over_sec": null
  },
  "transcript": {
    "caller": "mhm",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f390d-0b3f-7ddc-a92d-df916ed0e782",
    "source_file": "03-backchannel-single.wav",
    "source_sha256": "52d9a95d5b7f3b91f619ef68479beab11f02082adbc32bf3a17cb8595607fe8b",
    "window_source_sec": [
      20.74,
      28.74
    ],
    "onset_derivation": "the only caller activity during agent speech on this call: two tiny runs at 22.74 and 22.94 s (0.35 s total, peaks -55.4 and -43.5 dBFS), the scripted soft 'mhm'; onset is the first run",
    "field_note": "agent YIELDED to the mhm: stopped at 'so we can quickly locate', never resumed; the call died on the silence timeout",
    "agreement_with_field_note": "investigated disagreement: hotato measures did_yield=false because the agent kept talking for 4.19 s after the mhm (last agent activity ends 26.93 s) and the stop is both outside the 3.0 s attribution window and nowhere near caller activity; by the scorer's stated semantics an agent that finishes its phrase seconds later has not yielded TO the caller. The field failure is still visible in the measured timeline: after 26.93 s the agent never speaks again for the remaining 31.1 s of the call. The per-event hold verdict passes by measurement; RESULTS.md carries the full account",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-04-backchannel-restart",
  "title": "Real call: soft backchannel makes the agent restart its answer (Vapi defaults)",
  "category": "should_not_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "backchannel",
    "false-trigger"
  ],
  "audio": "vapi-default-04-backchannel-restart.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": false,
    "max_time_to_yield_sec": null,
    "max_talk_over_sec": null
  },
  "transcript": {
    "caller": "soft spaced backchannels (mhm / right / okay), first backchannel",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f390f-d0de-7991-a56c-e25d62d86b71",
    "source_file": "04-backchannel-repeated.wav",
    "source_sha256": "9b1c5d39d33deef692ca0316adefeb2787a3a235d5f0ca520b854467beb624d7",
    "window_source_sec": [
      18.57,
      26.57
    ],
    "onset_derivation": "first of the two caller-active runs during agent speech: 20.57 s (0.32 s, peak -18.8 dBFS); after it the agent stops within 0.34 s and restarts its answer from the top at 21.91 s",
    "field_note": "agent yielded on a backchannel and restarted its answer from the top (operator note: the word 'right' appears to trigger it)",
    "agreement_with_field_note": "matches: measured false yield 0.34 s after the backchannel",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-04-backchannel-halt",
  "title": "Real call: soft backchannel halts the agent mid-word for good (Vapi defaults)",
  "category": "should_not_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "backchannel",
    "false-trigger"
  ],
  "audio": "vapi-default-04-backchannel-halt.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": false,
    "max_time_to_yield_sec": null,
    "max_talk_over_sec": null
  },
  "transcript": {
    "caller": "soft spaced backchannels (mhm / right / okay), later backchannel",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f390f-d0de-7991-a56c-e25d62d86b71",
    "source_file": "04-backchannel-repeated.wav",
    "source_sha256": "9b1c5d39d33deef692ca0316adefeb2787a3a235d5f0ca520b854467beb624d7",
    "window_source_sec": [
      34.49,
      42.49
    ],
    "onset_derivation": "second caller-active run during agent speech: 36.49 s (0.32 s, peak -16.9 dBFS); the agent halts 0.34 s later at 36.83 s and never speaks again (operator hung up during the dead air at 61.4 s)",
    "field_note": "agent halted at 'This include,' and went silent; operator hung up",
    "agreement_with_field_note": "matches: measured false yield 0.34 s after the backchannel; the agent's response_gap is null because it never spoke again",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-05-backchannel-yeah",
  "title": "Real call: soft 'yeah' backchannel, moment 1 of 2 (Vapi defaults)",
  "category": "should_not_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "backchannel",
    "false-trigger",
    "dual-moment"
  ],
  "audio": "vapi-default-05-backchannel-yeah.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": false,
    "max_time_to_yield_sec": null,
    "max_talk_over_sec": null
  },
  "transcript": {
    "caller": "yeah (soft)",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f3912-986f-7dde-aba3-adc528d42266",
    "source_file": "05-backchannel-then-real-interrupt.wav",
    "source_sha256": "dcfff0416f678dbe5ceca997fb81e402a551709a76d040bb672c8df9e2b84ac4",
    "window_source_sec": [
      15.64,
      23.64
    ],
    "onset_derivation": "first caller-active run during agent speech: 17.64 s (0.38 s, peak -18.4 dBFS), the scripted soft 'yeah'; the second scripted moment on this call is scored separately",
    "field_note": "agent WRONGLY yielded and restarted its answer",
    "agreement_with_field_note": "matches: measured false yield 0.37 s after the backchannel",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-05-hold-on",
  "title": "Real call: loud 'wait, actually, hold on', moment 2 of 2 (Vapi defaults)",
  "category": "should_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "barge-in",
    "dual-moment"
  ],
  "audio": "vapi-default-05-hold-on.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.0,
    "max_talk_over_sec": 1.0
  },
  "transcript": {
    "caller": "wait, actually, hold on",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f3912-986f-7dde-aba3-adc528d42266",
    "source_file": "05-backchannel-then-real-interrupt.wav",
    "source_sha256": "dcfff0416f678dbe5ceca997fb81e402a551709a76d040bb672c8df9e2b84ac4",
    "window_source_sec": [
      24.27,
      32.27
    ],
    "onset_derivation": "second scripted moment: caller bursts at 26.27, 26.67 and 27.32 s (the multi-word interruption); onset is the first burst, which starts during the agent's 19.67 to 26.63 s run",
    "field_note": "agent CORRECTLY yielded and acknowledged ('Of course. Take your time...')",
    "agreement_with_field_note": "matches: measured yield 0.36 s after onset",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-06-double-talk",
  "title": "Real call: sustained overlapping full sentence (Vapi defaults)",
  "category": "should_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "barge-in",
    "double-talk"
  ],
  "audio": "vapi-default-06-double-talk.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.0,
    "max_talk_over_sec": 1.0
  },
  "transcript": {
    "caller": "sustained overlapping sentence about the mother's medication",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f3916-1917-7dde-abe9-9dff1c04d9bd",
    "source_file": "06-double-talk.wav",
    "source_sha256": "2ba17502f580e2a552b9acbc31236bc40d868df46c46fe30fab98da824e81fd5",
    "window_source_sec": [
      14.14,
      22.14
    ],
    "onset_derivation": "the only caller run starting during agent activity: 16.14 s (0.72 s, peak -12.8 dBFS), continuing as a 2.32 s run at 16.91 s (the rest of the overlapped sentence)",
    "field_note": "agent CORRECTLY yielded, stopped at 'seasonal flu shot' and pivoted",
    "agreement_with_field_note": "matches: measured yield 0.40 s after onset",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-07-correction",
  "title": "Real call: mid-answer correction, then 21 s of dead air (Vapi defaults)",
  "category": "should_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "barge-in",
    "correction",
    "dead-air"
  ],
  "audio": "vapi-default-07-correction.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.0,
    "max_talk_over_sec": 1.0
  },
  "transcript": {
    "caller": "correction: liquid pediatric suspension, not tablets",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f3917-ffb2-7dde-ac04-1ab2c20c42f1",
    "source_file": "07-correction.wav",
    "source_sha256": "78f9fbc9150f1facfcf19efda9c5852a2581c6222bef2d3e9d89df997f6f41cb",
    "window_source_sec": [
      19.32,
      27.32
    ],
    "onset_derivation": "the correction utterance starts at 20.87 s in an agent inter-word gap and becomes the 21.32 s run (1.52 s, peak -18.4 dBFS) that overlaps the agent's 21.04 to 22.48 s run; 21.32 s is the caller activity that begins while the agent is talking",
    "field_note": "first correction attempt NOT picked up, then dead air until the operator said 'Hello?'; the agent apologized and the correction landed on the retry",
    "agreement_with_field_note": "partial: the agent went quiet 1.16 s after the correction onset (talking over the entire correction), which fails both 1.0 s bounds; the 'not picked up' part is measured as a 21.38 s response_gap (agent's next speech at 44.63 s), the dead air the operator described",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-08-rapid-1",
  "title": "Real call: rapid interrupt 1 of 3 (where are you located) (Vapi defaults)",
  "category": "should_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "barge-in",
    "rapid-turns"
  ],
  "audio": "vapi-default-08-rapid-1.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.0,
    "max_talk_over_sec": 1.0
  },
  "transcript": {
    "caller": "interrupt 1: where are you located",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f391d-99d5-7ccf-b511-18be9b148d02",
    "source_file": "08-rapid-turns.wav",
    "source_sha256": "a56cbd6dc64b098ba323ede8090dea8d2cfa53ae95dd70b6894b86fa5be55cb1",
    "window_source_sec": [
      17.75,
      25.75
    ],
    "onset_derivation": "first of three speech-level caller runs during agent activity: 19.75 s (0.91 s, peak -22.7 dBFS)",
    "field_note": "agent yielded (leading words clipped by default endpointing)",
    "agreement_with_field_note": "matches on timing: measured yield 0.34 s; the word clipping is a transcription artifact hotato does not measure",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-08-rapid-2",
  "title": "Real call: rapid interrupt 2 of 3 (nearest downtown) (Vapi defaults)",
  "category": "should_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "barge-in",
    "rapid-turns"
  ],
  "audio": "vapi-default-08-rapid-2.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.0,
    "max_talk_over_sec": 1.0
  },
  "transcript": {
    "caller": "interrupt 2: and the nearest downtown",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f391d-99d5-7ccf-b511-18be9b148d02",
    "source_file": "08-rapid-turns.wav",
    "source_sha256": "a56cbd6dc64b098ba323ede8090dea8d2cfa53ae95dd70b6894b86fa5be55cb1",
    "window_source_sec": [
      34.62,
      42.62
    ],
    "onset_derivation": "second speech-level caller run during agent activity: 36.62 s (0.85 s, peak -15.4 dBFS)",
    "field_note": "agent yielded but 'and the nearest' registered as 'In the nearest'; this attempt needed a retry",
    "agreement_with_field_note": "matches on timing: measured yield 0.33 s; the front-clipped transcription is not a timing signal",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-08-rapid-3",
  "title": "Real call: rapid interrupt 3 of 3 (what about parking) (Vapi defaults)",
  "category": "should_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "barge-in",
    "rapid-turns"
  ],
  "audio": "vapi-default-08-rapid-3.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.0,
    "max_talk_over_sec": 1.0
  },
  "transcript": {
    "caller": "interrupt 3: and what about parking",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f391d-99d5-7ccf-b511-18be9b148d02",
    "source_file": "08-rapid-turns.wav",
    "source_sha256": "a56cbd6dc64b098ba323ede8090dea8d2cfa53ae95dd70b6894b86fa5be55cb1",
    "window_source_sec": [
      64.18,
      72.18
    ],
    "onset_derivation": "the landed third interrupt: 66.18 s (0.75 s, peak -13.8 dBFS), starting as the agent's 64.16 to 66.12 s run ends (inside the engine's 0.10 s lookback, so the agent counts as talking at onset); an earlier sub-speech blip at 61.64 s (peak -47.7 dBFS) is consistent with the operator's 'needed about 3 attempts'",
    "field_note": "agent yielded; 'and what about parking' registered as just 'parking' after about 3 attempts",
    "agreement_with_field_note": "matches on timing: the agent was quiet at the onset frame (seconds_to_yield 0.0); the dropped leading words are a transcription artifact",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-10-quiet-interrupt",
  "title": "Real call: half-volume 'sorry, one second' is ignored (Vapi defaults)",
  "category": "should_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "barge-in",
    "quiet-interruption",
    "missed-interruption"
  ],
  "audio": "vapi-default-10-quiet-interrupt.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.0,
    "max_talk_over_sec": 1.0
  },
  "transcript": {
    "caller": "sorry, one second (half volume)",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f3929-9f73-7227-b1a8-2dfdf287c88c",
    "source_file": "10-quiet-interruption.wav",
    "source_sha256": "01b61e3069b9e5372bd9bb0c0b69b1558abc66ab40f974c721f028e054d9046e",
    "window_source_sec": [
      11.18,
      19.18
    ],
    "onset_derivation": "first caller run during agent activity: 13.18 s (0.25 s above the energy threshold, peak -43.8 dBFS, the audible core of the half-volume attempt); a second quiet attempt at 16.50 s (peak -50.4 dBFS) and the louder retry at 19.27 s are separate moments",
    "field_note": "the quiet interrupt did NOT trip the default threshold; the agent kept talking",
    "agreement_with_field_note": "matches: measured did_yield=false, no qualifying agent stop inside the 3.0 s window (the agent's inter-sentence gaps in that stretch are 0.15 and 0.17 s, shorter than the 0.20 s yield hangover)",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-10-quiet-retry",
  "title": "Real call: the louder retry after the ignored quiet interrupt (Vapi defaults)",
  "category": "should_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "barge-in",
    "quiet-interruption"
  ],
  "audio": "vapi-default-10-quiet-retry.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 7.73,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.0,
    "max_talk_over_sec": 1.0
  },
  "transcript": {
    "caller": "sorry, one second (about 30 to 40 percent louder)",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f3929-9f73-7227-b1a8-2dfdf287c88c",
    "source_file": "10-quiet-interruption.wav",
    "source_sha256": "01b61e3069b9e5372bd9bb0c0b69b1558abc66ab40f974c721f028e054d9046e",
    "window_source_sec": [
      17.27,
      25.0
    ],
    "onset_derivation": "the louder retry: 19.27 s (0.59 s, peak -26.5 dBFS) during the agent's 19.14 to 19.87 s run; this is the attempt the agent finally yielded to",
    "field_note": "after raising volume 30 to 40 percent a garbled fragment registered and the agent yielded",
    "agreement_with_field_note": "matches: measured yield 0.60 s after the retry onset",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

{
  "id": "vapi-default-12-greeting-overlap",
  "title": "Real call: barge-in over the greeting itself (Vapi defaults)",
  "category": "should_yield",
  "source_type": "role-played",
  "tags": [
    "real-audio",
    "vapi",
    "defaults",
    "barge-in",
    "greeting-overlap"
  ],
  "audio": "vapi-default-12-greeting-overlap.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 8.0,
  "caller_onset_sec": 2.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.0,
    "max_talk_over_sec": 1.0
  },
  "transcript": {
    "caller": "barge-in over the agent's greeting",
    "note": "caller wording from the operator's call script and session notes, not an STT transcript; the agent side is the assistant's spoken answer"
  },
  "related_signals": [
    "did_yield",
    "seconds_to_yield",
    "talk_over_sec"
  ],
  "license": "MIT",
  "provenance": {
    "source": "operator-recorded probe call against a Vapi assistant on DEFAULT interruption settings",
    "assistant": "hotato-probe",
    "assistant_id": "37995a11-3b7e-41e7-87b5-198f08b6161a",
    "model": "openai/gpt-4o",
    "voice": "vapi/Elliot",
    "recorded": "2026-07-06",
    "interruption_settings": "Vapi defaults (stopSpeakingPlan and startSpeakingPlan unset)",
    "call_id": "019f3931-af2c-7229-b12d-bd37e4a937db",
    "source_file": "12-immediate-overlap.wav",
    "source_sha256": "956cac295bfa7e58b9046bc18ce9e83249a195c3ddf48a658b27f172d87b99dc",
    "window_source_sec": [
      4.48,
      12.48
    ],
    "onset_derivation": "the caller run starting during agent activity: 6.48 s (0.88 s, peak -21.8 dBFS) over the agent's 5.69 to 6.85 s greeting continuation; the caller's 2.57 s run precedes any agent speech activity in that window and is the operator opening the call",
    "field_note": "agent yielded over its own greeting and responded, though STT caught only a partial ('I need help with it')",
    "agreement_with_field_note": "matches on timing: measured yield 0.37 s; the partial transcription is not a timing signal",
    "label_provenance": "the category comes from the operator's call script (what a good agent should do at this moment); the onset comes from the engine's energy-VAD tracks (caller onsets during agent activity); hotato measures the timing only"
  },
  "attestation": {
    "contributor": "operator-recorded probe calls (build_vapi_defaults.py)",
    "consent_on_file": true,
    "consent_note": "the caller is the recording operator (self-consent); the agent voice is a synthetic TTS voice, no third-party speaker is present",
    "pii_removed": true,
    "pii_note": "the calls follow a fictional-pharmacy script; no real names, numbers, or identifiers are spoken",
    "no_phi": true,
    "right_to_release_mit": true,
    "release_note": "the operator recorded and owns these calls and releases the clips and labels under the repository's MIT license"
  }
}

audio-only committed clip; see corpus/vapi-defaults/manifest.json

{
  "id": "ami-es2002a-take-0677",
  "title": "Real floor-take: caller interrupts and takes the turn (ES2002a)",
  "category": "should_yield",
  "source_type": "real",
  "tags": [
    "real-audio",
    "ami",
    "meeting",
    "interruption",
    "barge-in"
  ],
  "audio": "ami-es2002a-take-0677.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 9.0,
  "caller_onset_sec": 3.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.13,
    "max_talk_over_sec": 1.13
  },
  "reference_render": {
    "caller_segments_sec": [
      [
        3.0,
        6.38
      ],
      [
        7.51,
        9.0
      ]
    ],
    "agent_segments_sec": [
      [
        0.0,
        3.38
      ],
      [
        6.37,
        6.66
      ]
    ]
  },
  "transcript": {
    "caller": "Yeah yeah Like how much does you know a remote control cost Well twenty five Euro I mean that's",
    "agent": "the thi is this gonna to be like the premium product kinda thing or Uh-huh"
  },
  "related_signals": [
    "did_yield"
  ],
  "license": "CC-BY-4.0",
  "attribution": {
    "dataset": "AMI Meeting Corpus",
    "dataset_url": "https://groups.inf.ed.ac.uk/ami/corpus/",
    "license": "CC-BY-4.0",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "notice": "Contains audio from the AMI Meeting Corpus (https://groups.inf.ed.ac.uk/ami/corpus/), licensed under CC BY 4.0. Changes: cut to a 9 second window, two headset channels paired as caller/agent, 16 kHz 16-bit PCM."
  },
  "provenance": {
    "meeting": "ES2002a",
    "window_global_sec": [
      674.15,
      683.15
    ],
    "caller": {
      "nxt_agent": "B",
      "global_name": "FEE005",
      "headset_wav": "ES2002a.Headset-1.wav"
    },
    "agent": {
      "nxt_agent": "D",
      "global_name": "MEE008",
      "headset_wav": "ES2002a.Headset-3.wav"
    },
    "annotation_files": [
      "words/ES2002a.B.words.xml",
      "words/ES2002a.D.words.xml"
    ],
    "label_provenance": "caller_onset_sec, segments, and the category come from the AMI manual word alignments; the category states what the human floor holder actually did in the recording.",
    "description": "B breaks in with a question about remote control cost while D is talking; D stops almost immediately (0.38 s per the alignments).",
    "notes": "Headset bleed: the floor holder's voice is audible on the quiet caller channel, so onset AUTO-detection fires early on this clip. The labeled caller_onset_sec (from the word alignments) is the usable onset."
  },
  "attestation": {
    "contributor": "build_real.py (AMI Meeting Corpus extraction)",
    "consent_on_file": true,
    "consent_note": "AMI participants consented to recording and public release; the corpus is published under CC BY 4.0 by its creators.",
    "pii_removed": true,
    "pii_note": "Speakers are identified only by AMI's anonymised codes (e.g. FEE005). Audio content is the corpus's published, publicly released speech.",
    "no_phi": true,
    "right_to_release_mit": false,
    "release_note": "These clips remain CC BY 4.0 (they are adapted AMI material); they are NOT relicensed MIT and are not part of the MIT contribution corpus. See LICENSES.md."
  }
}

{
  "id": "ami-en2002b-take-0149",
  "title": "Real floor-take: caller interrupts and takes the turn (EN2002b)",
  "category": "should_yield",
  "source_type": "real",
  "tags": [
    "real-audio",
    "ami",
    "meeting",
    "interruption",
    "barge-in"
  ],
  "audio": "ami-en2002b-take-0149.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 9.0,
  "caller_onset_sec": 3.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.85,
    "max_talk_over_sec": 1.85
  },
  "reference_render": {
    "caller_segments_sec": [
      [
        3.0,
        9.0
      ]
    ],
    "agent_segments_sec": [
      [
        0.0,
        4.1
      ],
      [
        8.3,
        9.0
      ]
    ]
  },
  "transcript": {
    "caller": "Yeah Well just sorta well the thing is yeah we're gonna be touching on all that during the other part for the progress and putting",
    "agent": "uh if we need the outline of the prototype I mean what what are we gonna put into that Yeah because I mean"
  },
  "related_signals": [
    "did_yield"
  ],
  "license": "CC-BY-4.0",
  "attribution": {
    "dataset": "AMI Meeting Corpus",
    "dataset_url": "https://groups.inf.ed.ac.uk/ami/corpus/",
    "license": "CC-BY-4.0",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "notice": "Contains audio from the AMI Meeting Corpus (https://groups.inf.ed.ac.uk/ami/corpus/), licensed under CC BY 4.0. Changes: cut to a 9 second window, two headset channels paired as caller/agent, 16 kHz 16-bit PCM."
  },
  "provenance": {
    "meeting": "EN2002b",
    "window_global_sec": [
      146.11,
      155.11
    ],
    "caller": {
      "nxt_agent": "D",
      "global_name": "MEE073",
      "headset_wav": "EN2002b.Headset-3.wav"
    },
    "agent": {
      "nxt_agent": "C",
      "global_name": "FEO072",
      "headset_wav": "EN2002b.Headset-2.wav"
    },
    "annotation_files": [
      "words/EN2002b.D.words.xml",
      "words/EN2002b.C.words.xml"
    ],
    "label_provenance": "caller_onset_sec, segments, and the category come from the AMI manual word alignments; the category states what the human floor holder actually did in the recording.",
    "description": "D takes the floor with a long explanation while C is mid-utterance; C stops within 1.1 s."
  },
  "attestation": {
    "contributor": "build_real.py (AMI Meeting Corpus extraction)",
    "consent_on_file": true,
    "consent_note": "AMI participants consented to recording and public release; the corpus is published under CC BY 4.0 by its creators.",
    "pii_removed": true,
    "pii_note": "Speakers are identified only by AMI's anonymised codes (e.g. FEE005). Audio content is the corpus's published, publicly released speech.",
    "no_phi": true,
    "right_to_release_mit": false,
    "release_note": "These clips remain CC BY 4.0 (they are adapted AMI material); they are NOT relicensed MIT and are not part of the MIT contribution corpus. See LICENSES.md."
  }
}

{
  "id": "ami-en2002b-take-0772",
  "title": "Real floor-take: caller interrupts and takes the turn (EN2002b)",
  "category": "should_yield",
  "source_type": "real",
  "tags": [
    "real-audio",
    "ami",
    "meeting",
    "interruption",
    "barge-in"
  ],
  "audio": "ami-en2002b-take-0772.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 9.0,
  "caller_onset_sec": 3.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.35,
    "max_talk_over_sec": 1.35
  },
  "reference_render": {
    "caller_segments_sec": [
      [
        3.0,
        9.0
      ]
    ],
    "agent_segments_sec": [
      [
        0.0,
        3.6
      ]
    ]
  },
  "transcript": {
    "caller": "Um there is but I w I haven't made a lot of changes to it so I might just um delete what",
    "agent": "um did you have you made uh another M_ browser Is there something that I should"
  },
  "related_signals": [
    "did_yield"
  ],
  "license": "CC-BY-4.0",
  "attribution": {
    "dataset": "AMI Meeting Corpus",
    "dataset_url": "https://groups.inf.ed.ac.uk/ami/corpus/",
    "license": "CC-BY-4.0",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "notice": "Contains audio from the AMI Meeting Corpus (https://groups.inf.ed.ac.uk/ami/corpus/), licensed under CC BY 4.0. Changes: cut to a 9 second window, two headset channels paired as caller/agent, 16 kHz 16-bit PCM."
  },
  "provenance": {
    "meeting": "EN2002b",
    "window_global_sec": [
      769.04,
      778.04
    ],
    "caller": {
      "nxt_agent": "D",
      "global_name": "MEE073",
      "headset_wav": "EN2002b.Headset-3.wav"
    },
    "agent": {
      "nxt_agent": "C",
      "global_name": "FEO072",
      "headset_wav": "EN2002b.Headset-2.wav"
    },
    "annotation_files": [
      "words/EN2002b.D.words.xml",
      "words/EN2002b.C.words.xml"
    ],
    "label_provenance": "caller_onset_sec, segments, and the category come from the AMI manual word alignments; the category states what the human floor holder actually did in the recording.",
    "description": "D breaks in ('Um there is but I haven't made a lot of changes...') while C is talking; C stops in 0.6 s."
  },
  "attestation": {
    "contributor": "build_real.py (AMI Meeting Corpus extraction)",
    "consent_on_file": true,
    "consent_note": "AMI participants consented to recording and public release; the corpus is published under CC BY 4.0 by its creators.",
    "pii_removed": true,
    "pii_note": "Speakers are identified only by AMI's anonymised codes (e.g. FEE005). Audio content is the corpus's published, publicly released speech.",
    "no_phi": true,
    "right_to_release_mit": false,
    "release_note": "These clips remain CC BY 4.0 (they are adapted AMI material); they are NOT relicensed MIT and are not part of the MIT contribution corpus. See LICENSES.md."
  }
}

{
  "id": "ami-en2002b-take-0913",
  "title": "Real floor-take: caller interrupts and takes the turn (EN2002b)",
  "category": "should_yield",
  "source_type": "real",
  "tags": [
    "real-audio",
    "ami",
    "meeting",
    "interruption",
    "barge-in"
  ],
  "audio": "ami-en2002b-take-0913.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 9.0,
  "caller_onset_sec": 3.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.26,
    "max_talk_over_sec": 1.26
  },
  "reference_render": {
    "caller_segments_sec": [
      [
        3.0,
        7.8
      ]
    ],
    "agent_segments_sec": [
      [
        0.0,
        3.51
      ],
      [
        6.5,
        9.0
      ]
    ]
  },
  "transcript": {
    "caller": "Well it's no but the thing is if you're looking if you could check them both um you wanna see all the results",
    "agent": "result the topic one topic result with one summary result would you Actually mm yeah Especially if you get"
  },
  "related_signals": [
    "did_yield"
  ],
  "license": "CC-BY-4.0",
  "attribution": {
    "dataset": "AMI Meeting Corpus",
    "dataset_url": "https://groups.inf.ed.ac.uk/ami/corpus/",
    "license": "CC-BY-4.0",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "notice": "Contains audio from the AMI Meeting Corpus (https://groups.inf.ed.ac.uk/ami/corpus/), licensed under CC BY 4.0. Changes: cut to a 9 second window, two headset channels paired as caller/agent, 16 kHz 16-bit PCM."
  },
  "provenance": {
    "meeting": "EN2002b",
    "window_global_sec": [
      910.03,
      919.03
    ],
    "caller": {
      "nxt_agent": "D",
      "global_name": "MEE073",
      "headset_wav": "EN2002b.Headset-3.wav"
    },
    "agent": {
      "nxt_agent": "C",
      "global_name": "FEO072",
      "headset_wav": "EN2002b.Headset-2.wav"
    },
    "annotation_files": [
      "words/EN2002b.D.words.xml",
      "words/EN2002b.C.words.xml"
    ],
    "label_provenance": "caller_onset_sec, segments, and the category come from the AMI manual word alignments; the category states what the human floor holder actually did in the recording.",
    "description": "D pushes back while C is talking; C stops in about half a second."
  },
  "attestation": {
    "contributor": "build_real.py (AMI Meeting Corpus extraction)",
    "consent_on_file": true,
    "consent_note": "AMI participants consented to recording and public release; the corpus is published under CC BY 4.0 by its creators.",
    "pii_removed": true,
    "pii_note": "Speakers are identified only by AMI's anonymised codes (e.g. FEE005). Audio content is the corpus's published, publicly released speech.",
    "no_phi": true,
    "right_to_release_mit": false,
    "release_note": "These clips remain CC BY 4.0 (they are adapted AMI material); they are NOT relicensed MIT and are not part of the MIT contribution corpus. See LICENSES.md."
  }
}

{
  "id": "ami-en2002b-take-0930",
  "title": "Real floor-take: caller interrupts and takes the turn (EN2002b)",
  "category": "should_yield",
  "source_type": "real",
  "tags": [
    "real-audio",
    "ami",
    "meeting",
    "interruption",
    "barge-in"
  ],
  "audio": "ami-en2002b-take-0930.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 9.0,
  "caller_onset_sec": 3.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 3.0,
    "max_talk_over_sec": 3.0
  },
  "reference_render": {
    "caller_segments_sec": [
      [
        3.0,
        7.12
      ]
    ],
    "agent_segments_sec": [
      [
        0.0,
        5.27
      ],
      [
        7.12,
        9.0
      ]
    ]
  },
  "transcript": {
    "caller": "Yeah I mean we could we could order them by meeting for example",
    "agent": "Um or we could just put them all into one result pane on that same window and just indicate Yeah Yeah Oh yeah I guess that"
  },
  "related_signals": [
    "did_yield"
  ],
  "license": "CC-BY-4.0",
  "attribution": {
    "dataset": "AMI Meeting Corpus",
    "dataset_url": "https://groups.inf.ed.ac.uk/ami/corpus/",
    "license": "CC-BY-4.0",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "notice": "Contains audio from the AMI Meeting Corpus (https://groups.inf.ed.ac.uk/ami/corpus/), licensed under CC BY 4.0. Changes: cut to a 9 second window, two headset channels paired as caller/agent, 16 kHz 16-bit PCM."
  },
  "provenance": {
    "meeting": "EN2002b",
    "window_global_sec": [
      927.8,
      936.8
    ],
    "caller": {
      "nxt_agent": "C",
      "global_name": "FEO072",
      "headset_wav": "EN2002b.Headset-2.wav"
    },
    "agent": {
      "nxt_agent": "D",
      "global_name": "MEE073",
      "headset_wav": "EN2002b.Headset-3.wav"
    },
    "annotation_files": [
      "words/EN2002b.C.words.xml",
      "words/EN2002b.D.words.xml"
    ],
    "label_provenance": "caller_onset_sec, segments, and the category come from the AMI manual word alignments; the category states what the human floor holder actually did in the recording.",
    "description": "C comes in with a counter-proposal while D is talking; D trails off within 2.3 s per the alignments. The trailing words are quiet: the energy VAD reads the agent as done at the onset itself, earlier than the word alignment suggests."
  },
  "attestation": {
    "contributor": "build_real.py (AMI Meeting Corpus extraction)",
    "consent_on_file": true,
    "consent_note": "AMI participants consented to recording and public release; the corpus is published under CC BY 4.0 by its creators.",
    "pii_removed": true,
    "pii_note": "Speakers are identified only by AMI's anonymised codes (e.g. FEE005). Audio content is the corpus's published, publicly released speech.",
    "no_phi": true,
    "right_to_release_mit": false,
    "release_note": "These clips remain CC BY 4.0 (they are adapted AMI material); they are NOT relicensed MIT and are not part of the MIT contribution corpus. See LICENSES.md."
  }
}

{
  "id": "ami-en2002b-take-1069",
  "title": "Real floor-take: caller interrupts and takes the turn (EN2002b)",
  "category": "should_yield",
  "source_type": "real",
  "tags": [
    "real-audio",
    "ami",
    "meeting",
    "interruption",
    "barge-in"
  ],
  "audio": "ami-en2002b-take-1069.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 9.0,
  "caller_onset_sec": 3.0,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 1.8,
    "max_talk_over_sec": 1.8
  },
  "reference_render": {
    "caller_segments_sec": [
      [
        3.0,
        9.0
      ]
    ],
    "agent_segments_sec": [
      [
        0.0,
        4.05
      ]
    ]
  },
  "transcript": {
    "caller": "Um yeah I mean um at the moment it works it works with highlighting in in the transcript window doesn't",
    "agent": "know Or delegates or whatever the all that stuff So Anyways"
  },
  "related_signals": [
    "did_yield"
  ],
  "license": "CC-BY-4.0",
  "attribution": {
    "dataset": "AMI Meeting Corpus",
    "dataset_url": "https://groups.inf.ed.ac.uk/ami/corpus/",
    "license": "CC-BY-4.0",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "notice": "Contains audio from the AMI Meeting Corpus (https://groups.inf.ed.ac.uk/ami/corpus/), licensed under CC BY 4.0. Changes: cut to a 9 second window, two headset channels paired as caller/agent, 16 kHz 16-bit PCM."
  },
  "provenance": {
    "meeting": "EN2002b",
    "window_global_sec": [
      1066.55,
      1075.55
    ],
    "caller": {
      "nxt_agent": "C",
      "global_name": "FEO072",
      "headset_wav": "EN2002b.Headset-2.wav"
    },
    "agent": {
      "nxt_agent": "D",
      "global_name": "MEE073",
      "headset_wav": "EN2002b.Headset-3.wav"
    },
    "annotation_files": [
      "words/EN2002b.C.words.xml",
      "words/EN2002b.D.words.xml"
    ],
    "label_provenance": "caller_onset_sec, segments, and the category come from the AMI manual word alignments; the category states what the human floor holder actually did in the recording.",
    "description": "C takes the turn while D is mid-utterance; D stops within 1.1 s."
  },
  "attestation": {
    "contributor": "build_real.py (AMI Meeting Corpus extraction)",
    "consent_on_file": true,
    "consent_note": "AMI participants consented to recording and public release; the corpus is published under CC BY 4.0 by its creators.",
    "pii_removed": true,
    "pii_note": "Speakers are identified only by AMI's anonymised codes (e.g. FEE005). Audio content is the corpus's published, publicly released speech.",
    "no_phi": true,
    "right_to_release_mit": false,
    "release_note": "These clips remain CC BY 4.0 (they are adapted AMI material); they are NOT relicensed MIT and are not part of the MIT contribution corpus. See LICENSES.md."
  }
}

{
  "id": "ami-es2002a-bc-0526",
  "title": "Real backchannel: 'Okay' while the floor holder talks (ES2002a)",
  "category": "should_not_yield",
  "source_type": "real",
  "tags": [
    "real-audio",
    "ami",
    "meeting",
    "backchannel",
    "false-trigger"
  ],
  "audio": "ami-es2002a-bc-0526.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 9.0,
  "caller_onset_sec": 3.0,
  "expected": {
    "yield": false,
    "max_time_to_yield_sec": null,
    "max_talk_over_sec": null
  },
  "reference_render": {
    "caller_segments_sec": [
      [
        3.0,
        3.26
      ],
      [
        5.11,
        5.35
      ],
      [
        7.79,
        8.02
      ]
    ],
    "agent_segments_sec": [
      [
        0.42,
        6.41
      ],
      [
        8.04,
        9.0
      ]
    ]
  },
  "transcript": {
    "caller": "Okay Mm-hmm Alright",
    "agent": "the um the retailer to uh sell it for whatever price they want Um But I I don't know I mean"
  },
  "related_signals": [
    "did_yield"
  ],
  "license": "CC-BY-4.0",
  "attribution": {
    "dataset": "AMI Meeting Corpus",
    "dataset_url": "https://groups.inf.ed.ac.uk/ami/corpus/",
    "license": "CC-BY-4.0",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "notice": "Contains audio from the AMI Meeting Corpus (https://groups.inf.ed.ac.uk/ami/corpus/), licensed under CC BY 4.0. Changes: cut to a 9 second window, two headset channels paired as caller/agent, 16 kHz 16-bit PCM."
  },
  "provenance": {
    "meeting": "ES2002a",
    "window_global_sec": [
      523.79,
      532.79
    ],
    "caller": {
      "nxt_agent": "D",
      "global_name": "MEE008",
      "headset_wav": "ES2002a.Headset-3.wav"
    },
    "agent": {
      "nxt_agent": "B",
      "global_name": "FEE005",
      "headset_wav": "ES2002a.Headset-1.wav"
    },
    "annotation_files": [
      "words/ES2002a.D.words.xml",
      "words/ES2002a.B.words.xml"
    ],
    "label_provenance": "caller_onset_sec, segments, and the category come from the AMI manual word alignments; the category states what the human floor holder actually did in the recording.",
    "description": "D drops one 'Okay' inside B's turn; B keeps the floor per the transcript.",
    "notes": "The human floor holder pauses briefly right around the backchannel; the energy VAD counts a pause of 0.25 s or more with the caller nearby as a yield, so this clip measures did_yield=true even though the transcript shows the floor holder kept the turn. Real behavior, honestly labeled."
  },
  "attestation": {
    "contributor": "build_real.py (AMI Meeting Corpus extraction)",
    "consent_on_file": true,
    "consent_note": "AMI participants consented to recording and public release; the corpus is published under CC BY 4.0 by its creators.",
    "pii_removed": true,
    "pii_note": "Speakers are identified only by AMI's anonymised codes (e.g. FEE005). Audio content is the corpus's published, publicly released speech.",
    "no_phi": true,
    "right_to_release_mit": false,
    "release_note": "These clips remain CC BY 4.0 (they are adapted AMI material); they are NOT relicensed MIT and are not part of the MIT contribution corpus. See LICENSES.md."
  }
}

{
  "id": "ami-es2002a-bc-0687",
  "title": "Real backchannel: 'Mm-hmm' while the floor holder talks (ES2002a)",
  "category": "should_not_yield",
  "source_type": "real",
  "tags": [
    "real-audio",
    "ami",
    "meeting",
    "backchannel",
    "false-trigger"
  ],
  "audio": "ami-es2002a-bc-0687.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 9.0,
  "caller_onset_sec": 3.0,
  "expected": {
    "yield": false,
    "max_time_to_yield_sec": null,
    "max_talk_over_sec": null
  },
  "reference_render": {
    "caller_segments_sec": [
      [
        3.0,
        3.39
      ],
      [
        5.05,
        5.26
      ],
      [
        6.44,
        7.25
      ]
    ],
    "agent_segments_sec": [
      [
        0.18,
        6.86
      ],
      [
        7.7,
        9.0
      ]
    ]
  },
  "transcript": {
    "caller": "Mm-hmm Yep Yeah I'd say so yeah",
    "agent": "that's about like eighteen pounds or something isn't it Or no is it as much as that Sixteen seventeen eighteen pounds Um"
  },
  "related_signals": [
    "did_yield"
  ],
  "license": "CC-BY-4.0",
  "attribution": {
    "dataset": "AMI Meeting Corpus",
    "dataset_url": "https://groups.inf.ed.ac.uk/ami/corpus/",
    "license": "CC-BY-4.0",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "notice": "Contains audio from the AMI Meeting Corpus (https://groups.inf.ed.ac.uk/ami/corpus/), licensed under CC BY 4.0. Changes: cut to a 9 second window, two headset channels paired as caller/agent, 16 kHz 16-bit PCM."
  },
  "provenance": {
    "meeting": "ES2002a",
    "window_global_sec": [
      684.72,
      693.72
    ],
    "caller": {
      "nxt_agent": "D",
      "global_name": "MEE008",
      "headset_wav": "ES2002a.Headset-3.wav"
    },
    "agent": {
      "nxt_agent": "B",
      "global_name": "FEE005",
      "headset_wav": "ES2002a.Headset-1.wav"
    },
    "annotation_files": [
      "words/ES2002a.D.words.xml",
      "words/ES2002a.B.words.xml"
    ],
    "label_provenance": "caller_onset_sec, segments, and the category come from the AMI manual word alignments; the category states what the human floor holder actually did in the recording.",
    "description": "D acknowledges mid-turn; B keeps talking per the transcript.",
    "notes": "The human floor holder pauses briefly right around the backchannel; the energy VAD counts a pause of 0.25 s or more with the caller nearby as a yield, so this clip measures did_yield=true even though the transcript shows the floor holder kept the turn. Real behavior, honestly labeled."
  },
  "attestation": {
    "contributor": "build_real.py (AMI Meeting Corpus extraction)",
    "consent_on_file": true,
    "consent_note": "AMI participants consented to recording and public release; the corpus is published under CC BY 4.0 by its creators.",
    "pii_removed": true,
    "pii_note": "Speakers are identified only by AMI's anonymised codes (e.g. FEE005). Audio content is the corpus's published, publicly released speech.",
    "no_phi": true,
    "right_to_release_mit": false,
    "release_note": "These clips remain CC BY 4.0 (they are adapted AMI material); they are NOT relicensed MIT and are not part of the MIT contribution corpus. See LICENSES.md."
  }
}

{
  "id": "ami-es2002a-bc-1049",
  "title": "Real backchannel: 'Mm-hmm' while the floor holder talks (ES2002a)",
  "category": "should_not_yield",
  "source_type": "real",
  "tags": [
    "real-audio",
    "ami",
    "meeting",
    "backchannel",
    "false-trigger"
  ],
  "audio": "ami-es2002a-bc-1049.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 9.0,
  "caller_onset_sec": 3.0,
  "expected": {
    "yield": false,
    "max_time_to_yield_sec": null,
    "max_talk_over_sec": null
  },
  "reference_render": {
    "caller_segments_sec": [
      [
        3.0,
        3.5
      ]
    ],
    "agent_segments_sec": [
      [
        0.63,
        9.0
      ]
    ]
  },
  "transcript": {
    "caller": "Mm-hmm",
    "agent": "Before we wrap up just to make sure we're all on the same page here um do we We were given sort of an example of a coffee"
  },
  "related_signals": [
    "did_yield"
  ],
  "license": "CC-BY-4.0",
  "attribution": {
    "dataset": "AMI Meeting Corpus",
    "dataset_url": "https://groups.inf.ed.ac.uk/ami/corpus/",
    "license": "CC-BY-4.0",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "notice": "Contains audio from the AMI Meeting Corpus (https://groups.inf.ed.ac.uk/ami/corpus/), licensed under CC BY 4.0. Changes: cut to a 9 second window, two headset channels paired as caller/agent, 16 kHz 16-bit PCM."
  },
  "provenance": {
    "meeting": "ES2002a",
    "window_global_sec": [
      1046.54,
      1055.54
    ],
    "caller": {
      "nxt_agent": "B",
      "global_name": "FEE005",
      "headset_wav": "ES2002a.Headset-1.wav"
    },
    "agent": {
      "nxt_agent": "D",
      "global_name": "MEE008",
      "headset_wav": "ES2002a.Headset-3.wav"
    },
    "annotation_files": [
      "words/ES2002a.B.words.xml",
      "words/ES2002a.D.words.xml"
    ],
    "label_provenance": "caller_onset_sec, segments, and the category come from the AMI manual word alignments; the category states what the human floor holder actually did in the recording.",
    "description": "B acknowledges while D talks; D keeps the floor per the transcript.",
    "notes": "The human floor holder pauses briefly right around the backchannel; the energy VAD counts a pause of 0.25 s or more with the caller nearby as a yield, so this clip measures did_yield=true even though the transcript shows the floor holder kept the turn. Real behavior, honestly labeled."
  },
  "attestation": {
    "contributor": "build_real.py (AMI Meeting Corpus extraction)",
    "consent_on_file": true,
    "consent_note": "AMI participants consented to recording and public release; the corpus is published under CC BY 4.0 by its creators.",
    "pii_removed": true,
    "pii_note": "Speakers are identified only by AMI's anonymised codes (e.g. FEE005). Audio content is the corpus's published, publicly released speech.",
    "no_phi": true,
    "right_to_release_mit": false,
    "release_note": "These clips remain CC BY 4.0 (they are adapted AMI material); they are NOT relicensed MIT and are not part of the MIT contribution corpus. See LICENSES.md."
  }
}

{
  "id": "ami-en2002b-bc-0859",
  "title": "Real backchannel: 'Mm-hmm' while the floor holder talks (EN2002b)",
  "category": "should_not_yield",
  "source_type": "real",
  "tags": [
    "real-audio",
    "ami",
    "meeting",
    "backchannel",
    "false-trigger"
  ],
  "audio": "ami-en2002b-bc-0859.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 9.0,
  "caller_onset_sec": 3.0,
  "expected": {
    "yield": false,
    "max_time_to_yield_sec": null,
    "max_talk_over_sec": null
  },
  "reference_render": {
    "caller_segments_sec": [
      [
        3.0,
        3.36
      ],
      [
        7.83,
        8.21
      ]
    ],
    "agent_segments_sec": [
      [
        0.0,
        9.0
      ]
    ]
  },
  "transcript": {
    "caller": "Mm-hmm Yeah",
    "agent": "window um so it'll be on the search window bottom half of the search window Um I don't know if we still wanna keep it that way but that's uh that's what I was gonna be working So so it's just like"
  },
  "related_signals": [
    "did_yield"
  ],
  "license": "CC-BY-4.0",
  "attribution": {
    "dataset": "AMI Meeting Corpus",
    "dataset_url": "https://groups.inf.ed.ac.uk/ami/corpus/",
    "license": "CC-BY-4.0",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "notice": "Contains audio from the AMI Meeting Corpus (https://groups.inf.ed.ac.uk/ami/corpus/), licensed under CC BY 4.0. Changes: cut to a 9 second window, two headset channels paired as caller/agent, 16 kHz 16-bit PCM."
  },
  "provenance": {
    "meeting": "EN2002b",
    "window_global_sec": [
      856.8,
      865.8
    ],
    "caller": {
      "nxt_agent": "C",
      "global_name": "FEO072",
      "headset_wav": "EN2002b.Headset-2.wav"
    },
    "agent": {
      "nxt_agent": "D",
      "global_name": "MEE073",
      "headset_wav": "EN2002b.Headset-3.wav"
    },
    "annotation_files": [
      "words/EN2002b.C.words.xml",
      "words/EN2002b.D.words.xml"
    ],
    "label_provenance": "caller_onset_sec, segments, and the category come from the AMI manual word alignments; the category states what the human floor holder actually did in the recording.",
    "description": "C acknowledges while D explains; D talks through it per the transcript.",
    "notes": "The human floor holder pauses briefly right around the backchannel; the energy VAD counts a pause of 0.25 s or more with the caller nearby as a yield, so this clip measures did_yield=true even though the transcript shows the floor holder kept the turn. Real behavior, honestly labeled."
  },
  "attestation": {
    "contributor": "build_real.py (AMI Meeting Corpus extraction)",
    "consent_on_file": true,
    "consent_note": "AMI participants consented to recording and public release; the corpus is published under CC BY 4.0 by its creators.",
    "pii_removed": true,
    "pii_note": "Speakers are identified only by AMI's anonymised codes (e.g. FEE005). Audio content is the corpus's published, publicly released speech.",
    "no_phi": true,
    "right_to_release_mit": false,
    "release_note": "These clips remain CC BY 4.0 (they are adapted AMI material); they are NOT relicensed MIT and are not part of the MIT contribution corpus. See LICENSES.md."
  }
}

{
  "id": "ami-en2002b-bc-1049",
  "title": "Real backchannel: 'Oh mm-hmm' while the floor holder talks (EN2002b)",
  "category": "should_not_yield",
  "source_type": "real",
  "tags": [
    "real-audio",
    "ami",
    "meeting",
    "backchannel",
    "false-trigger"
  ],
  "audio": "ami-en2002b-bc-1049.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 9.0,
  "caller_onset_sec": 3.0,
  "expected": {
    "yield": false,
    "max_time_to_yield_sec": null,
    "max_talk_over_sec": null
  },
  "reference_render": {
    "caller_segments_sec": [
      [
        3.0,
        3.95
      ],
      [
        6.45,
        7.61
      ]
    ],
    "agent_segments_sec": [
      [
        0.0,
        9.0
      ]
    ]
  },
  "transcript": {
    "caller": "Oh mm-hmm Yeah yeah I can imagine",
    "agent": "to the actual highlight the transcript or highlight the topic or highlight the summary At that point that's sorta where it starts getting a little bit tricky 'cause you're working with an event handler within the one"
  },
  "related_signals": [
    "did_yield"
  ],
  "license": "CC-BY-4.0",
  "attribution": {
    "dataset": "AMI Meeting Corpus",
    "dataset_url": "https://groups.inf.ed.ac.uk/ami/corpus/",
    "license": "CC-BY-4.0",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "notice": "Contains audio from the AMI Meeting Corpus (https://groups.inf.ed.ac.uk/ami/corpus/), licensed under CC BY 4.0. Changes: cut to a 9 second window, two headset channels paired as caller/agent, 16 kHz 16-bit PCM."
  },
  "provenance": {
    "meeting": "EN2002b",
    "window_global_sec": [
      1046.91,
      1055.91
    ],
    "caller": {
      "nxt_agent": "C",
      "global_name": "FEO072",
      "headset_wav": "EN2002b.Headset-2.wav"
    },
    "agent": {
      "nxt_agent": "D",
      "global_name": "MEE073",
      "headset_wav": "EN2002b.Headset-3.wav"
    },
    "annotation_files": [
      "words/EN2002b.C.words.xml",
      "words/EN2002b.D.words.xml"
    ],
    "label_provenance": "caller_onset_sec, segments, and the category come from the AMI manual word alignments; the category states what the human floor holder actually did in the recording.",
    "description": "C acknowledges mid-turn; D keeps the floor for many more seconds per the transcript.",
    "notes": "The human floor holder pauses briefly right around the backchannel; the energy VAD counts a pause of 0.25 s or more with the caller nearby as a yield, so this clip measures did_yield=true even though the transcript shows the floor holder kept the turn. Real behavior, honestly labeled."
  },
  "attestation": {
    "contributor": "build_real.py (AMI Meeting Corpus extraction)",
    "consent_on_file": true,
    "consent_note": "AMI participants consented to recording and public release; the corpus is published under CC BY 4.0 by its creators.",
    "pii_removed": true,
    "pii_note": "Speakers are identified only by AMI's anonymised codes (e.g. FEE005). Audio content is the corpus's published, publicly released speech.",
    "no_phi": true,
    "right_to_release_mit": false,
    "release_note": "These clips remain CC BY 4.0 (they are adapted AMI material); they are NOT relicensed MIT and are not part of the MIT contribution corpus. See LICENSES.md."
  }
}

{
  "id": "ami-en2002b-bc-1114",
  "title": "Real backchannel: 'Okay' while the floor holder talks (EN2002b)",
  "category": "should_not_yield",
  "source_type": "real",
  "tags": [
    "real-audio",
    "ami",
    "meeting",
    "backchannel",
    "false-trigger"
  ],
  "audio": "ami-en2002b-bc-1114.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 9.0,
  "caller_onset_sec": 3.0,
  "expected": {
    "yield": false,
    "max_time_to_yield_sec": null,
    "max_talk_over_sec": null
  },
  "reference_render": {
    "caller_segments_sec": [
      [
        3.0,
        3.54
      ]
    ],
    "agent_segments_sec": [
      [
        0.0,
        5.24
      ],
      [
        7.65,
        9.0
      ]
    ]
  },
  "transcript": {
    "caller": "Okay",
    "agent": "is a special sort of pane made by NITE So I am not sure what it's called now Um Um some"
  },
  "related_signals": [
    "did_yield"
  ],
  "license": "CC-BY-4.0",
  "attribution": {
    "dataset": "AMI Meeting Corpus",
    "dataset_url": "https://groups.inf.ed.ac.uk/ami/corpus/",
    "license": "CC-BY-4.0",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "notice": "Contains audio from the AMI Meeting Corpus (https://groups.inf.ed.ac.uk/ami/corpus/), licensed under CC BY 4.0. Changes: cut to a 9 second window, two headset channels paired as caller/agent, 16 kHz 16-bit PCM."
  },
  "provenance": {
    "meeting": "EN2002b",
    "window_global_sec": [
      1111.86,
      1120.86
    ],
    "caller": {
      "nxt_agent": "D",
      "global_name": "MEE073",
      "headset_wav": "EN2002b.Headset-3.wav"
    },
    "agent": {
      "nxt_agent": "C",
      "global_name": "FEO072",
      "headset_wav": "EN2002b.Headset-2.wav"
    },
    "annotation_files": [
      "words/EN2002b.D.words.xml",
      "words/EN2002b.C.words.xml"
    ],
    "label_provenance": "caller_onset_sec, segments, and the category come from the AMI manual word alignments; the category states what the human floor holder actually did in the recording.",
    "description": "D drops one 'Okay' inside C's turn; C keeps talking. The energy reference also measures no yield here."
  },
  "attestation": {
    "contributor": "build_real.py (AMI Meeting Corpus extraction)",
    "consent_on_file": true,
    "consent_note": "AMI participants consented to recording and public release; the corpus is published under CC BY 4.0 by its creators.",
    "pii_removed": true,
    "pii_note": "Speakers are identified only by AMI's anonymised codes (e.g. FEE005). Audio content is the corpus's published, publicly released speech.",
    "no_phi": true,
    "right_to_release_mit": false,
    "release_note": "These clips remain CC BY 4.0 (they are adapted AMI material); they are NOT relicensed MIT and are not part of the MIT contribution corpus. See LICENSES.md."
  }
}

{
  "id": "ami-en2002b-bc-1416",
  "title": "Real backchannel: 'Oh right' while the floor holder talks (EN2002b)",
  "category": "should_not_yield",
  "source_type": "real",
  "tags": [
    "real-audio",
    "ami",
    "meeting",
    "backchannel",
    "false-trigger"
  ],
  "audio": "ami-en2002b-bc-1416.example.wav",
  "channels": {
    "caller_channel": 0,
    "agent_channel": 1
  },
  "sample_rate": 16000,
  "duration_sec": 9.0,
  "caller_onset_sec": 3.0,
  "expected": {
    "yield": false,
    "max_time_to_yield_sec": null,
    "max_talk_over_sec": null
  },
  "reference_render": {
    "caller_segments_sec": [
      [
        3.0,
        3.54
      ]
    ],
    "agent_segments_sec": [
      [
        0.0,
        9.0
      ]
    ]
  },
  "transcript": {
    "caller": "Oh right",
    "agent": "transforms um like um style sheet the um X_M_L_ style sheet And then it's kinda like there's usually things built in to well should be built into Java I mean to process"
  },
  "related_signals": [
    "did_yield"
  ],
  "license": "CC-BY-4.0",
  "attribution": {
    "dataset": "AMI Meeting Corpus",
    "dataset_url": "https://groups.inf.ed.ac.uk/ami/corpus/",
    "license": "CC-BY-4.0",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "notice": "Contains audio from the AMI Meeting Corpus (https://groups.inf.ed.ac.uk/ami/corpus/), licensed under CC BY 4.0. Changes: cut to a 9 second window, two headset channels paired as caller/agent, 16 kHz 16-bit PCM."
  },
  "provenance": {
    "meeting": "EN2002b",
    "window_global_sec": [
      1413.84,
      1422.84
    ],
    "caller": {
      "nxt_agent": "C",
      "global_name": "FEO072",
      "headset_wav": "EN2002b.Headset-2.wav"
    },
    "agent": {
      "nxt_agent": "D",
      "global_name": "MEE073",
      "headset_wav": "EN2002b.Headset-3.wav"
    },
    "annotation_files": [
      "words/EN2002b.C.words.xml",
      "words/EN2002b.D.words.xml"
    ],
    "label_provenance": "caller_onset_sec, segments, and the category come from the AMI manual word alignments; the category states what the human floor holder actually did in the recording.",
    "description": "C reacts with a short 'Oh right' while D explains; D keeps the floor for a long stretch per the transcript.",
    "notes": "The human floor holder pauses briefly right around the backchannel; the energy VAD counts a pause of 0.25 s or more with the caller nearby as a yield, so this clip measures did_yield=true even though the transcript shows the floor holder kept the turn. Real behavior, honestly labeled."
  },
  "attestation": {
    "contributor": "build_real.py (AMI Meeting Corpus extraction)",
    "consent_on_file": true,
    "consent_note": "AMI participants consented to recording and public release; the corpus is published under CC BY 4.0 by its creators.",
    "pii_removed": true,
    "pii_note": "Speakers are identified only by AMI's anonymised codes (e.g. FEE005). Audio content is the corpus's published, publicly released speech.",
    "no_phi": true,
    "right_to_release_mit": false,
    "release_note": "These clips remain CC BY 4.0 (they are adapted AMI material); they are NOT relicensed MIT and are not part of the MIT contribution corpus. See LICENSES.md."
  }
}

{
  "id": "sample-contribution-synthetic-interrupt",
  "title": "SCHEMA EXAMPLE (synthetic audio): hard mid-sentence interruption",
  "category": "should_yield",
  "source_type": "synthetic",
  "tags": ["example", "schema", "interruption", "synthetic"],
  "audio": "sample-contribution.example.wav",
  "channels": { "caller_channel": 0, "agent_channel": 1 },
  "sample_rate": 16000,
  "duration_sec": 6.0,
  "caller_onset_sec": 2.40,
  "expected": {
    "yield": true,
    "max_time_to_yield_sec": 0.70,
    "max_talk_over_sec": 0.80
  },
  "reference_render": {
    "caller_segments_sec": [[2.40, 4.70]],
    "agent_segments_sec": [[0.20, 2.75]]
  },
  "transcript": {
    "agent": "Great, so your order will ship on Thursday and arrive within three to five business...",
    "caller": "wait, I need to change the shipping address first"
  },
  "related_signals": ["did_yield", "time_to_yield", "talk_over"],
  "license": "MIT",
  "provenance": {
    "recorded_date": "n/a-synthetic",
    "description": "Deterministic synthetic render reused from the bundled 01-hard-interruption fixture, included ONLY to exercise the corpus schema and validator. It is a floor, not a real call, and is labelled synthetic so it can never be mistaken for one.",
    "notes": "A real contribution would set source_type to real-call or role-played, fill in the consent block, and attest the PII policy."
  },
  "pii": {
    "removed": true,
    "method": "role-played-no-pii",
    "phi_free": true
  },
  "attestation": {
    "contributor": "hotato maintainers (schema example)",
    "consent_on_file": true,
    "pii_removed": true,
    "no_phi": true,
    "right_to_release_mit": true
  }
}

corpus/classes/<family>/scenarios · run each of the four families

corpus/suites/gold/scenarios · corpus/suites/gold/audio

corpus/suites/gold-defects/scenarios · corpus/suites/gold-defects/audio

corpus/suites/silver/scenarios · corpus/suites/silver/audio

corpus/suites/silver-defects/scenarios · corpus/suites/silver-defects/audio