Reported interface observations: historical evidence boundary¶
Status: Historical reported results, not independently reproduced from a pinned dataset/run manifest in this checkout. These summaries are retained for provenance and negative findings; they are not current API specifications, admitted physical validation or a research execution queue.
The numerical tables were consolidated from the former
docs/EMPIRICAL_CONFRONTATION_EEG.md and the historical result sections of
docs/STRUCTURAL_INTERFACE_THEORY.md. Moving them does not rerun, authenticate
or upgrade any result. Current behavior belongs to the
observational interface guide;
physical admission and priorities belong to the
execution plan.
The passive transport and
TCLab owners retain their separately scoped
models, protocols and evidence; the summaries below do not replace those records.
Reported static ranking¶
Historically reported ranking power (ROC-AUC) against classifier errors. The reported target differs from local disagreement; independent replay must still verify the split, label availability, graph construction and preprocessing. The circular "local-disagreement" target gives ≈ 1.0 for all local scores and is used only as a localization sanity check.
| Dataset | TNFR | local disagreement | graph TV | local entropy | label-prop residual | errors / N |
|---|---|---|---|---|---|---|
| WDBC (breast cancer) | 0.9590 | 0.9493 | 0.9493 | 0.9345 | 0.9563 | 12 / 569 |
| Iris | 0.9860 | 0.9820 | 0.9820 | 0.9695 | 0.9850 | 7 / 150 |
| Digits | 0.6984 | 0.6962 | 0.6962 | 0.6980 | 0.8200 | 146 / 1797 |
| Wine quality (red) | 0.8739 | 0.8623 | — | — | 0.9423 | — |
The reported rankings include stronger label-propagation performance on digits and wine red. Small differences on WDBC and Iris lack an uncertainty statement here, and the low error counts matter. These summaries establish neither superiority nor a non-noise mechanism without a pinned replay and appropriate uncertainty analysis.
Reported grid-frequency trends¶
The historical power-grid report gives a classical variance trend (Kendall-τ ≈ 0.255) slightly stronger than the strongest TNFR channel (Φ_s, τ ≈ 0.184). Both reported correlations are below 0.26. Those statistics do not establish the physical mechanism behind the signal or a general limit of either method.
Variance and autocorrelation are necessary comparators for this preparation. Multichannel data permit additional graph-local observations; their availability does not itself establish added predictive value.
Reported EEG Eye State ranking¶
Historically reported discrimination (ROC-AUC) of eyes-open versus eyes-closed labels; the table does not supply a pinned replay or uncertainty interval:
| Indicator | AUC |
|---|---|
| phase dispersion (baseline) | 0.641 |
| ξ_C (TNFR) | 0.615 |
| Kuramoto R (baseline) | 0.559 |
| mean PLV (baseline) | 0.530 |
These reported AUCs order the selected scores in this preparation. The gap of about 0.026 between phase dispersion and ξ_C does not establish statistical equivalence or superiority. Neither this ranking nor the formulas alone prove that ξ_C adds information after controlling for the other observables.
The maintained multichannel adapter now reports orientation-free discrimination
max(AUC, 1-AUC) on its supplied labels. The old table has no pinned replay
establishing its exact orientation convention, preprocessing or estimator branch;
do not reconstruct those details from current code or treat its values as a
reserved forecast.
Reported CHB-MIT findings (external empirical arm)¶
The prior report attributes these findings to the separate TNFR-IA empirical arm using the public CHB-MIT scalp-EEG corpus. No pinned external commit or complete replay envelope is supplied here.
| Confrontation | Result | Reading |
|---|---|---|
Local phase tetrad \|∇φ\|, \|K_φ\| vs seizure/interictal |
cross-patient (leave-one-patient-out) AUC ≈ 0.66–0.67, higher at ictal | reported association with labels; independent replay remains pending |
Global Kuramoto R vs seizure |
AUC ≈ 0.49 (chance) | reported lack of label discrimination in this preparation; does not by itself identify a local mechanism or exclude global synchrony effects |
νf (single structural frequency) |
tracks the measured dominant EEG frequency (Spearman ρ ≈ 0.86); fingerprints subjects | reported association does not identify structural mobility with physical oscillation frequency |
Single-νf conservative-wave forecaster (few-shot, leave-one-patient-out) |
ties a strong Gaussian process and a stabilised DMD/Koopman model (within ~1%); beats a parameter-matched neural net | reported accuracy with 1–3 model parameters and ~20× lower cost than the neural TNFR; physical identification and replay remain pending |
Face diagnosis (verify_overdamped_projection at the data-fitted γ) |
reported under-damped classification | a fitted auxiliary-model result, not certification of physical conservative dynamics |
| Pre-ictal (5 min before onset) tetrad | every magnitude reported at chance | no predictive evidence in the reported five-minute preparation; concurrent association does not establish a general detection or prediction rule |
Reported cross-subject confrontation (PhysioNet eegmmidb)¶
The earlier summary describes a second confrontation through the shipped instrument
(confront_signal, historical data-fitted face) on the PhysioNet eegmmidb eyes-open
(R01) vs eyes-closed (R02) baselines — 10 subjects, 64-channel, drift removed
(fft_bandpass 1–40 Hz), paired across subjects with a two-sided binomial sign
test. Data are not bundled (fetched to a scratch dir); the run is
not independently reproducible from the supplied description alone: exact
run scripts, subject/file hashes, preprocessing and evaluation records remain
to be pinned and replayed.
| Confrontation (open→closed) | Result | Reading |
|---|---|---|
Modal root classifier (AR(2) on L_sym) |
reported WAVE / under-damped in 18/20 conditions | descriptive classification; the reported thermal comparison (97–99% diffusive labels) also needs acquisition/replay provenance and is not a physical certificate |
K_φ (phase curvature, local) |
rises in 10/10 subjects (paired dz ≈ +1.1, sign-test p ≈ 0.002) | reported cross-subject association in this sample; independent replay remains pending |
\|∇φ\| (phase gradient, local) |
rises in 9/10 (p ≈ 0.021) | reported cross-subject association, pending replay |
C_static (pressure snapshot) / Q (quality factor) |
both rise (8/10 each) | reported trends; C_static is pressure-only (p ≈ 0.11), not dynamic total C(t) or proof of a sharper physical rhythm |
Global Kuramoto R |
falls in 9/10 (dz ≈ −0.9) | reported direction differs from the local-field trends in this sample; no general diagnostic ordering follows |
Collective pulse ω₀=√λ₂ |
structural ω₀ ∈ [0.12, 0.27] vs measured 1–11 Hz (Spearman ≈ −0.3); flat open↔closed (dz ≈ −0.2, p = 1.0) |
graph-derived auxiliary wave scale, without an identified conversion to temporal Hz; the reported absence of discrimination does not prove topology is state-insensitive |
Reported correction. The cross-subject summary gives the opposite local-field
direction to the earlier single-subject reading (eyes-closed ⇒ |∇φ| down).
If replay confirms those results, they would reject that directional
generalization for this preparation. A local/global diagnostic difference does
not establish a travelling wave, show that one observable has the correct sign
and the other the wrong one, or identify a TNFR mechanism. Fixing k in a PLV
neighbor graph does not fix which neighbors its data-dependent construction
selects. Both acquisition provenance and the actual constructed graphs are
needed to interpret the reported result.
Reported same-window diffusion comparisons¶
These scores concern the supplied isolated EPI-diffusion realization; the bare nodal identity alone does not select that pressure law.
The reported in-sample diffusion-direction comparison used
nodal_prediction_skill:
the predictor x̂(t+1) = x(t) − c·L_rw·x(t) fits a single diffusion step
c = ν_f·dt and scores the one-step increment variance explained beyond
persistence, against a per-channel AR-1 baseline. Both models were fitted and
scored on the same supplied data. The table preserves reported scores; it does
not establish out-of-sample forecasting.
| Signal | nodal 1-step skill | AR-1 | Reading |
|---|---|---|---|
| synthetic graph diffusion | +0.16–+0.21 (recovers c) |
+0.03 | reported synthetic same-window recovery; not rerun here or physical validation |
| real EEG (eyes-closed) | +0.02 | +0.05 | reported same-window AR-1 score exceeds the selected diffusion-direction fit |
| real thermal field (60–120 min steps) | +0.015–+0.029 | +0.013–+0.025 | reported same-window scores are close; this comparison alone does not certify a diffusive physical regime |
Limits of these reports¶
- Reported baseline comparison. The external wave-model summary claims accuracy near GP and stabilized DMD. That comparison needs a replay manifest; it does not derive the auxiliary wave from canonical nodal evolution or establish physical meanings for its fitted parameters.
- Concurrent association, not an admitted predictor. The CHB-MIT summary reports concurrent AUC ≈ 0.67 and chance-level pre-ictal results. No clinical detection/prediction capability follows without replay and the appropriate evaluation protocol.
- Modest reported margins. Relative-MSE differences near 1% and the listed AUCs need uncertainty and information-matched comparisons; these summaries establish no competitive-performance claim.
- Limited cross-subject preparation. The eegmmidb report covers ten subjects in one eyes-open/closed comparison. Its local/global trends, if reproduced, would remain observations in that sample, not identification of spatial alpha dynamics or of the complete nodal law.
- Modal diagnosis has limited scope. Adequate samples do not turn AR(2) root classification into a certificate of diffusion, conservative dynamics or TNFR correspondence. Failure and unresolved fits now retain explicit statuses, and the old physical face labels are no longer generated.
- Reported associations need provenance. Label discrimination, model-fit scores and timing claims remain externally reported until their precise experiments are independently replayable. Even successful replay would establish only the measured association or restricted prediction.