Recursive causal structure across states of consciousness

Xiaoba Yang · Independent researcher · xiaobayang01@gmail.com

A preregistered, open, and so far underpowered test of one narrow question: in the same person, does the directed causal graph of the cortex contain a different proportion of short directed cycles when conscious than when not?

This repository is the working record of that test — preregistrations written before the data were opened, the analysis code, the results, and the mistakes found and corrected along the way. Most documents are in Chinese; this page summarises them.

Status, in one paragraph

Three studies so far. The first two failed for reasons that turned out to be methodological, not empirical, and are documented as such. The third, on intracranial single-pulse electrical stimulation data (PRIOS, OpenNeuro ds004370), is the first with a design that can actually answer the question — and it is underpowered. Six usable participants. The effect ran in the hypothesised direction at every one of six graph density thresholds tested, in every version of the analysis, but the preregistered primary test did not reach significance (p > 0.2 in the current committed results). The verdict is not confirmed, not falsified, underpowered. The bottleneck is sample size; I am seeking access to larger intracranial datasets with stimulation in more than one state of consciousness.

The three studies

	Data	Contrast	Method	Verdict
01 · P1–P4	ds003768, scalp EEG, sleep	wake vs sleep stages	integration index `I_global` on 13 participants / 660 epochs	Predictions tested; see VER-001
02 · Panel C	ds005620, scalp EEG (19 ch)	wake vs propofol sedation	conditional transfer entropy → cycle proportion	Untestable, not falsified. The binary cycle metric saturates at 1.0 for any random graph above density ≈0.23; all real graphs sat above it. Diagnosed in DIAG-002
03 · PRIOS	ds004370, intracranial ECoG + single-pulse stimulation	wake vs propofol anaesthesia	intervention-defined directed graph, density-matched, weight- and density-preserving permutation null	Not confirmed, not falsified, underpowered (n = 6). Full report: RPT-001

Why the switch from study 2 to study 3 matters: a correlational edge ("B's future is predictable from A's past") is exactly what slow-wave-dominated signals inflate. A stimulation edge ("stimulate A, B responds") is not. Study 3 is the first version of this question where the causal graph comes from an intervention.

What the PRIOS result is and is not

Preregistered before any two-condition comparison was seen: hypothesis direction, primary metric, gates, and falsification criteria (PRE-003).
Detection threshold calibrated against the pre-stimulus false-positive rate, not chosen by hand (CAL-001).
Graphs density-matched between conditions so the comparison cannot be driven by edge count; null model preserves the weight multiset and — after a correction, see below — the set of electrodes that can emit edges.
Direction consistent across all six density thresholds in every analysis version.
Primary paired test not significant. With n = 6 the best two-sided p a Wilcoxon signed-rank test can return is 0.031, so "not significant" here is a statement about power, not evidence of absence.

Corrections made in the open

Two defects in my own analysis were found after the first PRIOS report and fixed with the old and new results kept side by side (DEF-001, defect log):

The null model let never-stimulated electrodes emit edges; the observed graph cannot. A pure random graph with no structure reproduced the "effect" this produced. Fixed by restricting the null's candidate edge slots to stimulated sources.
The node set was all common channels; the preregistration said dual-role (stimulated and recorded) electrodes. Fixed to match the preregistration.

Each correction moved the primary p away from significance (0.06 → 0.09 → 0.22). That is the honest direction: the earlier, better-looking numbers were partly artefacts. A third issue — run-to-run nondeterminism from Python hash ordering feeding a shared random generator — was fixed with a regression test that fails when the fix is reverted.

What is in this repository

docs/studies/     three studies: preregistration → calibration → verdict → report
docs/methods/     cross-study methodological diagnoses
docs/data-sourcing/数据源清单.md   which public repositories were searched, and how
registers/        document register and defect log (living tables)
code/core/        ce.py — cycle-proportion metric and null model (the only copy)
code/pipelines/   per-study analysis pipelines
code/simulations/ one-off simulations backing specific claims in the documents
results/          derived results (JSON/CSV/figures); raw data are not included

Every document under docs/ carries a YAML header (id, type, status, date). Filenames are Chinese; the register maps IDs to files.

Reproducing

python -m venv .venv && .venv/bin/pip install -r requirements.txt
# PRIOS raw data: OpenNeuro ds004370 (CC0), expected at data/prios/ (symlink is fine)
PRIOS_ROOT=/path/to/ds004370 .venv/bin/python code/pipelines/prios/prios_分析.py
.venv/bin/python code/pipelines/prios/test_reproducibility.py   # determinism check

Disclosures

AI assistance. Analysis code, simulations, and most documents were written with an LLM coding assistant (Claude Code) under my direction. The theoretical predictions tested in study 1 came out of a dialogue with an LLM (Tencent Yuanbao); documents from that study refer to it by name.
Preregistrations are self-archived. They carry the dates on which they were written, but no third-party timestamp. This repository's git history begins on 2026-09-02, after all three studies, so it cannot independently establish that the preregistrations preceded the results. Future studies will be registered externally.
No affiliation. Independent, unfunded, single-person project.
Language. Working documents are in Chinese. Section and table headers in the main report are literal enough to follow with machine translation; the numbers are the numbers.

Data sought

The question needs, within the same participants, intracranial recordings with stimulation delivered in both a conscious and an unconscious state. A sweep of eleven public repositories (documented in the data-source register) found exactly one such public dataset — the one already used. Access requests are pending with the groups holding three candidate datasets. If you hold data of this kind and would consider sharing it under a data-use agreement, I would be glad to hear from you.