Participant-disjoint does not mean deployment-valid preprocessing¶

Experimental longitudinal transformation provenance, 21 October 2026.

A held-out participant may be disjoint from training while a normalization procedure uses that participant's complete future recording. This can be legitimate for retrospective descriptive work, but is not valid evidence for first-use prediction.

The research-only certify_preprocessing_availability() consumes actual fit-observation lineage: transform ID, source participant and source time; plus prediction time, target participant and declared training participants. It cannot detect unrecorded internal uses of data.

Synthetic falsification test¶

import pandas as pd
from research.october_2026.information_availability import certify_preprocessing_availability

fits = pd.DataFrame({
    "transform_id": ["z"] * 3,
    "source_participant": ["heldout"] * 3,
    "source_time_s": [0, 5, 10],
})
predictions = pd.DataFrame({
    "transform_id": ["z"],
    "target_participant": ["heldout"],
    "prediction_time_s": [4],
})
audit = certify_preprocessing_availability(
    fits, predictions, training_participants=["train"], mode="online")
print(audit[["future_target_fit_uses", "deployment_valid"]])

The same logged observations may be disclosed as retrospective, in which case first-use validity is deliberately false.

Plot recipe: time of information availability¶

import matplotlib.pyplot as plt
fig, ax = plt.subplots(figsize=(6, 2.5))
ax.scatter([0, 5, 10], [1, 1, 1], s=100, label="Scaler fit observations")
ax.axvline(4, color="red", linestyle="--", label="Prediction at t=4")
ax.set(xlabel="Synthetic recording time", yticks=[], ylim=(0, 2))
ax.legend()
fig.tight_layout()

The certificate supports four declared regimes: population training statistics, prespecified prior baseline, online past-only normalization, and retrospectively fitted whole-recording normalization. It is not an independent audit of hidden vendor preprocessing, smoothing, imputation, subject identity correctness, or calibration quality. Those require independent tests and external evidence before release.

Publication and scientific qualification boundary¶

This source-repository research method is not an installed public PyPI 1.2 API or an independent certificate of preprocessing validity. The illustrative figure is synthetic. Declared training/first-use source observations and future participant information must be validated against real preprocessing provenance before any prospective leakage-free claim. The research-to-release issue tracks independent empirical qualification. This article does not resolve the separate W3 Bayesian simultaneous-band calibration blockers and does not authorize an eyetrajectoriespy 1.2 release.