What this instrument can, and cannot, tell you.

What we measure

The engine reads four channels of behaviour from ordinary video. No wearables. No survey. Every number on a report traces back to a function in the code. For the room-level signals it also traces to a citation on the Methods page.

• Face. For one person: 468 landmarks per frame, from MediaPipe FaceMesh, giving expression strength, head pose, and blink and gaze proxies. A proxy is a stand-in, not a direct reading. A room read does not run this model. It gets the room-level signals below.
• Body. Pose keypoints on the same footage.
• Voice. Prosody: pitch, timing and voice quality. Only when there is audio.
• Words. Word features from the transcript, such as word variety and hedging. Only when there is speech.
• Room-level, used in Proof of Impact. Room-level means the room is read as one body, never a named person. Five signals: attention on the speaker, from head pose. Room synchrony, which is how far faces react together. Audible reaction, from peaks in the sound. Speaking balance. And engagement across the session.

What we do not measure

These are the limits. We would rather you learn them here than assume we can do more.

• No heart rate. No pulse, heart-rate variability or anything vagal is read from video. We do not do rPPG, which is reading a pulse from camera footage. We removed every earlier mention of a heart or rPPG channel from the product.
• No inner emotion. We do not read what a person feels, detect lying, or judge whether warmth is real or performed. A face signal describes the expression, not the feeling behind it.
• No per-person scoring in an audience read. Proof of Impact is aggregate-only. It reads the room as one body. It never scores a named person. There is no leaderboard.
• No clinical or diagnostic output. This is not a medical tool, a psychological test, or a hiring test.

Conditions that change a reading

A read is only as good as the footage. Some clips cannot carry one. The engine says so, rather than hand back a confident-looking number. These matter most:

• Face size. Faces too small in frame, like the far rows of a big room, fall below the floor for a reliable read. They are left out, not guessed.
• Lighting. Low or uneven light makes the landmarks less reliable. The engine scores image quality, then gives that window less weight or abstains.
• Motion. Heavy movement, big laughter included, degrades a window. That window can then be dropped.
• Eye openness. Gaze is timed only on frames where the eyes pass one fixed openness level. Narrow eyes, and eyes in a laugh or smile, fall below it more often. When too few frames pass, look-aways are withheld for that take. In our test of eight faces, this withheld look-aways for two of the six in view that a per-person level would have measured. We are moving to a per-person level. It is not yet tested across skin tones, ages, glasses, facial hair or head coverings.
• Footage length. Very short clips carry less signal, and the confidence shows it.
• Frame timing. The pipeline seeks by nominal time, which a variable frame rate can shift. Any claim tied to a clock time can drift as a result. Checking frame timing is on our hardening list.

Indices are a design choice, not a physical measurement

Several headline numbers are composite indices: blends of other measures, with weights the GRW team chose. They are not a physical unit, and they are not fitted to any outside outcome. They are useful, comparable and ours. We publish the weights so nothing is hidden.

One composite exists for audience engagement, worked out per time bucket. Its weights are attention 40%, positive expression 25%, audible reaction 20%, and crowd retention 15%. The aggregate Proof of Impact report does not roll these into an invented "Impact Index". It reports the strongest index we truly measured, by its real name. That is attention on the speaker, or room synchrony. When nothing is measurable, it abstains.

The bar below shows the Overall Score weights for an individual report. Presence Quotient, Play IQ and the rest are our own indices too. None of them has been tested against an outside outcome to date.

Congruence Signal is marked AT-RISK. It is a proxy for expression congruence: how well the parts of a facial expression line up. Its inputs are genuine-smile activation, low suppression, low leakage, and how well affect and engagement agree. It does NOT infer whether a person is genuine or performing. It will not carry a genuineness claim unless it is re-specified and validated as one.

What the confidence level means

Every score carries a confidence level. It measures precision: how self-consistent the reading is on its footage. It does not measure agreement with an outside ground truth. We have run no formal accuracy study against a reference standard. So we claim no validated accuracy, and we keep "a new standard of accuracy" language off the product.

When the footage cannot carry a claim, the engine withholds the number. It does not invent one. A withheld score is a feature. A number you cannot trust on hard footage is a number you cannot decide with.

Found something that does not match?

A claim or a number that does not match this page is a bug. Email info@boostinnovation.ca, and we will fix the product or the page.