The Oscillome · 24 — Version history
The sections of this site always state the current best reading. This page records how that reading changed. Three things are versioned separately: the acquisition protocol (what was physically done in a recording), the dataset release (which recordings and labels existed) and the analysis release (which pipeline produced the numbers).
| Analysis release | Date | Dataset | What changed |
|---|---|---|---|
| 4.26 | 2026-10-13 | 2026.10.10 · 56 recordings | A new section, §22: the recordings as a whole-head dynamical system. The wave modes of each band as collective variables, with their literature names; the instability named (a supercritical Hopf bifurcation of the alpha front–back mode) with every regime border calibrated on a simulated linear system; phase portraits of the modes; the unstable minutes listed and shown as raw EEG beside rest; brain states as regimes of the phase diagram, the transitions between them and their shape; the control plane of eye state and time; the maximum-entropy phase diagram; a ring of Jansen–Rit columns with a gradient that crosses the same bifurcation, the field model with spatial modes that cannot, and the ring inverted onto every minute of the recordings; the five models as dynamical systems, the eyes-closing transition, single-electrode drift and jumps, the complex-systems repertoire; and the learned representations on the regimes and transitions, with the benchmark of signals with one spectrum and two regimes. The L2 model is refitted inside its stable zone for that section; §20 keeps its published constants and says so. The statement that the alpha envelope requires no limit cycle is now made of whole recordings, with the window-level result beside it. Sections 22 to 26 become 23 to 27. Recording pages: every panel is now computed for all 42 recordings: window-level alpha for Ayahuasca experiment 005 (the original run had stopped at one flat window), channel parameterization and hemispheric asymmetry for Resting baselines 002 and 003 (part 1) and Workday recording 014 (the archive tables predated them), a bounded refit of the whole-recording spectral fit for Workday recording 002 (stated on its page), the asymmetry-state metrics for DMT experiments 003, 012 and 013, and a stated reason wherever a recording is too short for a panel. The site is now nine pages: the home page keeps the overview, the evidence ledger, the instrument and the archive explorer, and each research program has its own page (A to E), with program F in three: neural-mass models, learned representations, brain states and phase diagrams. Section numbers, anchors and every cross-reference are unchanged; a link to a section lands on the page that holds it. The October 2026 update: fourteen recordings added (Workday recordings 020 to 028, Daily-life recordings 004 and 005, Psilocybin experiment 017, Ayahuasca experiments 018 and 019) and Workday recording 009 extended from three minutes to its full 95; 56 recordings and 100 hours, February to October; every recording in the analysis cohort. Every Program A to E analysis was re-run on the 56 with the archive’s own scripts and the sections rewritten to the new numbers; the ten new flicker sweeps and seven light-noise runs join §18 and §19; §09 gains the frontal–posterior coupling analysis; Program F keeps its published cohort statistics and gains its first prospective test (§21, Result 22), the regime check of the new sessions (§22, Result 20) and two retained psychedelic sessions fitted post-freeze (§20). |
| 4.24 | 2026-09-27 | 2026.09.16 · 42 recordings | The model comparison of §21 becomes the standing one: every representation read by one calibrated, tuned CatBoost read-out with whole sessions held out (decisions at 0.5, session-paired bootstrap tests), the foundation models LoRA fine-tuned by masked reconstruction with a label-free health check before evaluation, the 41 × 14 fitted neural-mass model constants given to the networks as inputs during fine-tuning and concatenated at the read-out, three open-ended targets no network was trained on, ensembles of pre-specified sets, and the LoRA rank sensitivity. Section 20 is recomputed from the per-electrode fits: the tables of fitted constants, the four predictions, the constants as a representation and the per-compound read-outs. The posterior-average results, the segment-agreement fine-tuning, the supervised fine-tuning arm, the greedy ensemble and the scaling study leave the standing results (their numbers remain in earlier releases). Results 7 to 15 (the label-free scorecard, probes, geometry, agreement between spaces, input importance, attribution and robustness) are recomputed for every standing representation, with the constants of the five neural-mass models refitted on every perturbed window so that their robustness is measured rather than assumed, and the per-compound attribution is re-pooled over every standing network with the same CatBoost read-out. The compute note is rewritten for the standing analysis. Results 16 to 20 follow a review of the section on 2026-09-27: the scaling study restated under the standing read-out with the sessions split into seen and unseen, a matched-control ablation of the constants, seed variance and fold-wise refits of the LoRA rows, phase-randomized perturbations that keep every amplitude spectrum, and an oscillatory-against-aperiodic probe of every representation, with a graph transformer that adds a scalp and coupling prior to the replica’s attention, with and without the constants (Result 20); Result 14 traces the five best-read and five worst-read sessions with the CatBoost per-window probabilities. The vocabulary is fixed: the fitted quantities of the neural-mass and mean-field models are the constants of their equations, networks have weights, read-outs have hyperparameters. |
| 4.23 | 2026-09-23 | 2026.09.16 · 42 recordings | Section 20 gains the five neural-mass models fitted to each of the fourteen electrodes separately: all 41 parameters for every electrode of every clean two-minute window, with scalp maps and electrode-by-electrode psychedelic contrasts under both eye states. |
| 4.22 | 2026-09-23 | 2026.09.16 · 42 recordings | Wording: the transformers trained in-house on this archive are no longer described as foundation models; the overview and the ledger now say they borrow REVE's design and were trained on one person's recordings. |
| 4.21 | 2026-09-23 | 2026.09.16 · 42 recordings | A site-wide consistency audit. Stale figures recomputed from the current data: the §06 alpha ranges (nine of sixteen acute sessions, not all, sit above the non-psychedelic maximum), the §11 harmonic cards and the §17 locking values, the artifact yield in §04 and the clean hours in §10. The reference is stated once, in Methods. The eye-state classifier's dependence on alpha is stated wherever an eye-matched comparison carries a result; the blink detector is named as proprietary; §12 gains its sign convention, and its wavelength conversion is folded away as illustrative. “Pre-registered” is replaced by “pre-specified”, since nothing was lodged with a registry. Corrections: two added (the withdrawn first-to-last comparison, the superseded in-house result), Correction 02 and 03 figures fixed. References: the three uncited neural-mass models cited, citation numbers generated by the build, one link and journal style. Recording pages moved to neutral addresses (/r/session-NN/) with their tables rendered without JavaScript, and links to the satellite pages repaired. |
| 4.20 | 2026-09-23 | 2026.09.16 · 42 recordings | The compound comparison split by kind of model. §20 gains the per-compound read-out over the 41 fitted neural-mass parameters, every parameter written out. Result 16 of §21 is replaced by the learned models’ own account: which frequency band in which scalp region carries each compound’s evidence, pooled over the learned models whose read-out finds that compound. |
| 4.19 | 2026-09-23 | 2026.09.16 · 42 recordings | Correction to Result 16 of §21: the per-compound chart plotted the read-out’s coefficients and labeled them SHAP values; it now shows the mean SHAP value on each compound’s own sessions, all three on one axis. The apparent reversal of resonance sharpness between LSD and ayahuasca is shown to be largely two correlated parameters dividing one signal, and the raw fits are added to show what the compounds share. |
| 4.18 | 2026-09-23 | 2026.09.16 · 42 recordings | Result 16 of §21 rebuilt from the models themselves: class-conditioned variable importance in the mechanistic and hand-built spaces, beside the class offsets in the learned ones. |
| 4.17 | 2026-09-23 | 2026.09.16 · 42 recordings | What the models can say about which compound was taken: three-way class read-outs and per-class mechanistic parameters, as Result 16 of §21, following §17. |
| 4.15 | 2026-09-23 | 2026.09.16 · 42 recordings | Ablation of the fused neural-mass target: each of the five models used alone, three seeds each, with per-parameter importance against the identifiability check (Result 11 of §21). Editorial: the five models named wherever the 41-parameter target is described. |
| 4.13 | 2026-09-22 | 2026.09.16 · 42 recordings | A 60-model scaling experiment (four sizes, three data fractions, with and without the neural-mass target, three seeds): physics-informed training versus scale, as Result 15 of §21 and in the overview. |
| 4.12 | 2026-09-21 | 2026.09.16 · 42 recordings | Privacy: recording pages and the corrections page are no longer indexed by search engines; public dates are given as study day and month and clock times as time of day; timestamp-named runs renumbered; participant identity not disclosed; neutral tagline and page descriptions. |
| 4.11 | 2026-09-21 | 2026.09.16 · 42 recordings | Editorial: the dose amount replaced by “a known, small quantity”; spatial-phase, foundation-model and harmonic statements in the reviewer's wording, with the nuisance caveat beside every 39-of-40 result; classifier eye-state labels marked as estimates; model fit distinguished from mechanism identification throughout; wavelength conversions labeled illustrative; headline counters carry their values in the HTML. |
| 4.10 | 2026-09-21 | 2026.09.16 · 42 recordings | A parameter-recovery test for every mechanistic parameter, with each labeled identified, partially constrained, model-coordinate only or prior-dominated, and the parameter-level interpretation of §20 limited accordingly; the linear-resonator conclusion stated at the resolution the test supports; stale statements corrected (eye-state annotation in Methods, artifact yield recomputed on the current archive, a second-headset reference in §07); phase-gradient terminology on recording pages; alpha generators and electrode projections described as approximate; the harmonic and foundation-model headlines stated exactly; the published page now loads in pieces on phones. |
| 4.9 | 2026-09-21 | 2026.09.16 · 42 recordings | t-SNE maps of every representation in §21, with the large map switchable between models; result headings made prominent; the input-perturbation tests called robustness tests; the version-history panel removed from §21; the roadmap section removed as redundant (its eye-state-classifier panel now lives in §16) and later sections renumbered; the headline result of §20 and of §21 added to the overview. |
| 4.8 | 2026-09-21 | 2026.09.16 · 42 recordings | The mechanistic models now precede the learned representations (§20, §21), and both sections were reorganized: the models compared are introduced one by one (with a description of each foundation model), every result has its own heading, and the mechanistic section explains what is new, how each parameter is estimated from the recordings, and what the fitted values say. An ensemble of every label-free model, fully nested, with an analysis of which members it depends on. Missing values in the feature-importance and comparison charts filled. |
| 4.7 | 2026-09-21 | 2026.09.16 · 42 recordings | REVE added to every comparison (frozen, adapted, fine-tuned, feature importance, robustness). The comparison reorganized as five lines: in-house, foundation, fine-tuned, foundation + neural mass, fine-tuned + neural mass, with the neural-mass models joined to the learned models both at the read-out and as a training target. A 9-million-parameter replica of REVE's architecture trained on these recordings only. Training times and hardware stated. |
| 4.6 | 2026-09-21 | 2026.09.16 · 42 recordings | Model evaluation widened beyond AUC (log loss, Brier, confusion matrices with precision and recall, two conditional tests against the time of day and the eye-state estimate); a label-free scorecard; permutation feature importance on the raw signal and exact SHAP for named features; robustness curves under six kinds of input corruption; five neural-mass and mean-field models fitted to every window, tested on their own predictions and compared with the learned representations as embeddings; architecture schematics and the model equations. |
| 4.5 | 2026-09-20 | 2026.09.16 · 42 recordings, 40 session groups in the embedding study | Learned representations redone: in-house transformer retrained on all 42 recordings (three seeds) and with a second objective; LaBraM, CBraMod, BIOT and EEGPT embedded frozen, adapted without labels and fine-tuned with low-rank adapters; one battery of probes, controls and geometry applied to all, with time-of-day, eye-state and hand-built baselines. The in-house 29-of-37 result is superseded: that model does not beat the time of day and fails eye-state matching. Also: an analysis-basis line under every analysis section, a session manifest download, recording pages renamed to “spatial phase structure”, long-range-memory wording in §07. |
| 4.4 | 2026-09-20 | 2026.09.16 · 42 recordings, 41 in cohort | Editorial consolidation. One overview; sections grouped into six research programs; the evidence ledger generated from a claim registry with a controlled status vocabulary; version notes moved here. Corrected on the page: the embedding benchmark is 29 of 37 everywhere (27 of 30 is the superseded earlier figure); one headset unit was used throughout (Methods had wrongly mentioned a second); dose records exist for the September protocol; the acute sessions are not matched on eye closure; the equal-frequency argument was removed from the asymmetry section; the code repository is not yet public. |
| 4.3 | 2026-09-18 | 2026.09.16 | Waves in motion: clip player and cortex view; direction bias split by eye state; 21-band scan; pairs of coexisting waves; cross-band tests. |
| 4.0–4.2 | 2026-09-17 | 2026.09.16 (four mid-September workdays added) | Four mid-September workdays were added, each with the fixed eyes-open / eyes-closed block, four more photic sweeps and six runs of a new experiment, broadband light noise (§19): its headline is a perceptual echo at 10.8 Hz, and a claim that did not survive, that brighter light speeds alpha up. Every archive-level measure for recordings that contain stimulation is now computed with the stimulation masked. The embedding (§21) was retrained and says less; the long-term section (§09) was refitted with curved forms and now covers the psychedelic sessions; the 5 Hz line of LSD experiment 016 (§11) turns out to be mostly a head movement locked 2:1 to alpha; the wave measure of §12 is re-read against eye state; and an eye-state classifier now labels every window (§16). |
| 3.0–3.8 | 2026-09-09 | 2026.09.08 · 38 recordings, 37 in cohort | Five recordings were added after the v2 cut-off: an ordinary workday, a resting baseline split by an application restart, a second workday, and one LSD-labeled session. Every one opened with a fixed eyes-open / eyes-closed block, and three of them ran the photic resonance protocol, five sweeps in all. The new section §18 reports those sweeps; every earlier section that posed a question the sweeps bear on now carries a short panel with the answer. Per-recording metrics, 1/f-corrected asymmetry, the alpha harmonic, per-channel band dynamics and the wave portraits include the new recordings; the label-free embedding (§21) and the archive-wide longitudinal tests (§09) were not re-run and remain on the original cohort, as their sections state. Later 3.x releases added the four-class comparison, the band primer, the dissection of the 5 Hz line in LSD experiment 016, and the change from “drug” to “psychedelic” throughout. |
| 2.0 | 2026-09-04 | 2026.08 · 33 recordings, 30 in the embedding | Labels corrected (three recordings archived as baselines were DMT sessions), which collapsed the resting comparison group; accessibility pass; per-recording pages. |
| 1.0 | 2026-09-04 | 2026.08 | First public version. |
| Protocol | From | What a session contains |
|---|---|---|
| P1 · legacy | February 2026 | Free-form recording; eye state, dose and timing not logged. Eye state reconstructed afterwards by the classifier of §16. |
| P2 | 4 September 2026 | Opens with a fixed block (30 s settling, 120 s eyes open, 120 s eyes closed); stepped flicker sweeps under the analysis plan; dose amount recorded. |
| P3 | 10 September 2026 | P2 plus broadband light-noise runs with a fixed intensity schedule. |
| P4 · planned | — | P3 plus dose clock time, subjective ratings at each block, a photodiode channel, eye recording and head support. |
A single-subject EEG archive: 56 recordings, 100.1 hours, 19 acute sessions, four psychedelic classes, 34 photic resonance sweeps, thirteen light-noise runs, one brain, February to October 2026. All 56 recordings form the analysis cohort.