On 16 November 2026 Ryan Mitchell stopped for 720 ms in the middle of naming a payment date, and Hazel, Steelyard Credit's AI collections agent, decided his turn was over after 300 ms and asked him to pay that day, speaking over the date (trace T05). The same day Sofia Ramirez cut in to say she had already paid while Hazel was reading out her balance; Hazel talked on for 900 ms and then asked her for the payment anyway (trace T02).
T05 is a question of where the endpointing threshold sits. T02 is two questions: how fast the agent stops, and whether it changes course afterwards. This page gives each its own number.
What happened on the two calls?
In T05 the turn detector ended Ryan's turn inside a hesitation before a date, and Hazel's audio began 230 ms after he had started talking again. In T02 the barge-in monitor needed 900 ms to silence Hazel, and her next turn carried on as if Sofia had said nothing.
T05, 16 November, AUTO-early-nov, still on its old 300 ms silence budget:
t_ms ch event / text
51200 caller vad_start REF: "I can pay it... next month on the... [pause 720 ms] ...fifth, I get paid then"
52950 caller pause begins after "next month on the"
53250 agent eot_commit (silence 300 ms; AUTO-early-nov budget was 300 ms)
53670 caller pause ends (720 ms); REF: "...fifth, I get paid then"
53900 agent TTS starts over the caller: "Could you make a payment today? That way you'd avoid a late fee."
Diagnosis: the detector committed at 53,250 ms on a pause the caller meant to hold, and playback started at 53,900 ms on top of "fifth, I get paid then", so the date never reached the STT cleanly. That is two defects: the commit itself, and no cancel of playback when caller speech came back in the 650 ms between eot_commit and play_start. A cancel-on-resume rule would have turned this false cutoff into a short extra wait instead of talk-over over the one word the call was for.
T02, 16 November, POS-short-nov, kitchen noise and a TV in the room:
t_ms ch event / text
30500 agent TTS (playing): "Sofia, el pago de su financiamiento, ciento treinta dolares, que vencia el cinco de noviembre..."
gloss: "Sofia, your financing payment, one hundred thirty dollars, which was due on November fifth..."
32160 caller vad_start REF: "oiga, ya lo pague, el tres, con mi tarjeta"
gloss: "listen, I already paid it, on the third, with my card"
32160 agent still playing (TV noise had raised the barge-in threshold)
33000 agent barge_in detected; tts_stop at 33060 (stop latency 900 ms)
34980 caller vad_end
35420 agent TTS: "Si, entiendo. Entonces, puede hacer el pago hoy?"
gloss: "Yes, I understand. So, can you make the payment today?"
Diagnosis: a missed barge-in, followed by a recovery failure. Of the 900 ms stop latency, 840 ms was detection, slowed because the TV had pushed the monitor's threshold up, and 60 ms was the path from detection to tts_stop. The recovery failure matters more for the account: no check_payment_status, no log_dispute, and a request for money straight after an "already paid" claim, which fails Steelyard's C08 policy check. Every span name used here is defined in what a voice agent trace contains.
Why is endpointing a curve and not a number?
Every turn detector trades cutoffs against waiting. Holding out longer after the caller goes quiet interrupts fewer callers mid-thought, but every caller who has finished waits longer. Sweeping the setting traces a curve of false cutoffs against dead air, and the setting you ship is one point you chose on it.
Endpointing is the call that the caller's turn is over and the agent may speak. In a cascaded pipeline that call is the turn detector's eot_commit. Any setting is described by two numbers:
- False-cutoff rate: of the labeled hold pauses (the caller was going to keep talking), the share on which the detector committed. Each such commit is one false cutoff.
- Dead air: on pauses labeled as true ends, the silence from speech end to the commit. It is the endpointing share of the end-of-turn delay, a term measuring voice agent latency defines alongside the rest of the latency vocabulary.
Silence is a poor signal because people barely leave any. Across languages the most common gap between turns is about 200 ms 1. In a 10-language study of answers to yes/no questions, the mean response offset was 208 ms and the median 100 ms, counting overlaps as negative offsets 2. A detector that waits for silence alone cannot answer that quickly without cutting callers off.
The most complete public measurement of the trade-off comes from a vendor: LiveKit built eot-bench, its dataset and the top-ranked model. It scores end-of-turn detectors on the false-cutoff rate at a 300 ms and a 600 ms latency budget, and on latency at a 5% and a 10% false-cutoff budget; latency here is dead air after the true end of turn, not compute time. The dataset holds 7,342 real human-to-agent user turns in 14 languages, in one validation split (English 2,288 rows), and labels every silence of 100 ms or more inside a turn as eot (true end) or hold (mid-turn pause) 3.
CAUTION
Vendor-reported. Treat it as a hypothesis to test on your own calls.
Four rows from the eot-bench README as of 19 November 2026, pooled English (general, accented and short_turns cohorts, v1.1), all vendor-reported 3:
| Detector | False cutoffs at 300 ms | At 600 ms |
|---|
| LiveKit Turn Detector v1.1 | 13.4% | 7.5% |
| Deepgram Flux | 22.3% | 13.5% |
| SmartTurn v3.2 | 46.8% | 22.9% |
| Silence-only VAD baseline | 79.3% | 33.4% |
The figures depend on the set as much as on the detector. LiveKit's June 2026 English set gave Turn Detector v1 a 9.9% false-cutoff rate at a 300 ms mean-latency budget, with Deepgram Flux at 12.9% and ultraVAD at 27.7% 4. On the harder November 2026 v1.1 set the same column reads 13.4% for Turn Detector v1.1, 22.3% for Flux and 39.7% for ultraVAD 3. ultraVAD's rate rose 12 points between the two sets.
The silence-only row is the one to read first, because it measures the pauses rather than any model. In the same benchmark's 400 Hindi rows the baseline cuts off 49.5% of holds at 300 ms, against 79.3% in English (vendor-run) 3: two languages differ in how people pause before any detector runs. We have not checked eot-bench's results for Spanish, and its dataset is 16 kHz audio of user turns, not 8 kHz collections calls with a TV on. So Steelyard keeps eot-bench's four headline metrics and its labeling rule, draws its own curve for each language group, and reuses none of the published operating points. A Spanish-dominant caller on a mobile line needs a curve drawn on Spanish-dominant callers on mobile lines.
Figure: A detector is a curve and a setting is one point on it. Synthetic Steelyard data, drawn for shape only: detector B sits below A and below silence-only at every budget, and the same detector on the slower pauses of the SDL book's older borrowers sits higher, which is why SDL-biwk-nov runs a longer budget than AUTO-early-nov.
Which three failures should you count separately?
Talk-over, early jump-in and missed barge-in are three metrics with three causes and three fixes. Talk-over is the agent speaking over a caller who had the floor. Early jump-in is a false cutoff on a hold pause. Missed barge-in is the agent failing to stop, or to change what it was doing, when the caller cuts in.
Steelyard computes each from the trace, on one clock:
- Talk-over: agent audio overlaps caller speech by more than 300 ms while the caller held the floor. For each agent playback interval and each caller speech segment that began before
play_start_ms, overlap = min(caller end, agent end) - max(caller start, play_start_ms). Reported as talk-over events per agent turn.
- Early jump-in (false cutoff): the agent takes the floor during a hold pause. On a labeled set, that is a commit on a pause labeled hold. On unlabeled production calls the proxy is caller speech that resumes after
eot_commit_ms and before, or just after, play_start_ms. Reported as false cutoffs over hold pauses at the campaign's budget.
- Missed barge-in: the caller cuts in with new information and the agent either keeps talking for more than 500 ms or keeps its old plan. Stop latency =
tts_stop minus the caller's vad_start for that barge-in. Talking on is a stop-latency failure; keeping the plan is a recovery failure, scored separately further down.
Figure: Three turn-taking failures on one clock (gray = caller speech, accent = agent audio). Synthetic Steelyard data; the middle lane has the timing of Ryan's T05 call and the bottom lane the timing of Sofia's T02 call. The causes differ by lane: playback that does not yield, a detector that commits too early, and a barge-in monitor that stops too late.
Keeping them apart pays off at fix time. Playback that yields when the caller speaks again cures talk-over. The operating point cures early jump-in: a longer budget or a higher threshold, in the call states where it matters. The barge-in monitor cures missed barge-in: faster detection, noise thresholds that do not creep upward the way they did in Sofia's kitchen, and a stop path with no buffered audio left to drain. A longer budget does nothing for T02, and a faster monitor does nothing for T05. One blended "interruption rate" moves when any of the three moves and cannot tell you which. The three also pull against each other: a monitor that stops on the first sound stops on every "uh-huh".
Measure all three on the platform clock from spans, then again at the handset on the dual-channel recording. The carrier legs bias them in opposite directions. After a barge-in the caller keeps hearing Hazel for the uplink plus downlink delay after playback stopped, so the stop latency the caller lives through is longer than the span says. For talk-over on a caller who held the floor, the same round trip makes the handset overlap shorter. Measuring voice agent latency shows how to recover the carrier offset. Report both clocks, and treat the handset number as the one that counts.
How do you build an end-of-turn test set for English and Spanish callers?
Start from real calls, cut out every silence of 100 ms or more inside a caller turn, and label each one end or hold with a content tag. Keep eot-bench's schema so its four metrics apply unchanged, and oversample the pauses that cost the most: hesitations right before an amount or a date.
Steelyard's gold-v7 holds 400 labeled pauses per language group (English, Spanish-dominant, Spanglish), built in six steps:
- Take production audio. Consented, redacted dual-channel recordings from the real phone path: 8 kHz, real carriers, real kitchens and cars. A set recorded on headsets at 16 kHz measures some other detector.
- Segment the caller channel alone, so agent audio cannot mask caller pauses, and keep every silence of at least 100 ms.
- Label with hindsight. Hold if the same speaker went on with the same turn; end if the floor changed hands. Record the duration, the call state (default, amount-asked, date-asked), the language group and a content tag (date, amount, small talk, dispute).
- Oversample the costly pauses. Hesitations before amounts and dates ("next month on the... fifth"), one-word replies ("yeah", "si" [yes], where a finished turn looks unfinished to a text model), side talk, and TV or road noise.
- Double-label a slice. Two reviewers label the same pauses, bilingual English-Spanish ones for the Spanish-dominant and Spanglish groups, and you compute Cohen's kappa before anyone tunes a threshold. A label two people cannot agree on is not fit to set a threshold.
- Keep the call id on every row. A hesitant speaker hesitates all through the call, so intervals must resample whole calls.
Two traps are specific to Spanish and Spanglish callers. A pause straight after "pero" [but] or "so" at a language switch usually means the caller is about to change language, not stop; a text model trained on single-language turns reads the clause before it as complete. In T08's opening, Valeria Cruz says "Si, pero my paycheck comes el viernes" [Yes, but my paycheck comes on Friday], and a pause after "pero" in a line like that is a hold. And Spanish amounts are long phrases with natural breaks inside them: "cuatrocientos veinticinco dolares" [four hundred twenty-five dollars] can carry a pause after "cuatrocientos" [four hundred], and a cutoff there splits one entity into two wrong ones. English has the same trap in a milder form: Steelyard's SBL guarantors say "a hundred twenty thousand" with a breath in the middle.
How do you choose an operating point per campaign and call state?
Put both errors in one unit, compute expected cost per 1,000 turns at each candidate setting, and take the cheapest, separately for each campaign and call state. A cutoff just before an amount or a date costs more than one in small talk, so states where Hazel has just asked for a number get longer budgets.
Steelyard's unit is one second of dead air on a finished turn. A false cutoff costs 1.0 unit during small talk, 1.5 during a dispute and 2.5 just before an amount or a date, because a cutoff there loses the very entity the call was placed to collect. The campaign owner signs these weights.
Here is a worked example on 40 labeled pauses from AUTO-early-nov, synthetic and invented for this page, at four silence budgets with the detector at thr 0.70. A hold pause is cut off if it lasts at least as long as the budget and the detector does not hold it by itself. Dead air on a true end equals the budget, plus 300 ms on one-word replies the detector is unsure about.
state ends holds (ms, tag, H = detector holds at thr 0.70)
default 12 130 st, 180 st, 320 st, 370 st, 450 st, 560 st H, 340 dispute H, 610 dispute
(2 of the 12 ends are one-word replies: +300 ms dead air each)
amount-asked 8 140, 210, 280, 330, 360 H, 390, 440, 480, 550, 720 H, 760, 940 (all before an amount or date)
(1 of the 8 ends is a one-word reply)
The false-cutoff rate on holds and the cost, per state:
| State | Budget | False cutoffs | Weighted cutoffs | Dead air (s) | Cost per 1,000 turns |
|---|
| Default (12 turns) | 300 ms | 4/8 = 50.0% | 4.5 | 4.2 | 725 |
| 400 ms | 2/8 = 25.0% | 2.5 | 5.4 | 658 |
| 500 ms | 1/8 = 12.5% | 1.5 | 6.6 | 675 |
| 600 ms | 1/8 = 12.5% | 1.5 | 7.8 | 775 |
| Amount-asked (8 turns) | 300 ms | 7/12 = 58.3% | 17.5 | 2.7 | 2,525 |
| 400 ms | 5/12 = 41.7% | 12.5 | 3.5 | 2,000 |
| 500 ms | 3/12 = 25.0% | 7.5 | 4.3 | 1,475 |
| 600 ms | 2/12 = 16.7% | 5.0 | 5.1 | 1,263 |
Cost per 1,000 turns = (weighted cutoffs + dead air) x 1,000 / turns. For the default state at 400 ms that is (2.5 + 5.4) x 1,000 / 12 = 658. Pooled over both states, false cutoffs fall from 11/20 = 55% at 300 ms to 15% at 600 ms, while mean dead air climbs from 345 to 645 ms.
Reading the table:
- The default state settles at 400 ms. Going from 300 to 400 ms removes 2.0 units of cutoffs for 1.2 s more dead air. Going on to 500 ms removes 1.0 unit for another 1.2 s, a net loss. That is the 17 November change Steelyard made after Ryan's call.
- The amount-asked state keeps paying off up to 600 ms, since every cutoff removed is worth 2.5 units and each extra 100 ms costs 0.8 s spread over 8 turns.
- One pooled curve would pick a single compromise that is wrong for both states.
- Two amount holds are still cut off at 600 ms (760 and 940 ms). The closed read-back is what catches those, not a still longer budget.
- The set is oversampled on purpose, toward long holds. Reweight each stratum to production frequencies before comparing a rate with a tripwire, or the tripwire will fire on your sampling plan.
Forty pauses cannot settle the choice. In the synthetic run of the endpointing recipe on at least 600 pauses per language group, AUTO-early-nov's default-state choice moves between 350 and 450 ms across call-clustered resamples, so the team's 400 ms is consistent with the data rather than forced by it. In the same run, doubling the amount and date weights moves AUTO's amount-asked choice by 200 ms: the weights are the real decision and deserve more argument than the threshold they produce.
Steelyard's operating points, from 17 November 2026:
| Campaign | Default silence budget | Amount-asked / date-asked | Detector threshold |
|---|
| AUTO-early-nov | 300 ms (before T05) -> 400 ms | 300 ms (before T05) -> 600 ms | td-1.1 thr 0.62 -> 0.70 |
| PIL-digital-nov | 400 ms | 600 ms | thr 0.70 |
| POS-short-nov | 400 ms | 600 ms | thr 0.70 |
| SBL-guar-nov | 450 ms | 650 ms | thr 0.72 |
| SDL-biwk-nov | 600 ms | 800 ms | thr 0.78 |
SDL-biwk-nov gets the longest budgets because the credit-builder book's borrowers skew older and pause longer, so its curve sits higher (the dashed line in the first figure). SBL runs a little longer than the rest because its guarantors state six-figure amounts with breaks inside them.
Two rules keep these settings honest. First, an operating point belongs to one detector version, so the trace pins it (td-1.1@thr0.62 in the versions block of the T01 trace, before the fix) and every reported rate names it. Second, redraw the curve whenever the detector, the STT model or the agent SDK changes. td-1.1 scores STT text as well as audio, so an STT upgrade shifts its curve, and an SDK upgrade can change a default nobody chose. The general version of this problem is covered in environmental drift.
How do you tell a barge-in from a backchannel?
Label every caller sound during agent speech as barge-in, backchannel, side talk or background, and score the stop decision against each label. A barge-in (new information or a request to stop) should stop the agent. A backchannel, speech aimed at someone else in the room, or a TV should not.
Backchannels on Steelyard's calls are "uh-huh", "mm-hm", "yeah", "aja" [uh-huh] and "si si" [yes, yes]; background is the TV or traffic. The score is a four-row table: for each label, the share of events on which Hazel stopped. Stop latency is reported on the barge-in row only, as a P95 sliced by noise condition, because T02's 840 ms of detection came from a threshold the TV had raised.
Full-Duplex-Bench v1.5 shows why every row needs a number. It puts five full-duplex systems through four overlap events (user interruption, user backchannel, user talking to others, background speech) with 200 synthesized interruptions. GPT-4o Realtime stops fastest on real interruptions (stop latency 0.23 s, respond rate 0.78) but also responds to background speech 93% of the time and to speech meant for someone else 91% of the time; Gemini and Nova Sonic hold the floor through side speech but take more than 2 s to stop when the user really does interrupt. The authors report the same trade-off across systems, between fast responses and steady floor control 5. These are mid-2025 endpoints, English only, with simulated overlaps, and current products may behave differently. The lesson carries over: tune a monitor on the barge-in row alone and it will fail the side-talk and background rows.
Collections adds a rule no benchmark scores: a backchannel never confirms a changed amount. In T01, Ryan's "uh-huh, okay" after Hazel said she was noting two hundred fifteen dollars meant he was listening, not that he agreed to $215.00 instead of the $250.00 he had offered. During a read-back, an "uh-huh" while Hazel is still reading is a backchannel; the explicit yes that C12 requires comes after the closed question has ended. So one label feeds two decisions, whether to stop talking and whether the words count as an answer, and the second decision is gated on timing: words that overlap the question cannot answer it.
Polarity matters too. "No, no" during a read-back is a barge-in, though it has the same shape as "si si": two short repeated words. A classifier that learns "short and repeated means backchannel" lets a refusal through as attention.
One kind of barge-in carries a legal deadline. Steelyard collects its own debts in its own name, so the FDCPA and Regulation F largely do not bind it 6 7; it adopts their conduct rules as company policy, and state law, the TCPA and the CFPB's ban on unfair, deceptive or abusive practices still apply. The TCPA consent-revocation rule that takes effect on 25 November 2026 keeps the requirement to honor a revocation within a reasonable time of no more than ten business days 8. A caller who says "stop calling me" or "deje de llamarme" [stop calling me] over Hazel's speech has made a revocation request (Steelyard's C14, honored at once), and the monitor's job is the same as for any other barge-in: stop, and route the words to the intent check even though they overlapped agent audio. Steelyard keeps stop requests as their own slice in the barge-in suite, because a raised noise threshold that swallows one of them turns a turn-taking defect into a compliance one.
Why is recovery its own axis?
Stopping in time is half of handling a barge-in. The other half is changing course: the next agent turn must reflect what the caller said, and the right tool must run. In T02 even a 500 ms stop would have been followed by a payment request after "ya lo pague" [I already paid it], so recovery gets its own pass rate.
Current systems are weak here. In the EchoChain benchmark (Modi et al., arXiv preprint, April 2026), four real-time voice models (GPT-realtime-2025-08-28, Gemini Live 2.5 flash native audio, Grok Voice Agent, Amazon Nova Sonic 2) had to revise their answer after a scripted user barge-in mid-response. Across 200 interrupted conversations none fully passed more than half; Grok Voice Agent did best at 48.5% and Gemini Live worst at 16.5% 9. Those are speech-to-speech systems on scripted interruptions, not Hazel's cascade, but they show that stopping and revising are separate skills.
Steelyard runs its recovery check on every scenario card that has an interruption script (triggered by content, not by a timer) and on flagged production barge-ins. A barge-in passes only when all four hold:
- History truncation. The conversation history the LLM sees holds only the words that actually played before
tts_stop, not the full planned sentence. Otherwise the model believes Sofia heard the whole balance statement, and later turns build on that.
- No resumption of the interrupted plan. The next agent turn does not repeat or finish the interrupted content unless the caller asks for it.
- The new information is reflected. The next turn responds to the claim in substance ("Un momento, lo estoy revisando." [One moment, I'm checking it.]), not with a filler followed by the old plan ("Si, entiendo. Entonces, puede hacer el pago hoy?" [Yes, I understand. So, can you make the payment today?]).
- The required tool runs, in order. For "already paid":
check_payment_status before any payment request; if it finds the payment, the Customer Advocate details, send_disclosure_sms and end_call(paid); if it finds nothing, log_dispute(already_paid), then the Customer Advocate details and send_disclosure_sms before the call ends. That is the F02 sequence, enforced by policy check C08.
Items 3 and 4 are trajectory checks, scored on what the agent did after the event; why a final-state check misses them is covered in trajectory vs outcome. The CRM diff for F02 belongs to task success and outbound call outcomes, and generating callers who cut in on content, with noise, belongs to simulating callers.
Report stop latency, recovery and their joint pass rate. A release that halves stop latency while recovery stays flat makes the call sound better, not end better.
Do full-duplex models fix turn-taking?
No. A full-duplex model listens and speaks at once, which removes the cascade's fixed silence budget, but published benchmarks show the same trade-off in a new form: the systems that reply fastest most often take the floor during a user's pause. Report reply latency only with the takeover rate beside it.
Full-Duplex-Bench measures turn-taking latency from the end of the user's speech to the start of the model's reply, counted only when the model takes the turn. On the Candor set Moshi replies in 0.265 s and dGSLM in 0.352 s, against 0.953 s for Freeze-Omni and 1.301 s for Gemini Live. Moshi also takes over during user pauses almost every time (takeover rate 0.985 synthetic, 0.980 Candor), while Gemini Live takes over far less often (0.255 and 0.310) 10. These are 2024-2025 system versions, tested on English.
A September 2026 benchmark covering ten languages spoken in India found the same trade-off, and its authors write that no system they evaluated escaped it 11; the India edition works through those numbers. We found no comparable full-duplex result on Spanish or Spanglish calls; those two groups hold about a third of Steelyard's accounts by language preference.
A full-duplex model moves the operating point out of a config file and into the weights. You cannot set 600 ms for the amount-asked state; you can only measure what the model does there, and the labeled pause set, the overlap labels and the recovery suite become the only instruments that show the trade-off. What else changes if Hazel's cascade is replaced by the hazel-s2s-pilot speech-to-speech model, including which spans disappear, is covered in designing the voice evaluation platform.
What release gate would have caught both calls?
Three release gates, each computed per language group and call state: the false-cutoff rate at the campaign's budget on the labeled pause set, the stop-latency P95 on the barge-in suite, and the recovery pass rate on the same suite. Each carries an interval and is compared with the incumbent on the same items.
gate: turn-taking
gold_set: gold-v7 # 400 labeled pauses per language group; scenario cards with interruption scripts
detector: td-1.1 # threshold and budgets from the campaign table, pinned in the trace
segments: [english, spanish-dominant, spanglish]
states: [default, amount-asked, date-asked]
metrics:
false_cutoff_rate:
at: campaign_budget # per campaign and call state
weights: {small_talk: 1.0, dispute: 1.5, amount: 2.5, date: 2.5}
interval: call-clustered bootstrap, 95%
stop_latency_p95:
events: barge-in labels only # stop requests reported as their own slice
slices: [quiet, tv, street]
interval: call-clustered bootstrap, 95%
limit_ms: 500
recovery_pass_rate:
checks: [history_truncated, no_resume, info_reflected, required_tool_in_order]
interval: wilson # one scenario, one pass or fail
compare: mcnemar # candidate vs incumbent on the same scenarios
report_always: [talk_over_rate, early_jump_in_rate, stop_rate_by_label, dead_air_mean]
block_if:
- false_cutoff_upper_bound > campaign_limit
- stop_latency_p95_upper_bound > 500
- recovery_pass_rate_lower_bound < campaign_floor
- paired_difference_worse_than_margin
The interval rules follow from the unit being counted. Pauses inside one call are correlated, so the false-cutoff rate gets a bootstrap that resamples whole calls; a Wilson interval over pauses would be too narrow. Stop latency is a percentile, so it is call-clustered too. Recovery is one pass or fail per scenario, so a segment's pass rate gets a Wilson interval, and the candidate-versus-incumbent comparison is McNemar's test on the same scenarios. At nightly sizes the intervals are wide: with 20 scenarios in a segment, even 20 of 20 passing gives a Wilson 95% lower bound of 0.839. Pass rates over repeated runs, and how many runs a gate needs, are in reliability statistics for voice agents.
Both calls fail these gates. T05's pattern, a long hold before a date, is exactly the slice the pause set oversamples; at AUTO's old 300 ms budget the amount-asked and date-asked states fail the false-cutoff gate before release, and the talk-over check points at the missing cancel-on-resume. T02 fails twice on a suite with a TV-noise slice: stop latency of 900 ms against the 500 ms limit, and recovery item 4, because check_payment_status never ran. In production the F02 tool-sequence check, which runs on 100% of calls, fails T02 the moment the call ends. Targets for the rest of a voice agent's metrics are in the voice task-type playbook.
What goes in the decision memo?
The memo records one endpointing operating point per campaign and call state, the cost weights that picked it and who signed them, and the tripwire that reopens the decision. Steelyard's memo after Ryan's call:
Decision. AUTO-early-nov moves from 300 ms at td-1.1 thr 0.62 to 400 ms at thr 0.70 in the default state and 600 ms in the amount-asked and date-asked states. PIL and POS run 400 / 600 ms at thr 0.70, SBL 450 / 650 ms at thr 0.72, SDL 600 / 800 ms at thr 0.78. Playback is canceled when caller speech resumes before play_start.
Context. T05 (16 November): a false cutoff on a 720 ms hold before a date, on AUTO-early-nov, the highest-volume campaign.
Options considered. (1) One budget for every campaign: wrong for SDL's older borrowers and for the amount-asked state. (2) Per campaign only: a compromise that is wrong for both states. (3) Per campaign and call state: chosen. (4) A full-duplex model: no evidence on Spanish or Spanglish calls.
Evidence. gold-v7 labeled pauses (400 per language group), cost per 1,000 turns with weights 1.0 / 1.5 / 2.5, call-clustered bootstrap intervals; the recipe's synthetic run places AUTO's default choice between 350 and 450 ms.
Owner and cadence. The turn-taking owner on the eval platform team; each campaign owner signs the weights. Rerun on every detector, STT or SDK version.
Tripwires. Spanish-dominant false-cutoff rate above 4% at the chosen operating point for two weeks; stop-latency P95 above 500 ms on any noise slice.
Dated facts. eot-bench README (19 November 2026); Full-Duplex-Bench v1.5 (April 2026); EchoChain (April 2026); TCPA revocation rule effective 25 November 2026.
What to do this week
Label a small pause set from your own calls, draw your detector's curve on it, split your interruption metric three ways, and score recovery apart from stop latency.
- Pull 30 calls per language group you serve and label every caller pause of 100 ms or more as end or hold, with call state and a content tag; keep the call id on each row.
- Plot false cutoffs against dead air for your current detector on those pauses, and mark where your production setting sits on each curve.
- Replace your single "interruptions" chart with talk-over, early jump-in and missed barge-in, using the formulas above on your own trace fields.
- Write five interruption scenarios in which the caller says something that must change the plan ("I already paid", "wrong number", "I lost my job", "stop calling me", and one in Spanish if you serve Spanish speakers), and score stop latency and recovery separately.
- Confirm that your agent's conversation history is cut to the words that actually played before a barge-in.
NOTE
What will date on this page. The eot-bench figures come from its README as of 19 November 2026 (v1.1 released 13 November 2026); rows are added as detectors arrive, so recheck them and cite the commit date. The June 2026 English figures are from LiveKit's blog. Full-Duplex-Bench v1.5 is arXiv v4 (April 2026) and tested mid-2025 endpoints; Full-Duplex-Bench is v3 (August 2025); IndicFDB is v1 (September 2026); EchoChain is an April 2026 preprint. The TCPA consent-revocation rule takes effect on 25 November 2026, after this page's date. eot-bench's Spanish results were not checked for this page, and our finding of no full-duplex result on Spanish or Spanglish calls is as of 20 November 2026.