78/78
scoring cells matched exactly to the canonical engine (single-subject core)
0
variance across repeated runs (deterministic output)
5
landmarks per face on a room read, enough for a count and a coarse orientation. The 468-point mesh runs on a single-subject read, not on the room.
60 min
the sweep ceiling for footage no analysis is reading. A finished run has its footage deleted the moment it finishes. Footage the engine is still reading is kept until that run ends.
What you can rely on today
Five things are true right now, each stated at the strength the evidence supports.
A reliable, reproducible engine
The shipped scoring core matches the canonical engine on 78 of 78 cells, exactly. It returns identical output on repeated runs. That parity check covers the single-subject core, not the group aggregate step on top of it. Same footage in, same read out, every time.
We measure the room, not the person
The read is room-level. It reports attention, energy, and synchrony for the group as a whole. Synchrony is how much of the room reacts at the same moment. No individual is ever scored. The engine infers no emotion. It reads behaviour and geometry, never feelings.
Private by construction
Your footage is processed, then deleted as soon as the analysis finishes. No faces are kept. Only the aggregate numbers survive: group totals, with nothing tied to a person. There is no individual record, because none is ever created.
Grounded in the science
The attention, energy, and synchrony signals rest on published science. We did not invent them. A room read measures in a coarse way, on purpose. It finds the faces it can, counts them, and works out which way each one points. It reads a few points per face, not a fine facial mesh. At room distance a fine read would be noise. Grounding in science is not proof of outcome, and we do not claim it is.
In step with the regulators
The engine infers no emotion. The EU AI Act holds that emotion recognition lacks a scientific basis. We take the same position. Any emotion-framed output fails closed, which means the engine returns nothing rather than guess.
Held to an honest bar
Proof of Impact measures how a room responds. We hold the word impact to a pre-registered validation program rather than claim it before it is earned, and we publish the full record in the open, failures and all. That is the point of the name: proof you can check.