# Signals (https://orgasm.dev/docs/concepts/signals)

orgasm.dev sends two kinds of signal. **Imminence** looks ahead: how likely orgasm is to begin soon. **Orgasm calls** look back: an onset, a peak or an end that has just happened. From the two, orgasm.dev also counts each **edge**, an approach that receded. A live session also describes what the voice, the body and the face are doing, for a **display**. Your product decides what to do with any of them; [presets](/docs/concepts/presets) are worked examples.

**A woman's orgasm: onset, then end** (sandbox scenario `woman-onset-end`). The face reads as a woman’s after about 30 seconds, imminence climbs over nine minutes, and the onset and end of one orgasm are called a few seconds after each happens.

Imminence each minute:

| streamTimeS | p30 | p60 |
| --- | --- | --- |
| 62 | 0.019 | 0.037 |
| 122 | 0.025 | 0.049 |
| 182 | 0.029 | 0.058 |
| 242 | 0.048 | 0.094 |
| 302 | 0.052 | 0.102 |
| 362 | 0.117 | 0.22 |
| 422 | 0.14 | 0.261 |
| 482 | 0.322 | 0.541 |
| 542 | 0.868 | 0.983 |
| 602 | 0.087 | 0.167 |

Orgasm calls:

| type | data |
| --- | --- |
| `orgasm.onset` | orgasm 1 at 538 s, lag 4.6 s, score 0.71 |
| `orgasm.end` | orgasm 1 at 575 s, lasted 37 s, onset_end, lag 5.2 s |

## Imminence [#imminence]

`imminence.updated` arrives about once a second while the face is analysed.

| Field | Meaning |
| --- | --- |
| `p30` | Probability that the next orgasm begins within 30 seconds, given that it has not begun yet |
| `p60` | The same over 60 seconds; never lower than `p30` |
| `logit30`, `logit60` | The same probabilities as logits, to combine with a prior of your own |
| `featuresPresent` | Share of the model's inputs available this second |
| `faceVisible`, `baselineReady` | Whether the reading rests on a tracked face and on the person's own baseline |
| `stimulationAware` | Whether it accounts for time under stimulation you [reported](/docs/api/live-sessions/reportStimulation) |
| `recentCalls` | The session's last orgasm calls, repeated so a missed call is seen within a second |

"Begins" means the onset for a woman and the peak for a man.

Read it as a probability over a horizon, **not a countdown**. A `p30` of 0.6 does not mean "in 18 seconds"; it means the model estimates that orgasm is more likely than not to begin within the next 30 seconds. Compare it with thresholds you tune for your experience and for the person, rather than reading the number as exact. The reading is `null` until the person's resting baseline is in, about a minute after their face first appears, and whenever the face is out of view.

Imminence from the face alone does not know how long the person has been stimulated. If your product controls stimulation, report it with `reportStimulation`, and imminence takes time under stimulation into account.

Imminence is currently more informative for men than for women. For women, combine it with onset and end.

## Orgasm calls [#orgasm-calls]

| Event | Called by | Applies to | Needs |
| --- | --- | --- | --- |
| `orgasm.onset` | `onset_end` | Women | Face and microphone |
| `orgasm.end` | `onset_end` | Women | Face and microphone |
| `orgasm.peak` | `peak` | Men | Face and microphone |
| `orgasm.end` | `face_end` | Men, after a peak | Face |

Each call names a moment (`eventAtS`) and says when the detector was sure enough to call it (`calledAtS`). The difference, `lagS`, is typically a few seconds: a call tells you what happened, not what is about to. To act at a moment, act on imminence; to confirm what happened, act on calls.

`orgasmIndex` counts orgasms in the session from 1; the onset and end of the same orgasm share it, as do a man's peak and the end read from his face. `score` and `threshold` are the detector's own: a score at or above the threshold made the call.

`orgasm.end` also says whether the orgasm was `ruined`: your product reported stimulation `off` at the point of no return (`p30` at 0.75 or above) within the 30 seconds before it began. A recording takes no stimulation reports, so its orgasms are never ruined.

**A man's orgasm: peak, then end** (sandbox scenario `man-peak`). The face reads as a man’s, imminence climbs to a peak, which is called about three seconds after it happens; the end is then read from his face.

Imminence each minute:

| streamTimeS | p30 | p60 |
| --- | --- | --- |
| 62 | 0.01 | 0.02 |
| 122 | 0.017 | 0.033 |
| 182 | 0.025 | 0.05 |
| 242 | 0.043 | 0.084 |
| 302 | 0.074 | 0.142 |
| 362 | 0.086 | 0.165 |
| 422 | 0.125 | 0.234 |
| 482 | 0.206 | 0.369 |
| 542 | 0.285 | 0.488 |
| 602 | 0.467 | 0.716 |
| 662 | 0.667 | 0.889 |
| 722 | 0.318 | 0.535 |

Orgasm calls:

| type | data |
| --- | --- |
| `orgasm.peak` | orgasm 1 at 701 s, lag 3.2 s, score 0.66 |
| `orgasm.end` | orgasm 1 at 742 s, lasted 41 s, face_end, lag 3.5 s |

## Edges [#edges]

`edge.completed` arrives when an approach recedes: `p30` reached 0.6 and fell back below 0.3 with no orgasm called in between, the [edge preset's](/docs/concepts/presets#edge) back-off and resume levels. orgasm.dev counts it on its side, numbers it (`edgeIndex`), and says when the approach started, how high it went (`peakP30`) and how long it lasted. Without the microphone no orgasm can be called, so every approach that recedes is an edge. In live mode an edge is charged half a credit ([Credits and billing](/docs/credits)).

**Edge: three approaches, no orgasm** (sandbox scenario `edge`). Imminence rises past the hold and back-off levels three times and falls each time stimulation is backed off, and the session ends without an orgasm: three edges, each an edge.completed.

Imminence each minute:

| streamTimeS | p30 | p60 |
| --- | --- | --- |
| 62 | 0.02 | 0.039 |
| 122 | 0.039 | 0.077 |
| 182 | 0.083 | 0.158 |
| 242 | 0.143 | 0.265 |
| 302 | 0.29 | 0.496 |
| 362 | 0.605 | 0.844 |
| 422 | 0.171 | 0.312 |
| 482 | 0.145 | 0.268 |
| 542 | 0.526 | 0.776 |
| 602 | 0.26 | 0.452 |
| 662 | 0.116 | 0.219 |
| 722 | 0.389 | 0.627 |
| 782 | 0.396 | 0.635 |
| 842 | 0.133 | 0.249 |

## Signals for a display [#signals-for-a-display]

`signals.updated` arrives with each imminence reading of a live session. It describes what the voice, the body and the face are doing, so a heads-up display can show more than a probability. None of it is a call, and none of it is charged. It is only streamed: webhooks don't carry it, and recordings don't have it.

| Field | Meaning | Needs |
| --- | --- | --- |
| `vocalizationsPerMin` | Vocalizations a minute | Microphone |
| `loudnessAboveBaselineDb` | How much louder the voice is than at the person's baseline, in decibels | Microphone |
| `contractionsPerMin` | Pelvic floor contractions a minute | Body view |
| `movement` | How still the body is: `settled`, `shifting` or `unsettled` | Body view |
| `faceRegions` | How far the `brows`, `eyes`, `cheeks`, `nose`, `lips`, mouth `corners` and `jaw` have each moved from the person's own resting state, from 0 to 1: each the largest of its expression channels | Face |
| `face` | The face's 24 expression channels (`expression`, each 0 at the person's rest to 1 for a full movement) and how far the head has turned from its resting pose (`head`: `yawDeg`, `pitchDeg`, `rollDeg`, as someone facing them sees it), for a display that draws the face | Face |

Each field is `null` while its input is absent or still calibrating, so show only the fields that have a value. To act, act on imminence and the calls; these describe the moment.

The channels say how the face moves, not what it looks like: drawn on a generic face, as the [HUD's](/docs/guides/hud) `face` part does, they show the person's expression without their features. A display can read them five times a second as [face frames](/docs/guides/realtime#face-frames), which also carry the head as the camera sees it, from before the baseline. A view token reads them only when made with `includeFace`.

## The device's state [#the-devices-state]

`stimulation.updated` is not a signal of the person's: it is what your product [reported](/docs/guides/realtime#reporting-the-device) about the device it drives (on or off, its level and channels, its mode, who sets the level), merged across reports and sent to every reader, so a display can show it beside the signals. It is only streamed.

## The subject read [#the-subject-read]

`subject.identified` arrives once, about 30 seconds after the face first appears: whether it reads as a woman's or a man's, or `undetermined`. It decides which detectors apply. [Subject](/docs/concepts/subject) explains how, and how to treat it.

## Confidence [#confidence]

Every signal carries `confidence`: `high`, `medium` or `low`, from how complete its input was. `high` means the face was tracked throughout and every input the signal uses was present; `low` means degraded input, or a detector that ran while the subject was undetermined. It describes the input, not a calibrated probability that the call is right.

## What signals are not [#what-signals-are-not]

- **Not proof.** Detectors make false calls and miss real ones. Do not treat a call as evidence about a person, and do not attach consequences to one that the person has not agreed to in advance.
- **Not a medical measurement.** Nothing here diagnoses or monitors a condition.
- **Not identity.** No signal identifies who a person is, and the subject read is not a statement about anyone's sex or gender.
