“The blender smells of burning and has stopped working.”
01 / Choice
Which team should handle this?
Areturns0.56
Bshipping0.25
Cbilling0.12
Dtechnical0.07
Returns
P(returns) = 0.56A closed set. Every option receives a probability.
02 / Yes / no
Does the customer report a defect?
Ayes0.82
Bno0.18
A probability, directly.
P(yes) = 0.82Two options. The probability of yes is the readout.
03 / Ordered score
How urgent is this message?
Aroutine0.12
Bsoon0.45
Cimmediately0.43
Between soon and immediately.
E[level] = 1.31An expected level preserves uncertainty between options.
One pass / question
Illustrative values · not model output
The scalp illustrates one encoding and twenty decoder steps. The EEG view follows patches through spatial and temporal attention to a 14,000-value embedding. EEG and token values are synthetic. The decision view illustrates the V3 interface with authored probabilities and does not display V3 model predictions.
A returns: 0.998; B shipping: 0.001; C billing: 0.001.
Question · Yes or no
Does the customer report a defective product?
Recorded answerRecorded answer · T = 1
Yes, at 0.992.
A yes-or-no question is a two-option prompt; the answer is the probability of yes.
Probability by optionRe-weighted · T = 1.00Chance · 1/2
Ayes0.992
Bno0.008
P(yes)0.992
Response details
Returned0 output tokens
"defective": {"noul": 0.992}
A yes: 0.992; B no: 0.008.
Question · Score
How urgent is this message?
Recorded answerRecorded answer · T = 1
Between soon and immediately.
The two levels are 0.011 apart, so it returns the expected level instead of pretending to know.
Probability by optionRe-weighted · T = 1.00Chance · 1/3
Aroutine level 00.104
Bsoon level 10.454
Cimmediately level 20.443
expected level1.34confidence0.18
Response details
Returned0 output tokens
"urgency": {"score": 1.34, "confidence": 0.18}
Expected level
0routine1soon2immediately1.34
A routine: 0.104; B soon: 0.454; C immediately: 0.443.
Explore calibrationBrowser-side transform
0.51 · recorded23
A transform of the recorded response, not a model run: each probability becomes p1/T ÷ Σ p1/T, computed in your browser. Below 1 sharpens the distribution, above 1 flattens it; the order never changes. The model card’s readout uses T = 1.0, with no fitted calibration.
Archived V2.1 response · MLX 8-bit · Source & method. Replayed here; no model runs in your browser.View as table
Question
Option
Recorded probability
Which team? · Choice
A = returns
0.998457
B = shipping
0.000967
C = billing
0.000577
Is it defective? · Yes or no
A = yes
0.991509
B = no
0.008491
How urgent? · Score
A = routine
0.103718
B = soon
0.453758
C = immediately
0.442524
02 / Neural signals
EEG, ready for Apple.
Reconstruct missing channels with ZUNA. Extract representations and classify cue-paced motor imagery with CBraMod. Packaged in Core ML and checked against PyTorch.
ZUNA takes the channels that were recorded, with their 3D positions, and estimates the ones that were not by iterative denoising.
AF3F7F3FC5T7P7O1O2P8T8FC6F4F8AF4
5 cmTop view · t = 3.53 s
Denoising step 20/20
F3measured
F4estimated
T7estimated
P8measured
O2estimated
t = 3.53 s
012345 s
Seconds · synthetic, not recorded EEG
Measured
Not recorded
Estimated by the model
Synthetic scalp field: solid +, dashed −
The head is the FreeSurfer fsaverage average scalp, seen from above and drawn to scale, with the real EPOC X electrode positions (MNE standard_1005). Traces and scalp field are synthetic, drawn to show the task: not recorded EEG and not model output. The field is a spherical-spline interpolation at one instant, t = 3.53 s, chosen for illustration. The green discs around T7, F4 and O2 are a drawing convention for emphasis, not the model’s region of influence.
Every release includes its evidence, limitations, and upstream credits.
A new run from Qwen3.5-4B under the SemIf prompt, with public benchmark gains, a documented out-of-distribution regression, and full calibration limits.