TPlus
How sensing works
TPlus should notice launch moments the way a careful human does: many small cues, not one fragile detector. This page is the public sketch — what runs today, where we’re headed, and how collaborators can help.
Today
Spine
- Mission plan (Launch Library 2 → milestones)
- OCR on the webcast T+ / hold clock
- Emit when the clock hits a planned beat
Already weak
- Speech phrases as corroboration
- Light on-screen text heuristics
- Still frames as artifacts, not proof
Speech alone does not blast subscribers. The plan still gates public noise.
Next: a chain of weak indicators
Instead of “OCR or nothing,” treat the webcast as a bag of sensors that each cast a vote:
- Video classification — pad / plume / stage sep / landing scene, and similar
- Audio classification — roar, callout energy, engine character (alongside, not instead of, speech-to-text)
- More text / HUD readers — tighter than today’s heuristics
- Later: other attributable feeds, if they stay optional and honest
Fusion idea: several weak yeses near a planned milestone → confident alert; a lone flash → log it and stay quiet.
Contributing a sensor
We want people who love launch AV to plug in sensors, not rewrite Telegram fan-out.
Good fits
- Phrase / pattern packs mapped to milestones
- Small classifiers (frame or audio in → label + confidence)
- Offline scorers on saved webcasts
- Stream workers that emit evidence lines
Not the goal
- Posting straight to Telegram or the Worker
- Fighting OCR for clock ownership
- Scraping private or credentialed feeds
- Promising official telemetry
Sensors should emit evidence (what you think happened, when, how sure, which sensor). The desk decides whether fans see it.
The wire format isn’t frozen yet. Until a versioned evidence schema lands, the best start is an issue with modality, target milestones, sample clips, and how you’d run beside the live desk.
Open an issue on cue · Full SENSORS.md
Principles (short)
- Evidence in, alerts out — sensors don’t fan out alone
- One narrow clock; others propose what, carefully
- Optional: missing GPU/models must not kill the OCR path
- Footnote sources; never pretend to be the range