A silent assistant listens along and may show one card on the phone. Every card is scored against a hand-labelled need window with the real wall clock: transcript lag, gate, model latency, push. A card that arrives after the need is gone counts against the model.
1 · Speech to text
+15 s
transcript lag (streaming ASR)
→
2 · Gate
0.3 s
jev-latest (TypeSafe.ai), every 30 s on the last 90 s
→
3 · Card model
10 to 409 s
only when the gate fires
→
4 · Push
+5 s
phone out of the pocket
=
Speech to visible card
30 to 429 s
anything a person finishes faster is out of reach
Recordings
| Recording | Status | Gate ticks | Needs | Best model | Decisions | |
|---|---|---|---|---|---|---|
◈ Pokimane-Mafia-Game-1Processed on 2026-09-24 from Pokimane-Mafia-Game-1.mp4. continuous mode, 899 turns. | Not evaluated | – | – | – | – | |
◈ Recording, Sep 24, 2026 08:46A group of male streamers/friends are out in a city at night, having just been kicked out of what appears to be a sugar or chocolate museum, and they're walking around looking for someone named Lacey while fans approach them for photos. They discuss going back to talk to the museum manager about it being good promo, and mention that their stream was being watched by people in the venue. |
Evaluated | 4 of 15 | 0 long · 0 short | deepseek-v4-flash-07310 of 0 in time | 2 cards · 2 silent | |
◈ Los Angeles day, Sep 21, 2026A whole day on one phone: hotel, Uber rides, Griffith Observatory, Beverly Hills Hotel, airport, Delta flight to SF. 3.5 h of speech in 15.5 h, German and English, many strangers. | Not evaluated | – | – | – | – | |
◈ Evening (Cafeteria), Sep 20, 2026Late evening in a cafeteria, loud surroundings, German. |
Evaluated | 24 of 149 | 5 long · 3 short | gpt-5.6-sol3 of 5 in time | 64 cards · 80 silent | |
◈ Evening (Cafeteria) — Stan's pipelineThe same evening recording, processed with Stan's pipeline. Speakers are only distinguished within each section. | Not evaluated | – | – | – | – | |
◈ Tenten Founder Meeting, Sep 20, 2026Office conversation, mostly English. Four known voices and one new person. | Not evaluated | – | – | – | – | |
◈ Recording, Sep 18, 2026 20:53Team-Sync mehrerer Entwickler (Speaker 1–5), in dem es um die Inbetriebnahme einer Always-On-Voice-Capture-Funktion, die Camera-Ready-Fassung eines angenommenen Papers sowie die Planung des App-Store-Releases einer neuen App (inkl. Expedited-Review-Option bei Apple, UI-Anpassung wegen Copyright-Bedenken und Sequencing gegenüber dem laufenden Projekt QP) ging. Abschließend wurde ein mehrstufiges Gating-/Eval-Konzept für die App besprochen. |
Evaluated | 4 of 16 | 0 long · 0 short | gpt-5.6-sol0 of 0 in time | 2 cards · 2 silent | |
◈ Recording, Sep 18, 2026 21:09Eine Gruppe Tech-Gründer hängt in einem gemeinsamen Workspace ab und spricht über ein teures Shopping-Erlebnis, App-Entwicklung (Flutter vs. Native, AI-Prompting für Design) sowie Apps wie NGL und Duolingo. Nebenbei arbeiten sie an ihrer eigenen App „Sunny" und diskutieren Namensfindung, UI-Flows und ihren Coworking-Alltag. |
Evaluated | 8 of 75 | 0 long · 0 short | gpt-5.6-sol0 of 0 in time | 5 cards · 3 silent | |
◈ Dinner (Indian), Sep 17, 2026Five people at the table. German with English mixed in, lots of talking over each other. | Not evaluated | – | – | – | – | |
◈ Marlon 10 min (gemini, nemotron)Processed on 2026-09-24 from uploads/j-20260924-022419-eyqm/marlon_10min_2026-09-15.mp3. continuous mode, 169 turns. | Not evaluated | – | – | – | – | |
◈ Marlon 10 min (gemini, precision3)Processed on 2026-09-24 from uploads/j-20260924-022419-eyqm/marlon_10min_2026-09-15.mp3. continuous mode, 219 turns. | Not evaluated | – | – | – | – | |
◈ Marlon 10 min (gemini, window judge)Processed on 2026-09-24 from uploads/j-20260924-022419-eyqm/marlon_10min_2026-09-15.mp3. continuous mode, 233 turns. | Not evaluated | – | – | – | – | |
◈ Marlon 10 min (live test 2)Processed on 2026-09-24 from uploads/j-20260924-022419-eyqm/marlon_10min_2026-09-15.mp3. continuous mode, 266 turns. | Not evaluated | – | – | – | – | |
◈ Marlon 10 min (live test)Processed on 2026-09-24 from uploads/j-20260924-022419-eyqm/marlon_10min_2026-09-15.mp3. continuous mode, 266 turns. | Not evaluated | – | – | – | – | |
◈ Marlon 10 min (nemotron)Processed on 2026-09-24 from uploads/j-20260924-022419-eyqm/marlon_10min_2026-09-15.mp3. continuous mode, 177 turns. | Not evaluated | – | – | – | – | |
◈ Marlon 10 min (precision3)Processed on 2026-09-24 from uploads/j-20260924-022419-eyqm/marlon_10min_2026-09-15.mp3. continuous mode, 258 turns. | Not evaluated | – | – | – | – | |
◈ Marlon 10 min (qwen, window judge)Processing (speakers from diarization, text from alignment; the per-turn judge is still running). | Not evaluated | – | – | – | – | |
◈ Marlon x Lacy x Jason in New York (hour 4), Sep 15, 2026One hour cut from Marlon's IRL livestream in New York City (3:00-4:00 of the VOD): kicked out of the Museum of Sex, mobbed by fans, regrouping in a smoothie shop, deciding where to go next, an Uber ride with a talkative driver, Color Factory turns out closed. English, crowd noise, many strangers. |
Evaluated | 33 of 118 | 7 long · 2 short | gpt-5.6-sol5 of 7 in time | 133 cards · 65 silent | |
◈ Marlon NYC test (10 min)Processed on 2026-09-24 from marlon_10min_2026-09-15.mp3. continuous mode, 267 turns. | Not evaluated | – | – | – | – |
Models across evaluated recordings
| Model | Needs in time | Too late | No need | Repeat | Silent | Latency p50 | Latency max | Cost |
|---|---|---|---|---|---|---|---|---|
| gpt-5.6-lunaopenai | 9 / 12 | 20 | 3 | 6 | 11 / 57 | 8.1 s | 11.3 s | $0.168 |
| gpt-5.6-solopenai | 8 / 12 | 2 | 1 | 3 | 39 / 69 | 9.4 s | 26.3 s | $1.707 |
| glm-5.3-flashz-ai | 8 / 12 | 9 | 3 | 4 | 21 / 57 | 14.7 s | 408.7 s | $0.073 |
| kimi-k3moonshotai | 8 / 12 | 15 | 1 | 3 | 17 / 57 | 53.4 s | 221.0 s | $3.624 |
| deepseek-v4.1-flashdeepseek | 8 / 12 | 17 | 5 | 6 | 5 / 57 | 5.9 s | 103.8 s | $0.115 |
| gemini-3.8-flashgoogle | 2 / 12 | 0 | 0 | 0 | 55 / 57 | 9.5 s | 22.8 s | $0.658 |
| deepseek-v4-flash-0731deepseek | 0 / 0 | 0 | 0 | 0 | 0 / 4 | 6.0 s | 9.3 s | $0.001 |