Kṣipta from the jaw

Delve used to print a flag beside your reading. It fired when relative gamma passed 0.12, and it read:

Gamma Surge — Ajña/Sahasrāra activation: multi-sensory binding, peak insight, Spanda (divine pulse).

On one real sitting recorded through a real headband it fired on 36 of 65 epochs. It told a wearer they were having peak insight on more than half of their meditation. On 140 epochs of clean, wet-electrode laboratory EEG from five ordinary adults who were not meditating at all, it fired on 29 — one subject in five had a median relative gamma above the trigger.

It has been removed. Not re-tuned: removed, because the defect was the claim and not the level, and there is no value of relative gamma at these electrode sites from which insight follows.

A Muse records at TP9, AF7, AF8 and TP10 — two behind the ears and two on the brow. For a claim about high-frequency activity that is the single worst montage on the head, because those are the places the face’s muscles are.

AF7 and AF8 sit over frontalis, the muscle that raises the brow, whose electrical activity peaks at 30–40 Hz — inside the band the classifier calls gamma. TP9 and TP10 sit over temporalis, which spans 40–80 Hz. Masseter, the jaw, runs 50–60. Muscle at the temporal sites reaches around 100 µV against under 1 µV for genuine high-frequency neural activity: roughly a hundred to one in favour of the artefact.

The cleanest measurement of the size of this comes from neuromuscular blockade: with the muscles chemically paralysed, scalp power above 20 Hz falls between ten- and two-hundred-fold, while everything below 20 Hz is relatively spared. Restated at the frequency that matters, roughly eighty percent of scalp 40 Hz power in an awake adult is muscle rather than brain — and contamination rises with effort. Meditation is effort.

The strongest single result is from inside the meditation literature itself: in one dataset, midline and frontal gamma correlated with muscle components at r = 0.55–0.61 and was discarded by the authors as muscle, while occipital gamma correlated with muscle at r ≈ 0.11–0.17 and survived. Where genuine meditation gamma exists, it appears to be posterior. A Muse has no posterior electrode.

Consumer headbands have been compared against research amplifiers directly. Against a 64-channel laboratory system, a Muse tracks alpha at r = .87 and beta at r = .84 — genuinely good, for a band you can put on in four seconds. For gamma the same study reports r = .17, confidence interval −0.13 to 0.50, not significant, and its authors’ own recommendation is to use power below 30 Hz.

• alpha r = .87
• beta r = .84
• gamma r = .17, n.s.

It is worse than a weak correlation: the published comparisons do not agree on the direction of the error. One 2024 study finds Muse power higher than research-grade in every band, the gap widening into gamma, and had to discard the temporal electrodes outright on 17 of 19 recordings while the device’s own quality indicator showed green. Another finds Muse gamma lower.

A refusal is the strongest claim a classifier can make about an epoch. Delve’s used to rest on this number.

Three things, all of them one-directional. Gamma enters the score one-sided, so an epoch can be penalised for having too much of it and can never be rewarded for having little: the user is not clenching is not evidence of one-pointedness, and treating it as such was the inversion the old model made. Above 0.174 — the largest relative gamma the profile table attributes to any measured state — it discounts confidence in every state, continuously, on a floor that never reaches zero. And it no longer refuses an epoch at all; a derived distance threshold does that job on grounds that can be recomputed.

One flag replaced the three that went, and it is deliberately not a state:

Muscle Tension — gamma at the brow and temple sensors is jaw and brow muscle, not a state marker. Relax the jaw and brow.

It fires above 0.339, three standard deviations past the highest gamma any measured state carries, and imports that threshold from the classifier rather than retyping it, so the flag and the confidence ramp cannot drift apart. Measured: 2 of 140 clean laboratory epochs, 3 of 65 of the real sitting, 569 of 615 of the known-broken stored sessions. Quiet on signal, loud on noise — the opposite of what it replaced.

If Muscle Tension appears, it is about your face and not your mind. Unclench the jaw, let the teeth part slightly, drop the brow, soften the eyes behind the lids. Swallowing, frowning at a thought and straining to concentrate all show up here, and none of them is a failure of practice — but all of them make the four seconds they land in harder to read.

The reverse does not hold. A quiet gamma band is not evidence of depth, and Delve will never tell you it is. If you came to this product hoping for a gamma number to chase, that number is not available here, and the reason is that it was never measuring what it was being read as measuring.