score row → family → input → chains → decoder. keep that trail intact or this becomes a spreadsheet séance... which is a terrible debugging strategy September 17 evidence cutoff: local repository e2b603d5 , as recorded in the milestone...
C = 36 from the one-SE rule, raw max was 144. ESS/iteration jump looks huge, ESS/hour much less dramatic... use the latter for compute planning. recorded calibration: fifteen proposal weights × three seeds = 45 points, each with 2,000...
keep chain-specific points next to pooled results. starts, decoder, version, model, budget all differ here, so don't squint at a scatterplot and blame everything on one thing. diagnostic panel: 24 families, two 30-minute chains per family. The...
flat alignment length trace can still hide completely different homology. ugh, need diagnostics that actually notice that. Save idea: animate samples w/ residue identities fixed, watch the pairings move instead. why scalar traces are...
runtime fit ends way before the longest family, big extrapolation gap. Need an actual packed-VM throughput test before trusting the cost math; 3 chains on one machine might just slow each other down. runtime model: nonnegative fit using sequence...
tiny smoke runs for execution → decode → burn-in → pooling. raw-OK can still leave zero usable samples... useful failure to catch before paying for the bigger run. simulation panel: six original conditions, twenty datasets each, 27 taxa,...
fixed two-bar case, actually enumerate the source + diagonal options. ties count with multiplicity; don't lose one just because the costs happen to match Worked window in Supplement S2: a=0.049 , b=0.105 , midpoint c=0.077 , half-width h=0.028 ....
seed isn't the whole reproduction recipe. draw order + versions + tolerances + finite-bar conventions too, otherwise same seed / different diagrams somehow Protocol distinction: the supplement's 300 cloud-pair stress study uses a common ruler of...
Interesting experiment- how well does AI critique? useful ones: what am i preserving, what gets discarded, why choose this over the simpler calculation? source provenance: the files are stored under Claude Review . Their referee-report style and...
rehab game work + support groups + a few test users over the summer. Clinic access still slow; keep track of what was actually running in each session, live EEG vs mock vs camera stub recorded summer activity: adapting the EEG-control project...
sneaky normalization problem: cloud barely changes, its ruler does. suspicious now of comparisons where every object gets its own convenient ruler manuscript concept: a “ruler” is the positive scale dividing pairwise distances. If two clouds...
cloud budget = more than chains × CPU price. include parsing, retries, idle VM time, storage etc... all the tiny charges add up :/ retrospective summer 2026 direction: move the evaluation framework toward standard published protocols and cloud...
one headset = one person, setup included. Add calibration + rest + cleaning before trusting the 30–40 min slot... the timing adds up fast constraint in the July deployment plan: one EEG headset permits one EEG user at a time. The planning...
first things to track: setup time, comfort, instructions making sense, abandoned sessions. Save difficulty + normalization alongside scores or future me gets a graph of ??? First evaluation goal: setup feasibility, instruction comprehension,...
Unity → React/TS for the patient side, Electron console + Python signal engine. Keeping the folder map here; check the pending list before assuming a whole path is hooked up architecture revision in the engineering specification: replace the...
one pull for EEG + IMU, not two consumers fighting over the inlet! Check channel order/counters too. later: tiny timeline viewer for markers + sample gaps? code pointers: signal-engine/acquisition/ , MockUnicornOutlet , and...
artifact gates need the raw branch, before filtering/CAR can hide the thing being rejected. Also keep SOS state across chunks — check resets + chunk-size changes specified real-time chain: latched DC-offset subtraction → causal filter chain...
ERD formula + calibration timings here so I stop hunting around. near-zero baseline? check trial split BEFORE generating overlapping windows, sneaky leakage possibility Feature definitions recorded: ERD% = (Pbaseline - Ptask) / Pbaseline * 100 ;...
held score ≠ fresh score. Need artifact flag + null sub-scores, and don’t keep updating EMA/normalization while holding Specified score: IntentScore = w_eeg*EEG_Score + w_kin*Kinematic_Score - Penalty_Compensation , normalized to 0–100....
lost arm tracking ≠ arm stayed still, ugh this needs an explicit state. Check frame age all the way through WebRTC. Save for later: small camera-confidence indicator instead of making someone watch a skeleton overlay? target path: browser...
pairing codes: expiry, second browser, reconnect, sidecar dying. Need to try all the annoying paths, especially browser still connected while Python is gone lAN responsibilities: serve the React build over HTTPS, pair one browser to a live...
same packets through pure model + rendered game, compare behavior. Log the difficulty staircase too; higher score could just be easier targets game references: docs/gardengrow.md and docs/reachcollect.md ; patient-web/src/garden|reach/ for...
alignment accuracy, tree recovery, parameter recovery, mixing — separate axes. faster iterations are nice but ESS/time is the more useful number. Also come back to the fixed-tree vs fixed-alignment controls. code-reading pointers in the plan:...
need both biological + simulated cases. simulation gives known history; structural references are a different check, can't just call both “truth” and move on. saved reference: BAli-Phy biological-versus-simulation evaluation . The June 13...
trace ONE window all the way from filter to hand command. primer and code should agree on the intermediate objects, especially covariance / tangent space / readout primer architecture: filter EEG, estimate sliding covariance, log-map it to...
7 populations, signed connections, PT readout. check the stability notation... spectral radius can't be negative, need the largest real part where i'm talking about continuous-time stability model in the anatomical draft: a two-dimensional...
tangents parking lot: Takens, PLLs, bifurcations, time scales... a loop still doesn't tell me speed or phase. keep coming back to that manifold question: does low-dimensional activity reflect a circuit constraint, a restricted task, or both?...
TDA on EEG can mean direct signal, connectivity, or delay clouds. need to check which object each paper is actually using before borrowing the method three analysis routes preserved in the review notes: filtration directly on a signal, topology...
header fixes + video conversions, keep source → derived links. column reorder or just relabel? also frame timing, same recurring headache archived changes: b6b6d90 organizes March 1 trials, remaps electrode headers and fixes...
what do i mean by realization here, precisely. connect the state / observation / readout objects to the features in code, don't leave the math floating off by itself archived document (ac4a37c): realization.tex/pdf replaces an earlier proposal,...
March 1 recordings need an index. montage + clock info + matching files + exclusions, folders alone aren't enough for me to reconstruct the session Date provenance: the archive's March 2 commits 10aff2b , 0fca0e9 and b6b6d90 contain recordings...
Riemannian vs no-Riemannian runs: what actually changed besides the name? scale, dimension, distance... keep those matched before getting excited about different loops commit 09f599e: tda_eeg_pipeline.py , playground.py , extensions to tda.py ,...
folder rename ≠ new experiment, keep the mapping. also check what WebM → MP4 does to timestamps / frame count, don't quietly lose the original timing folder changes (4585bf7, 723d039): rhythmic February 22 runs become feb22lt_rhythmic/run1...
electrode rename vs physically different montage — need to be VERY clear which. model won't complain if i feed it the wrong channel order, it'll just use it Manual and automated changes ( 4aacf54 , aad3c28 ): electrode names are remapped in...
window → covariance → tangent features, interesting shift from looking at each channel alone. draw what shrinkage does to a nearly flat covariance ellipsoid later archived module (fb0bfe7): src/control/riemann_features.py implements...
collector moving out of the main repo. pin both versions with the session somehow, otherwise which frontend produced which analysis input?? archived change (f5ccae3): the React eeg-collector application is removed from the main repository and...
Feb 22 bundle: EEG, cue times, video, overlay, landmarks. make a per-trial map before this turns into matching files by vibes Archived acquisition ( 1f8e3eb ): the session includes rhythmic and sustained/practice-style trial groups, each with EEG...
feasibility / information bounds / proportional control — related but not the same problem. need to get from a continuous estimate to an actual command, smoothing class probabilities isn't the whole thing documents added (82921aa, c4d44f0):...
rhythmic trials + camera + landmarks, now several clocks to line up. maybe a flash visible in the video and logged by the collector for a timing check? collector changes (763c430, 46f1643): PeriodicMotionMode.jsx introduces a rhythmic/periodic...
keep raw video AND overlay. rerunning the tracker later is much easier if i haven't baked its guesses into the only copy archived change (a932083): CameraRecorder.jsx saves the unprocessed webcam feed alongside the MediaPipe-overlay recording....
proof/ → theory/, plus a place for open problems. separate what needs a derivation from what needs an experiment; otherwise i just keep circling the same questions reorganization (3fda9ec): src/proof/ becomes src/theory/ ; formal documents and...
can i build W and U around the feature geometry instead of hoping a random matrix behaves? keep the desired modes + input projection explicit, spectral radius alone isn't the full story february open problem: given Riemannian feature trajectories...
multiple primitives + rest + fast transitions, not just one nice stable cycle. wondering if faster switching makes the individual patterns harder to stabilize problem statement from the open-problems document: active movements are intended to...
write out latent dynamics → EEG observation → readout assumptions. where does the actual code fit into this, and what would make the model inapplicable? archived documents (5cb035f): m1_reservoir_formal.tex and technical_summary.md ,...
LaTeX crumbs everywhere... aux, log, fls etc can go. keep enough to rebuild, but check i'm not ignoring the only surviving output along with the junk cleanup (6a5adaf, 0640c79): LaTeX auxiliary files are removed and ignore rules expanded for...
README figures need breadcrumbs back to the run. data id + config + script near the image would save so much hunting later documentation commits: 4f4b43e adds the project title and status; 124da36 adds figures under img/ . The archive identifies...
prep / execution / rest: where exactly am i cutting these? cue time isn't necessarily movement onset, and windows crossing the boundary could get messy recorded extension (e7f5e25): the TDA pipeline adds multiphase analysis and generates another...
50 poses, 2 s holds, 1.2 s transitions — need a clear target without making the UI distracting. maybe previewing the next pose helps timing, or maybe it just encourages anticipation app split (0d32926): ClassificationMode.jsx ,...
window → delay cloud → persistence diagram, lots of knobs in between. keep delay / dimension / filtration / scale with the output or i'll just have a folder of mysterious plots archive (eb1afda): tda.py adds delay-coordinate embedding,...
anatomical prior could take some of the burden off EEG. how much motion comes from the signal vs constraints though? similar question in animation, good rig doing half the work archived artifact (1e6fa3d): anatomical.py , later reorganized as...
