The Verification Venue · a reflex you can break on purpose
The Touch Your Brain Saw Coming
Try to tickle yourself. You can't — and the usual story (your own touch is somehow "dulled") is wrong. Your cerebellum predicts the exact touch a fraction of a second before it lands, and subtracts the prediction from what arrives. Drag the prediction out of register and the touch comes back to life.
In 1998 and 1999, Blakemore, Frith and Wolpert built the experiment that pins this down. A subject moved one robot with the left hand; a second robot reproduced that movement as a foam touch on the right palm. Between the two they could insert a delay or a rotation — so the touch you commanded and the touch you got no longer matched. The instrument below is that experiment as a forward model: a moving touch, the brain's prediction of it, the real touch, and the error between them — every number live.
Prediction error E (rms)
0.00
E = 1 − cos θ · cos(ωτ)
Feels ticklish
0.0 / 10
verdict: can't tickle yourself
Attenuation
100%
how much the brain cancels
The 1999 paper tested 0, 100, 200, 300 ms. By ~200 ms the self-touch was as ticklish as an external one.
The paper tested 0, 30, 60, 90°. At 90° the twisted self-touch again felt external.
When the delay and rotation are both zero, the predicted trace sits exactly on the real one: the error is zero, the brain cancels the whole sensation, and you can't tickle yourself. Push either knob and the traces peel apart — the error vector opens up, cancellation fails, and the touch starts to tickle. The geometry is exact; the rating is a model of the paper's curve.
The check — every number recomputed in front of you
The touch moves sinusoidally at f = 2 Hz (the paper's value), so ω = 2πf = 12.566 rad/s. The mean-square error between the commanded (predicted) touch and the delayed-and-rotated real one has a clean closed form:
For your current settings:
The perceived-tickliness model (0–10 scale) is anchored to the paper's three results — it rises monotonically and reaches the external level by 200 ms delay or 90° rotation, either knob alone enough:
| condition | delay | rotation | tickly (0–10) | paper |
|---|
The animation is the exact instantaneous prediction error of the 2-D sinusoid; the closed form above is verified term-by-term against a brute-force average over a cycle. Run it: node research/the-touch-your-brain-saw-coming/verify-the-touch-your-brain-saw-coming.mjs
What's idealised here, and what's exactly true
Exactly true (and checked). The residual E = 1 − cos θ·cos(ωτ) is the exact normalised mean-square distance between a commanded sinusoidal touch and the same touch delayed by τ and rotated by θ — derived, not fitted, and confirmed against a numerical average. At θ = 0, τ = 0 it is exactly 0 (perfect prediction → full cancellation → you can't tickle yourself). The experiment's facts are quoted exactly: 16 subjects, a 2 Hz / 1.5 cm foam touch, delays of 0/100/200/300 ms and rotations of 0/30/60/90°, with tickliness rising significantly to a plateau at ~200 ms or ~90° that was statistically indistinguishable from an external touch.
A model, not a measurement. The 0–10 tickliness rating is our own monotone, saturating model of the paper's curve. The paper reports significance tests on ranked ratings (e.g. delay effect F = 24.93, p < 0.0005), not mean rating values, so we do not claim to reproduce specific data points — only the shape they established: starts near zero, rises, plateaus at the external level by the tested endpoints. The external level is set to 7/10 for legibility; the qualitative result is what's anchored.
Where the clean geometry breaks. The bare residual is periodic in delay: at τ = 250 ms the touch is exactly anti-phase (ωτ = π, E = 2, maximum), and at τ = 500 ms it has slid a full cycle and re-aligns (E → 0). Perception does not do this — once the delay exceeds the prediction window the touch simply feels external and stays that way. So the tickliness model is capped at the experiment's tested 0–300 ms range and does not follow the raw geometry's rebound. (The verifier checks both the periodicity of E and the monotone cap of T, so the gap is shown, not hidden.)
The mechanism, honestly. "Forward model in the cerebellum" is the dominant account, not the only one. The fMRI localisation (less somatosensory and cerebellar activity for self-produced touch) is correlational; the robot delay/rotation study is small-N (16); and individual ticklishness varies a lot. What is robust is the central finding the instrument is built on: a self-produced touch is attenuated, and breaking the prediction in time or space brings the tickle back.