The extraction layer on this rig

Status: experimental. Everything here uses the optional companion signal-detection tool (found through $OLIMEX_TOOLS) and private captures that are not in this repo (found only through $OLIMEX_RIG_CAPTURES). Without both, rig_noise_floor.py exits with a message saying which is missing. Every µV and nV figure below inherits the rig's 7.9 µV/count, which comes from the shield's built-in cal signal at its nominal amplitude and is not externally verified (rig.UV_PER_COUNT_PROVENANCE). See docs/experimental.md.

Part 1 of this write-up (the layer against a real human 40 Hz ASSR, OpenNeuro ds005185) lives with the companion signal-detection tool. What stays here is about THIS board.

OLIMEX_RIG_CAPTURES=… OLIMEX_TOOLS=… python3 rig_noise_floor.py. This is EXPLORATORY/RIG-CAN-DETECT-AT-ALL, and it is no longer UNRUN.

A private set of eight sessions recorded on the actual Olimex rig (not in this repo) (Oz–Cz, mastoid ground, 250/256 Hz, ~1.5 M samples) with sidecar JSON documenting acquisition. The script reads them and writes nothing — they are personal recordings and stay local. Everything below is an aggregate statistic.

The two unverified rig figures

The correction that matters: a broadband RMS is not the noise a lock-in fights

required_dwell_s() assumes white noise and therefore takes a broadband RMS. Real EEG is drift-dominated, and the two differ enormously:

The white-noise model is 3.5–24× pessimistic in amplitude (median 15×) — 12–594× in dwell time. That error runs in the dangerous direction: it says "unrunnable, go buy hardware" about a measurement that may be fine. local_noise() and budget_from_record() now measure the noise at the frequency of interest from neighbouring bins, assuming only local smoothness in frequency.

The answer

Correction — 2026-09-13. This section previously gave the median as 238 nV and the overstatement as 20–90×. Both came from rig_noise_floor.py output that was wrong: the median indexed sorted(floors)[n // 2], which on six values takes the upper-middle one, and the "90" was a hard-coded literal in a print statement that nothing computed. From the same six captures the median is 157 nV (range 64–750, 11.7×) and the overstatement is 3.5–24× in amplitude (median 15×; 12–594× in dwell). The spread was also attributed to electrode preparation; the session sidecars do not support that — the one capture scored good-contact (α SNR 36) has the highest 40 Hz figure and the one flagged as implausibly small sits mid-range — so these are in-band power of a scalp recording, not an instrument noise floor. No shorted-input measurement exists. The dwell table and the 712 nV figure are computed on one capture (the 750 nV one), not the median.

Median in-band noise at 40 Hz after 60 s of coherent dwell: 157 nV (range 64–750 nV, n = 6 captures).

A cortical 40 Hz ASSR (hundreds of nV upward) is plausibly within reach on this rig. A brainstem-class response — ABR, FFR, tens of nV — is about two orders of magnitude out of reach, and no amount of processing closes that. That is a decision, and it is the one this whole layer was built to produce.

Three operational facts worth more than the software

1. The spread across sessions is 11.7× (64 → 750 nV), larger than any algorithmic choice in this tool. Corrected 2026-09-13: this item was titled "Electrode prep dominates everything". The sidecars do not support that — the good-contact capture has the highest value — so the cause of the spread is open. Prepping electrodes well is still right; it is not shown here.
2. Mains is 60 Hz here and clearly present. Keep the target frequency and both sidebands off 60 Hz and its harmonics.
3. The sideband window at 256 Hz is 120–128 Hz, so the widest usable Δf is ~5 Hz. Same Nyquist wall as before, confirmed at the rig's real rate.

What the tri-state got right, stated accurately

budget() refused to give a verdict from the RECALLED floor, and that refusal earned its keep — but not by inverting a conclusion. Had the 34 µV figure been believed, the smallest response detectable in 10 min would have read 372 nV; measured on the noisiest capture, it is 712 nV. The two agree on the verdict at every target tested and differ ~2× on the smallest response worth chasing. The recalled figure was optimistic, not wrong-signed. Saying otherwise would be the same overclaiming this document exists to avoid.

The quality gate that matters more than the algorithm

Every dwell projection assumes noise falls as 1/√T. On these captures that assumption holds on 5 of 9 segments and fails outright on 4:

A failing session is drift-limited, not noise-limited: recording for longer buys nothing, and the fix is re-prepping the electrodes, not more time. integration_check() makes this a one-call check to run before committing to a long dwell. It is the cheapest thing in this tool and it is worth more than every algorithmic choice in it.

The pre-registered detection threshold

injection_curve_from_record() adds a known signal to real recorded baseline noise — including its drift, mains and artefacts, none of which a Gaussian model contains. Restricted to sessions that pass the integration gate (n=4):

Reported as a distribution, not a single number, because with a handful of segments the pooled "80% of trials" figure is a worst case set by the noisiest session — and quoting either one alone would misrepresent the rig in whichever direction happened to suit. The ~135× best-to-worst spread is the rig; the algorithm is not the variable.

A bug found in this function while writing it, worth recording: the first version bracketed each segment's search from the pooled median noise, so a segment much noisier than the median failed at the top of the bracket and was silently dropped — biasing the per-segment statistics optimistic by excluding exactly the hardest cases. Each segment now brackets from its own noise, the bracket widens rather than giving up, and anything still undetected is counted in segments_undetected_at_any_tested_amplitude. Same "no silent caps" rule the rest of the repo runs on, violated by its own author on first write.

These numbers were stated in a preregistration before any stimulus session existed, which is the only order in which a capability claim is a limit rather than an excuse for a null.

Still open

The rig's noise is characterised; the rig has never been shown to detect anything. None of these sessions (six with a measurable baseline, of twelve captures) played a stimulus at a known exact frequency, so there is no numerator. The next measurement is now well-posed and cheap: play a 40 Hz AM stimulus, record Oz–Cz, and test the response against the measured floor (median 157 nV). If it lands above ~700 nV it will be visible inside ten minutes even on the noisiest session.

---

Honest summary

The analysis chain has now recovered a real human evoked response, with correct controls, at a frequency it was told about in advance — and the exercise immediately found a defect that would have produced a false sustain finding. That is the whole argument for running on real data.

It validates the analysis, on someone else's better hardware. The cheap bench measurement of this rig still has not been made, and that remains the gate.

Provenance

Data is CC0 (Mikkelsen, Kidmose & Rezaei Tabar; doi:10.18112/openneuro.ds005185.v1.0.0), fetched version-pinned at v1.0.0 with the licence re-checked live at fetch time. validation/data/ is gitignored — the scripts reproduce it. numpy/scipy are used only to read the EEGLAB .set/.fdt container; every signal-processing call goes to the stdlib extract.py with plain Python lists, so the tool under test gains no dependency.

Sidenote worth keeping: the first-choice dataset was ds008065 (Optimizing parameters for multiband ABRs to continuous speech — an actual FFR study, the closer domain match). Every one of its S3 objects returns HTTP 403, including dataset_description.json; 25 other datasets tested identically returned 200. That is server-side and dataset-specific. Re-check before planning around it.