Cyborg Collective · Trust

How far ahead can a headband be read?

A pre-registered test of anticipating a Muse headband's activation index 2 to 16 s ahead, under the gates the lab's papers use.

On the owner's own head and on the lab's cold-exposure recordings, the question was how much of the next 2 to 16 s is already in the last 16 s, and whether a local language model, shown only the aggregate tables, says anything about the data that the same gate lets stand. Everything ran on one laptop; the raw recordings never left it; the design was written down and committed before the first run.

Pre-registered in PREDICTIONS.md, with two amendments made before the first run. Rerun it yourself: code.zip, pointed at a folder of MuseLog recordings, rebuilds every number on this page in a few minutes on a laptop.

6ense™ anticipation · pooled skill against how far ahead, 2 to 16 s, three pools of recordings
Four panels of pooled forecast skill against horizon (2, 4, 8, 16 s) for three pools, within a session and across sessions, against persistence and against the session mean, with bootstrap intervals

Skill = 1 − MSE(model)/MSE(baseline), ridge on the last 16 s of the index, 95 % intervals from 2000 session bootstraps. Left: against "it stays where it is". Right: against the session's own mean. Top: trained on the first 60 % of a session, tested on the last 40 %. Bottom: trained on other sessions. Two cold points at 4 s are off the scale: one session's within-session model, the variant with band and motion features, extrapolated badly; the lags-only model did not. Light version · PDF

What the data allowed

A recording entered a pool only if it was at least 3 minutes long and yielded 60 clean, contiguous 18-second windows under the gate: contact 2 or better on the electrode, under 200 µV peak to peak, 90 % of the samples arrived, two of four electrodes accepted. The gate, not the analyst, decided who got in.

13 of 115recordings of 3 minutes or more passed the gate
51 minof the owner's head in play, 6 sessions, one afternoon
84 minfrom the cold-exposure periods, 6 sessions, no plunge
0.11 vs 0.45usable fraction of the cold-period recordings against the owner's
0recordings labelled before or after a plunge that passed
PoolRecordings ≥ 3 minUsable fraction, meanEligibleMinutes used
the owner's head140.45651
cold periods890.11684
all1150.1613225
One afternoonThe six eligible sessions of the owner's head are all from 2025-12-13, recorded while the recorder was being written. "My head" here means one day of it.
No plungeNo recording labelled as before or after a plunge reached 60 clean windows. The cold pool is six recordings from the cold-exposure periods, wearers unnamed and not the owner. Nothing here is a plunge response.
Coverage, not skill, separates cold from calm7 % of the cold-period recordings qualify against 43 % of the owner's. The pre-registered P7 compared usable fractions inside the eligible pools and is refuted as worded; at the recording level the gap is 0.11 against 0.45.

What was found

Seven things. Every number is in the tables below and in FORECAST.md.

0.29 to 0.46of persistence's error removed by ridge on 16 s of history, 2 to 16 s ahead, every pool, held out
0.25 to 0.42removed by the session's own mean alone, owner and cold pools
+0.03 to +0.16beyond the mean at 2 s, owner and cold pools; nothing left by 16 s
0.409 vs 0.400other heads' model on the owner's sessions, against his own
  1. The index can be anticipated 2 to 16 s ahead, relative to "it stays where it is". Ridge on the last eight values removes 0.29 to 0.46 of persistence's squared error in every pool, on held-out time and on held-out sessions, every interval clear of zero.
  2. Most of that is drift back to the session's level. The session mean by itself, used as the forecast, removes 0.25 to 0.42 of persistence's error in the owner's and the cold pools. The game's quantity wanders about a level and persistence chases the noise.
  3. Beyond the level, 16 s of history adds a little at 2 s and nothing by 16 s. A within-session permutation test, which shuffles the older history while every target keeps its own current value, says the past does carry information (p = 0.005 in all six primary cells, Holm-adjusted 0.030); the amount is a few hundredths to 0.16 of the error.
  4. Nothing personal at these horizons. A model trained on other people's sessions anticipates the owner's as well as one trained on his own (0.409 against 0.400).
  5. The small recurrent network lost to ridge everywhere. Fifty to two hundred minutes per pool is not enough for it.
  6. The pre-registered primary model was the fragile one. Ridge with band powers and motion added extrapolated badly on two January sessions when trained on a session's first 60 %; lags alone did not. Three predictions read as refuted for that reason; the lags-only numbers are the ones to trust, and both are published.
  7. Asleep the picture inverts. In the one overnight recording persistence is a good baseline and the mean a bad one. One night; an observation.

Skill, ridge on the lags, trained on other sessions

Pooled skill with its 95 % interval; columns are how far ahead.

against persistence2 s4 s8 s16 s
the owner's head+0.38 [+0.27, +0.44]+0.41 [+0.30, +0.46]+0.42 [+0.35, +0.46]+0.38 [+0.26, +0.42]
cold periods+0.33 [+0.25, +0.38]+0.38 [+0.31, +0.42]+0.38 [+0.27, +0.46]+0.40 [+0.29, +0.46]
all+0.36 [+0.33, +0.41]+0.38 [+0.34, +0.43]+0.40 [+0.36, +0.45]+0.36 [+0.32, +0.42]
against the session mean2 s4 s8 s16 s
the owner's head+0.03 [-0.07, +0.15]+0.02 [-0.06, +0.10]+0.00 [-0.07, +0.05]-0.05 [-0.20, +0.03]
cold periods+0.16 [+0.06, +0.22]+0.08 [-0.05, +0.14]+0.00 [-0.11, +0.05]-0.03 [-0.20, +0.01]
all+0.46 [+0.04, +0.59]+0.44 [+0.01, +0.57]+0.41 [-0.00, +0.55]+0.40 [-0.07, +0.52]

The pre-registered predictions

PredictionVerdict
P1On the owner's head, within a session, skill over persistence at 2 s between 0.05 and 0.30refuted
P2Skill over the session mean falls with horizon and is at most 0.05 at 16 s, in every poolrefuted
P3The GRU never beats ridge by more than 0.05supported
P4Across-session ridge exceeds within-session ridge at 2 s by at most 0.03refuted
P5A model trained on other heads reads the owner's within 0.05 of one trained on his ownsupported
P6The cold pool's skill at 2 s is within 0.10 of the owner'srefuted
P7The cold pool's usable fraction is lower than the owner's by at least 0.15 (as worded, inside the eligible pools)refuted

P1 is refuted upward: the owner's head was more predictable than written, for the reason in point 2. P2, P4 and P6 fall to the fragile model in point 6 and to this n. The design, with the two amendments made before the first run, is in PREDICTIONS.md.

A local model proposes, the gate disposes

Ornith, a 9-billion-parameter model running on the laptop through Ollama, saw only a briefing: the column dictionary, counts per pool and condition, group means, and in the first round the forecasting summary. It was asked for twelve falsifiable claims in a fixed grammar (two groups compared on a statistic, a correlation across sessions, or a forecast-skill statement). Each claim met a permutation test or the held-out forecast result, with Holm correction across all its claims, groups under five sessions abstaining.

24claims in two rounds
8let through as empirical
2refuted, both textbook intuitions
5inconclusive at this n
7malformed: groups that overlap

Round one: the model saw the forecast numbers

12 claims: 6 supported, 0 refuted, 2 inconclusive at this n, 1 abstained, 3 malformed.

#The claimOutcomeWhat the gate sawThe model's reason
1usable_fraction in mine/any > cold/anymalformedreason groups overlap on 2 sessionsThe owner's own head has 0.45 usable fraction vs 0.11 for cold exposure, a 4-fold difference likely driven by better electrode contact and less motion.
2gyro_rms_mean - usable_fraction across allinconclusiverho -0.073; n 78; p_stated_holm 0.255Motion (gyro) is 21.4 average for all sessions; high motion should degrade signal, reducing usable fraction.
3mean_logbeta + mean_loggamma across allsupported (empirical)rho 0.966; n 39; p_stated_holm 0.000Beta and gamma are high-frequency bands; they should be positively correlated in band power.
4gyro_rms_mean in mine/any < cold/anymalformedreason groups overlap on 2 sessionsThe owner's head shows 10.8 mean gyro while cold exposure shows 25.3, consistent with more movement during cold plunges.
5mean_index in mine/any > mine_sleep/sleepabstainreason underpowered: n_a=7, n_b=1 (need 5 each)The owner's awake sessions have mean_index 0.078 vs -0.150 during sleep, consistent with activation index being positive in awake states.
6ridge_all on mine/A h=1: skill >= 0.2supported (empirical)pooled_skill 0.393; ci [+0.119, +0.507]; min_skill 0.200Mine h=1 skill is +0.393 with 95% CI [+0.119, +0.507] excluding zero, well above min_skill threshold.
7ridge_all on mine/A h=4: skill >= 0.2supported (empirical)pooled_skill 0.450; ci [+0.286, +0.528]; min_skill 0.200Mine h=4 skill is +0.450 with 95% CI [+0.286, +0.528] excluding zero, strong predictive power.
8ridge_all on cold/A h=1: skill >= -0.1inconclusivepooled_skill 0.155; ci [-0.078, +0.347]; min_skill -0.100Cold h=1 skill is +0.155 with 95% CI [-0.078, +0.347] which includes zero, but with min_skill of -0.100, it is marginally supported.
9mean_index + mean_logbeta across allsupported (empirical)rho 0.566; n 39; p_stated_holm 0.000Higher activation index should correlate with higher beta power.
10mean_logalpha in mine/any < cold/anymalformedreason groups overlap on 2 sessionsMine alpha is 0.48 while cold alpha is 0.82, suggesting more alpha (relaxation) during cold exposure.
11ridge_all on all/A h=4: skill >= 0.2supported (empirical)pooled_skill 0.249; ci [+0.030, +0.461]; min_skill 0.200All h=4 skill is +0.249 with 95% CI [+0.030, +0.461] excluding zero, supporting h=4 forecast for all sessions.
12gyro_rms_mean + mean_logtheta across allsupported (empirical)rho 0.786; n 29; p_stated_holm 0.000Motion may increase theta power (drowsiness/noise), expecting positive correlation.

Round two, blind: the forecast numbers withheld

12 claims: 2 supported, 2 refuted, 3 inconclusive at this n, 1 abstained, 4 malformed.

#The claimOutcomeWhat the gate sawThe model's reason
1usable_fraction in mine/any > cold/anymalformedreason groups overlap on 2 sessionsThe owner's own head (mine, 0.45) shows much higher usable fraction than cold-exposure sessions (cold, 0.11), likely because the owner can adjust the headband more carefully.
2gyro_rms_mean - usable_fraction across allinconclusiverho -0.073; n 78; p_stated_holm 1.000Higher head motion (gyro) should reduce the fraction of usable epochs due to motion artifacts corrupting the signal.
3mean_logalpha + mean_logtheta across allsupported (empirical)rho 0.890; n 39; p_stated_holm 0.001Alpha and theta bands typically co-vary as slow oscillations; positive correlation in log power is expected across sessions.
4mean_logbeta + mean_loggamma across allsupported (empirical)rho 0.966; n 39; p_stated_holm 0.001Beta and gamma bands often increase together during active states; a positive correlation in log power is plausible across sessions.
5mean_index in mine/any > cold/anymalformedreason groups overlap on 2 sessionsThe activation index (beta+gamma)/(theta+alpha) is higher in mine (0.078) than cold (0.062), suggesting the owner's baseline state is more aroused or the cold dampens high-frequency power.
6mean_logalpha in mine/any < cold/anymalformedreason groups overlap on 2 sessionsMine has lower mean logalpha (0.48) than cold (0.82), possibly because cold exposure increases alpha power (relaxation response) while the owner's baseline is more active.
7minutes + usable_fraction across allinconclusiverho 0.118; n 115; p_stated_holm 0.514Longer recording sessions may allow better electrode settling, leading to higher usable fraction over time.
8mean_logalpha in mine_sleep/any < all/anymalformedreason groups overlap on 1 sessionsSleep shows very low alpha (0.17) compared to all sessions (0.73), consistent with alpha suppression during sleep.
9gyro_rms_mean - mean_index across allinconclusiverho -0.107; n 29; p_stated_holm 1.000Higher head motion should disrupt the activation index by introducing noise that inflates theta/alpha or disrupts beta/gamma balance.
10usable_fraction in mine/any < mine_sleep/anyabstainreason underpowered: n_a=14, n_b=1 (need 5 each)Mine (0.45) has lower usable fraction than mine_sleep (0.96), which is expected since sleep epochs are cleaner and the overnight recording was well-worn.
11mean_logbeta - mean_logalpha across allrefutedrho 0.755; n 39; p_stated_holm 1.000Beta and alpha are often inversely related (beta up during alertness, alpha up during rest), so a negative correlation is plausible across sessions.
12mean_loggamma - mean_logalpha across allrefutedrho 0.649; n 39; p_stated_holm 1.000Gamma power is typically low during relaxed states where alpha dominates, suggesting a negative correlation between gamma and alpha across sessions.
Reading the two rounds together

Shown the forecast numbers, the model proposed them back; shown none, it proposed no forecast at all. Its supported claims are correlations that are nearly definitional on a four-electrode headband. Its refuted claims are textbook intuitions (beta and gamma fall as alpha rises) that run the other way across these sessions, where every band moves with amplitude and contact. It compared "the owner" with "cold periods" seven times in a way the gate could not test, after being told how. And once it set a negative skill threshold and called the result "marginally supported"; the gate returned inconclusive. The gate held every time. The ledger is the result, not the count of green rows.

What this does not show

It does not show that a plunge response can be anticipated, that any mental state is read, or that the index after a plunge means arousal rather than muscle. Anticipation at 2 to 4 s is real and is mostly reversion to the level, which is exactly what a game can use to cancel its one to two seconds of latency. Anticipating the response to cold needs marked plunges (in and out markers, a 60 s baseline, 10 to 15 minutes after exit) and a leave-one-plunge-out test, which is the next thing to record.

Files