Neurophenomenology pairs disciplined first-person reports with simultaneous physiological recording, so that neither channel has to be trusted alone. This repository builds that as personal instrumentation rather than laboratory apparatus: a reproducible pipeline for a single person over years. The question it was built to answer is a narrow one. The architecture is not — your question will differ, the instrumentation problem will not.

What this is

• A momentary sampling layer (randomised prompts, structured capture) that produces timestamped subjective annotations
• Adapters for continuous physiological streams — cardiac, electrodermal, electroencephalographic — normalised to a single schema
• A passive context layer (location, activity, environment) collected without user action
• An append-only local store with an event-locked analysis layer on top

Output is a queryable record in which any physiological window can be matched to what the subject reported experiencing at that moment.

The problem it's built around

Retrospective self-report is a biased sample of experience. What you believe your days consist of and what your days consist of are different datasets, and the difference is systematic rather than random. Measuring that difference requires both sides to be recorded independently — the momentary report and the later memory of it — which is an instrumentation problem before it is a psychological one.

Design principles

• Labels before sensors. Physiological data without concurrent subjective annotation is close to uninterpretable. The momentary sampling layer is what makes every other stream analysable, not one source among several.
• One clock. Streams arrive at rates from 256 Hz to six events per day. Every record carries a canonical UTC timestamp assigned at acquisition, not at import.
• Raw data is immutable. Ingestion is append-only. Transformations produce new derived layers and never overwrite. Analysis methods will change; unrecorded signal will not come back.
• Exploratory and confirmatory phases are separated. Collection runs in bounded waves. Hypotheses are written down between waves and tested on data collected afterwards. Without this, a system with fifty channels generates significant findings from noise indefinitely.
• Hardware is configuration, not architecture. Devices are adapters behind a common interface. The device list in hardware/ is one working instantiation, not a requirement.

Architecture

Everything reduces to two tables:

• signals — (timestamp, source, channel, value): continuous streams
• events — (start, end, type, annotations): interactions, sessions, sampling prompts, state transitions

Storage is append-only Parquet in a local lake, queried through DuckDB. Acquisition uses Lab Streaming Layer where devices support it. Nothing requires a server; nothing requires a vendor cloud.

The core analysis is event-locked: a baseline window before a marked event, a response window during and after, and per-event metrics — autonomic change, recovery time, concurrent subjective annotation — compared across event types.

Status

Early. See docs/protocol.md for the measurement protocol and hardware/ for the current instantiation.