Alice is a two-channel cognitive instrument built on daily writing.
It measures two orthogonal things at once: how you type and what you write. Each channel has its own null model, its own reproducibility guarantee, and its own methodological defensibility. Together, they produce a longitudinal cognitive record that no other architecture can sustain.
Three things that must be one thing
Alice is simultaneously a personal practice, a measurement instrument, and a research platform. These are not three products bundled together. They are three descriptions of the same architecture, and the architecture only works if all three are true at once.
One question per day. You write. The practice is valuable on its own terms. No dashboard, no score, no feedback loop. The writing is the thing.
Every session captures keystroke-level temporal data and submitted text across two channels. The behavioral channel measures motor/dynamical execution via a native Rust engine. The semantic channel measures propositional output via deterministic text analysis. Neither channel surfaces its measurements.
The longitudinal record accumulates over years. Self-referential baselines per channel. A reconstruction adversary that tests which behavioral dimensions are irreducible. The dataset that keystroke-cognition research needs but cannot produce from lab visits.
The personal practice sustains the measurement. The measurement produces the research data. The research validates the measurement. Remove any one and the others collapse.
Two channels, matched rigor
Keystroke timing, motor execution, nonlinear dynamics, pause architecture, revision behavior. Computed in a native Rust engine with bit-reproducible output. The null model is a reconstruction adversary (the ghost) that generates synthetic keystroke streams from your statistical profile. What remains in the residual is what the profile cannot explain.
Idea density, lexical sophistication, epistemic stance, cohesion, emotional valence. Computed in the same Rust engine as the behavioral channel via deterministic word-list methods — no LLM is invoked anywhere in the signal pipeline. Embeddings generated by a self-hosted model with SHA-256 archived weights and bit-reproducible inference. The null model is your own longitudinal baseline, topic-matched against prior sessions.
These channels measure different things at different timescales with different null models, but they share the same commitment: determinism, archivability, and reproducibility. A measurement instrument cannot have two standards of rigor.
What makes this structurally different
Alice does not. The black box stays black. Surfacing analysis contaminates the signal and creates optimization behavior. The instrument must be invisible to remain valid.
Alice does not. The assessment is embedded in a daily practice the participant already wants to do. No lab. No appointment. No awareness of being tested.
Alice captures both. Keystroke timing alone cannot distinguish retrieval difficulty from vocabulary contraction. You need the temporal process and the linguistic output, jointly, over time.
Alice enforces an unmediated contamination boundary. No autocomplete, no suggestions, no AI in the writing path. The generative AI is on the instrument side, never the participant side. This produces a clean cognitive corpus at exactly the moment when clean corpora are disappearing from the world.
Who this is for
You want a daily writing practice that is private, unmediated, and accumulates depth over time without performing measurement back at you. You want to write without being watched, and trust that the watching serves you in the long run.
You need longitudinal keystroke-cognition data from fluent typists, with intra-individual baselines, ecological validity, and dual-channel capture. No existing instrument produces this dataset under a matched reproducibility standard across both motor and semantic channels.
You're building passive cognitive assessment and need a data source that combines process-level temporal signals with linguistic content over months and years. Alice is the only architecture sustaining daily ecological data collection at this granularity, with archived provenance for every measurement.
You're concerned about AI mediation eroding cognitive capacities without detection. Alice captures unmediated baselines from the last cohort that will produce them at scale. The contamination boundary is the cryptographic proof that the corpus is what it claims to be.
Where this goes
Single-user prototype. Both channels instrumented. Six signal families computing daily. Reconstruction adversary validated on production data. Embedding model archived with SHA-256 identified weights. Contamination boundary attested per session.
Paper One: the methodological statement that defines the two-channel instrument. Multi-user architecture. Calibration protocol validation. First published longitudinal keystroke-cognition results from ecological data.
Research partnerships. Cohort studies using Alice as the reference instrument. Validation against clinical cognitive batteries. The semantic channel validated against years of accumulating data under the already-archived methodology.
The reference instrument for longitudinal cognitive measurement through writing. Pre-clinical detection of cognitive trajectory shifts years before annual screening would catch them. The cognitive health equivalent of what continuous glucose monitors did for metabolic health.
The window for capturing unmediated cognitive baselines is closing. Alice is the instrument designed to collect them while they still exist.
You're on the list.