public blackboard · GitHub Pages + Discord feed
Demian Lab is the evidence ledger for a stateful-systems research line.
This standalone preview mirrors the Astro site direction: curated research posts, a status blackboard,
and stable updates.json / rss.xml
feeds that a client can poll without scraping fragile page markup. EEG, Geo, and Zenith now have separate runnable public extractions; historical charts remain explicitly labeled appendices.
1 · runnable mechanisms Minimal EEG, Geo, and epistemic-state repositories expose code, synthetic fixtures, tests, and claim boundaries.
2 · curated posts Astro Markdown posts summarize what changed, what ran, what can be claimed, and what cannot.
3 · public feeds The site emits updates.json and rss.xml with titles, tags, metrics, artifact paths, and canonical URLs.
4 · Discord relay A bot can poll the feeds, route by tags, and organize updates into channels without bespoke scraping.
250
coupling streams
historical saved-run sample
1.0
positive growth fraction
recorded probe measure
101/250
reconvergent
bounded trajectories that reconverged
v3
architecture
active public progress phase
JSON
discord feed
/updates.json · tags + metrics
RSS
public syndication
/rss.xml · standard polling
metric legend — what each number means
tension reservoir
Accumulated internal pressure / memory. Rising means the model keeps integrating context instead of resetting.
uncertainty
How unstable or unresolved the current state is. Read beside tension: high tension + high uncertainty means stored pressure without clean resolution.
route tortuosity
How indirect the internal trajectory is. 1 is straight; higher values mean the state path bends, loops, or detours.
self/other cosine
Alignment between Demian's self-trace and other/input trace. Negative means they remain separated, not merged.
gate state norm
How much the gate has accumulated. Large non-zero values mean the gate is active and shaping channel flow.
prediction error
Mismatch between expected and incoming state. Spikes are windows where the model had to update harder.
local run appendix readout
reservoir change2.75 → 27.22
uncertainty final3.14
fast/slow ratio final1.37
prediction error final2.44
mean |coupling| final0.69
live vs gate-frozen L21.085
ordered vs shuffled L21.479
Interpretation: this local run appendix remains useful evidence, while the public blackboard headline is the 250-recording saved probe summary.
Plain-English verdict: the saved probe records a bounded observer response to a small input perturbation.
The local charts below show one detailed EEG-derived run appendix. Neither surface establishes brain-state decoding, diagnosis, biomarkers, or clinical relevance.
navigation — what each tab is for
Architecture
Internal channels, gate behavior, fast/slow competition, L2 ablations.
EEG Mapping
How raw channels become band, region, asymmetry, and quality features.
Experiments
Run-specific evidence: charts, findings, and sanity checks.
Blackboard
Public status, feed contract, claims, open questions, and next hypotheses without overclaiming.
Tension dynamics over time
Tension reservoir accumulation
Route tortuosity (global)
Self / other cosine (trisequence)
Prediction error (per window)
Mean absolute coupling (32-dim)
Final coupling weights — 32 hidden dimensions
■ negative ■ positive range: -0.98 → 0.98 · L2 4.15
internal channels — demian_v1_trisequence_tension_v3
select a channel →
Click any channel card to see its role, shape, and live final-state values.
runtime config
model config
| feature_mode | topographic |
| hidden_size | 32 |
| window_seconds | 2.0 |
| hop_seconds | 1.0 |
| strength | 0.25 |
| seed | 0 |
| shuffle_seed | 17 |
| runtime_variant | trisequence_tension_v3 |
summary L2 distances
live vs gate frozen1.085
surface only vs source resume1.648
ordered vs shuffled1.479
full vs source resume0.000
Large live/frozen gap confirms gate is actively modulating. Ordered/shuffled gap confirms temporal order carries signal.
fast / slow ratio over time
adaptive fast/slow scale ratio — shows timescale competition
how 64 EEG channels map into demian features
64
raw EEG channels
1000 Hz · 2s windows
109
feature entries
per 2s window
5
frequency bands
δ θ α β γ
5
scalp regions
frontal · central · temporal · parietal · occipital
feature pipeline
64 channels raw µV, 1000 Hz continuous
→ windowing 2s window, 1s hop → 217 frames
→ PSD power spectral density per channel
→ band split δ(1-4) θ(4-8) α(8-13) β(13-30) γ(30-50) Hz
→ global 19 global features: mean, std, range per band + quality
→ regional 5 bands × (5 regions + L/R/midline + 2 asymmetry) = 88
→ surface 109-entry vector → demian carrier input
frequency bands → relative power
| band | range | features extracted | global rel. mean (t=0) |
|---|
| delta δ | 1–4 Hz | mean · std · range + 8 regional | |
| theta θ | 4–8 Hz | mean · std · range + 8 regional | |
| alpha α | 8–13 Hz | mean · std · range + 8 regional | |
| beta β | 13–30 Hz | mean · std · range + 8 regional | |
| gamma γ | 30–50 Hz | mean · std · range + 8 regional | |
Bars are schematic UI values, not measurements or normative EEG claims. The adapter computes relative power as band power divided by total power.
scalp regions → demian topographic features
regional coverage — channels grouped into regions, each gets δ θ α β γ relative mean + coverage fraction
frontal
Fp1,Fp2
F3–F8
5 features / band
central
C3,Cz,C4
5 features / band
temporal
T7,T8
TP7,TP8
5 features / band
parietal
P3–P8
Pz
5 features / band
occipital
O1,Oz,O2
5 features / band
left hemi
odd channels
coverage + mean
right hemi
even channels
coverage + mean
midline
Fz,Cz,Pz,Oz
coverage + mean
L/R asymm
left − right
per band
F/O asymm
frontal − occipital
per band
global quality features
artifact / quality flags embedded in the feature vector
missing_fraction
channels dropped / total
flat_fraction
near-zero variance windows
log_rms
log root-mean-square amplitude
artifact_fraction
high-amplitude outlier fraction
These 4 global quality features are mixed directly into the 109-entry surface vector alongside spectral features. The model is not shielded from artifacts — they are part of the signal it observes.
current experiment — sub-003 · ses-1 · task-EyesClosed · acq-pre
demian_v1
runtime
trisequence_tension_v3
ds005385
dataset
sub-003 · ses-1
seed 0
deterministic run
shuffle_seed: 17
0.25
coupling strength
EEG surface → carrier injection
gate state norm — accumulation
Gate norm moves from 1.08 to 255.55 across 217 windows. It is active, not frozen.
fast / slow ratio — timescale competition
Ratio starts at 1.95 and ends at 1.37. Spikes mark fast-channel dominance; lower values mean slow integration catches up.
key findings — this run
confirmed
Tension reservoir stores persistent context
Reservoir starts at 2.75, peaks at 29.11, and ends at 27.22. It accumulates a session-level state rather than resetting per window.
reservoir: 2.75 → 27.22 · max 29.11
confirmed
Self/other traces stay separated
Cosine remains negative throughout, ranging from -0.96 to -0.50. The model does not collapse self-trace and input/other-trace into one representation.
cosine range: [-0.96, -0.50], final: -0.79
confirmed
Gate and order both matter
L2 ablations are non-zero: live vs gate-frozen is 1.085, ordered vs shuffled is 1.479, and surface-only vs source-resume is 1.648.
gate L2: 1.085 | shuffled L2: 1.479
open question
Route tortuosity — what drives the late complexity?
Tortuosity starts at 1.00, reaches 10.27, and ends at 10.25. Next run should mark whether spikes align with EEG-band shifts, gate pressure, or prediction error.
tortuosity: 1.00 → 10.25 · max 10.27
not claimed
This experiment does not claim: brain simulation · neural decoding · biomarker discovery · clinical relevance.
The model observes EEG-derived features and evolves internal dynamics. All observations are about the model's behavior, not the brain.
public blackboard + discord contract
GitHub Pages target
https://aeshma-daeva.github.io/Demian-Lab
The Astro site is the public source of truth. Markdown posts stay curated, while the blackboard and feeds make the current state easy to follow from Discord.
current status
v3 active · latest public result
A historical perturbation probe records the observer response across 250 seed-0 coupling streams.
recordings=250 · positive_growth_fraction=1.0 · reconvergent=101/250
automation
Draft posts from EEG artifacts
The local Astro helper can read an artifact from a configured EEG source and create an editable Markdown draft with artifact path, tags, and metrics.
npm run draft:post -- --source adaptation_probes/.../report.md --tags eeg,v3,probe
discord feed
Bot should poll updates.json first
Use JSON for reliable channel routing: title, date, description, tags, canonical URL, artifact paths, and metrics are emitted in one stable object.
GET /updates.json → updates[].tags → Discord channel rules
fallback feed
RSS is available for generic readers
RSS mirrors public log entries for standard feed readers or bots that already support RSS polling.
GET /rss.xml → item title/link/description/category
open
Keep blackboard data synchronized
When a new public result becomes the headline, update src/data/blackboard.json and publish a matching log post so humans and bots see the same state.
open
Discord organization rule
Recommended initial routing: eeg and probe updates into research-progress, tests and cli into engineering-log, architecture into architecture-notes.
feed preview
title: Lyapunov-style perturbation probe: v3 results
url: https://aeshma-daeva.github.io/Demian-Lab/log/2026-06-23-lyapunov-probe
tags: perturbation · v3 · eeg · probe
metrics: recordings=250 · positive_growth_fraction=1.0 · bounded_sensitive_reconvergent=101
artifacts: eeg-perturbation-summary.json
claims boundary
can claim
The saved probe records an observer response to a small input perturbation
The recorded response remains bounded within the observer route
The public site exposes curated posts plus machine-readable feeds
Discord can organize updates by tags and metrics
cannot claim
Brain simulation
Neural decoding or classification
Clinical biomarkers or medical relevance
A formal Lyapunov exponent