xevall

01 · the premise

Build voice and vision products people love.

Understanding is the new interface.

Does it catch the tone, the timing, the glance and the gesture, or just the words? xevall is the independent evals and self-improvement layer for human-AI interaction.

system interpretationintent · continueindependent grade · 0.92
human signal system interpretation independent grade
fig. 01a · a person is observed, the system is graded
  • processing · on this device
  • media transmitted · 0
  • system graded · never the person

02 · natural inputs

Interfaces are losing their keyboards.

What replaces typing is you: speech, tone, gaze, gesture. Products built around natural input enter the examiner's field. Their understanding is what gets tested.

before · explicit input
shiftcommandspaceenter
the person adapts to the interface
the system receives more of the meaning
  • gaze
  • blink
fig. 02a · gaze channel
fig. 02a · gaze · eyes that look back
  • speech
  • tone
fig. 02b · speech channel
fig. 02b · speech · lips mid-sentence
  • gesture
fig. 02c · gesture channel
fig. 02c · gesture · a hand saying hello
channel 01speechthe words that were spoken
channel 02tonehow those words were delivered
channel 03gazewhere attention was directed
channel 04gesturewhat movement added to the meaning
fig. 02 · natural input gives a system more to understandsignal field · shared scene

03 · system misfires

Nobody measures the understanding.

When an AI system misunderstands a person, the product feels broken and the reason stays hidden. xevall grades the system against the meaning a human reviewer established, and hands back the moment it went wrong.

sessionhuman meaningsystem actiongrade
kiosk 04:12a pause, not a yestook it as a yesmisfire · consent
agent 00:31still mid-thoughtjumped inmisfire · timing
telehealth 11:47said fine, meant not finenoticed, and askedunderstood · 0.92
signal comparison · 04:12divergence begins · 00:07
  • intended
  • inferred
fig. 03 · interpretation diverging
intended signal · stable inferred representation · displaced
fig. 03 · the system departs from the established meaningmisfire marked

04 · local demonstration

The demonstration is optional. Camera and microphone permissions stay separate, each can be declined and each can be revoked at any time.

No camera frame or audio sample is transmitted or stored. Media-derived values exist only as ephemeral on-device state, are never transmitted or sent to analytics, and are cleared on revoke or unmount.

optional demonstrationprocessing · on device

session · awaiting choice

camera · idle

Visual signals are processed locally.

microphone · idle

Voice level is processed locally.

local signal instrumentsession · awaiting choice
yaw0.0°pitch0.0°attentionunknownvoicegesturenone

Mapping: the mesh traces your face, raised hands are traced and named, attention and voice set the readings. Media stays on this device.

both paths · equal evidence · equal craft

fig. 04 · the same local instrument for live or recorded input

05 · the improvement loop

Every interaction creates traces to improve your product.

You already eval the words. Nobody is evaluating the coherence between the modalities: the tone, the distance, the body language. xevall grades whole interactions against human judgement, marks where the understanding broke, and re-scores the same sessions once you have fixed it. The grading exists to make products more human, not to mint a number.

  • agreement 0..1
fig. 05 · evidence returning
illustrative system result92xevall score · understood+7 since the last build

the Human Input Benchmark

The Human Input Benchmark goes public this autumn.

Building a product around voice, timing, gaze or gesture? Early access is open for teams testing natural interfaces.

Explore the benchmarkAsk about early access

xevall ·the independent evals and self‑improvement layer for human‑AI interaction.

xevall
aligning signalaxis checkfigure expected