Longitudinal keystroke-cognition data from ecological daily writing.

Alice is a running measurement instrument that captures dual-channel process data (keystroke timing + linguistic content) from daily naturalistic writing, with intra-individual baselines, same-day calibration, and participant-blind measurement. No existing instrument produces this dataset. We are looking for research partners to help validate it.

What the instrument captures

Behavioral channel

Key-down and key-up timestamps at millisecond precision. Full temporal microstructure: inter-key intervals, hold durations, flight times, pause architecture, deletion sequences, burst segmentation. Captured from unmediated character-by-character input. No autocomplete, no predictive text. Null model: a five-variant reconstruction adversary that generates synthetic keystroke streams from your statistical profile and measures the residual.

Semantic channel

Final submitted text. Deterministic linguistic analysis: idea density, lexical sophistication, epistemic stance, integrative complexity, cohesion, emotional valence, text compression ratio. NRC emotion word densities. MATTR vocabulary diversity. Null model: longitudinal self-referencing baselines, topic-matched against prior sessions via archived embedding model (Qwen3-Embedding-0.6B, SHA-256 identified weights, FP32 CPU inference, bit-reproducible).

The two channels are analyzed jointly. A decline in lexical diversity with stable production fluency means vocabulary contraction. The same decline with slowed production means retrieval difficulty. You need both to distinguish them. Each channel has its own null model and its own reproducibility guarantee, under a matched standard of mathematical rigor.

Signal pipeline

Six signal families computed per session. Native Rust engine for numerical estimation. All methods cited with validation status documented on the methodology page.

Dynamical Permutation entropy spectrum (orders 3-7), DFA alpha, MF-DFA (spectrum width, asymmetry, peak alpha), temporal irreversibility, spectral analysis (PSD slope, LF/HF ratio, peak frequency), ordinal statistics (statistical complexity, forbidden patterns, weighted PE, LZC), OPTN (transition entropy), RQA (6 metrics), recurrence networks (transitivity, clustering, assortativity), causal emergence (EI, CEI, PID), criticality (branching ratio, avalanche exponent), DMD (4 modes), pause mixture model, transfer entropy (KSG)
Motor Ex-Gaussian parameters (MLE/EM), Fisher trace, sample entropy, multiscale entropy / complexity index, motor jerk, lapse rate, tempo drift, digraph latency profile, IKI compression ratio, hold-flight rank correlation
Process Pause classification (within-word, between-word, between-sentence), R/I burst segmentation, abandoned thought detection, phase transition point
Semantic Idea density, lexical sophistication, epistemic stance, integrative complexity, deep cohesion, emotional valence arc, text compression ratio, discourse coherence (global, local, ratio, decay slope)
Cross-session Self-perplexity (personal trigram model), motor self-perplexity, NCD trajectory (lags 1/3/7/30), vocabulary recurrence decay, digraph stability, text network density, text network communities, bridging ratio
Behavioral state 7D state vector (fluency, deliberation, revision, commitment, volatility, thermal, presence), PersDyn dynamics (baseline, variability, attractor force), empirical coupling matrix

Current state

Scale n=1 since April 2026. Single-user prototype. Multi-user architecture in development.
Database PostgreSQL 17 + pgvector. 32 tables. 120+ metrics per session. 512-dimensional embeddings (Qwen3-Embedding-0.6B, self-hosted, HNSW-indexed).
Signal engine Native Rust via napi-rs. Single source of truth. Typed error variants (InsufficientData, ZeroVariance, DegenerateValue).
Calibration Same-day within-person controls. Neutral writing prompts. Calibration-relative deviation for all metrics. Life context extraction (sleep, stress, physical state).
Papers 2 published preprints. 3 in progress. Establishing the problem space, not yet reporting longitudinal results.
Reproducibility Bit-reproducible signal engine (pinned Rust toolchain, Neumaier compensated summation, deterministic iteration). Reconstruction residuals store PRNG seed, profile snapshot, and corpus hash for regeneration. Embedding weights archived by SHA-256 hash with versioned inference environment. CI enforces two-clean-build reproducibility on every PR.
Contamination No AI in the writing path. Unmediated character-by-character input. Contamination boundary attested per session with versioned specification, audited code paths, and git commit hash. Clean cognitive corpus by design.
Validation Not yet possible at n=1. The instrument is designed to produce the data that would enable validation. Reconstruction validity (computable from n=1) characterizes what the instrument captures; external-criterion validation requires longitudinal outcome data.

What partnership looks like

Data access

Anonymized longitudinal process + content data under data sharing agreement. Full signal pipeline output. Raw keystroke streams available for custom feature extraction.

Custom signal extraction

If your research requires specific features not in the current pipeline, we can implement them. The architecture is designed for extensibility. Every signal family — numerical and linguistic — runs in a single Rust engine; no LLM is invoked anywhere in the signal pipeline.

Protocol adaptation

The question schedule, calibration protocol, and session structure can be adapted for specific research populations or research questions.

Joint publication

We are interested in co-authoring with domain experts. The instrument produces the data. You bring the clinical or theoretical framework. The validation comes from the collaboration.

The gap between "running" and "validated" is the work that remains. If your research could use longitudinal keystroke-cognition data that no other instrument produces, let's talk.