Anthony Guzzardo April 2026 (v1, draft)
Alice collects two writing sessions per day from a single participant: a journal session (reflective writing in response to a generated question about the person's inner life) and a calibration session (neutral writing in response to a randomly selected prompt from a pool of 303 knowledge-telling tasks). The two sessions form a within-person, within-day matched pair. The delta (journal value minus calibration value) on any behavioral or linguistic signal isolates what the reflective question provoked beyond the person's same-day neutral writing baseline. The same-day pairing is motivated by evidence that 76-89% of stress response variance is within-day (Toledo et al. 2024), making same-day controls more appropriate than between-day baselines for isolating state-level provocation effects.
The calibration prompt pool contains 303 prompts across 12 cognitive task categories: observation (describe what's around you), recent memory (what did you do/eat), routine and process (walk through how you do X), object explanation (describe X to someone who's never seen one), procedural instruction (explain how to X), constrained description (describe X without using words Y/Z), fluent generation (name as many X as you can), sensory description (taste, smell, texture, sound), spatial memory (describe a place from memory), perspective shift (describe from an unusual vantage), and enumeration (count and describe specific things). All prompts are designed to elicit knowledge-telling, not knowledge-transforming. Prompts are selected without replacement until the pool is exhausted, then oldest-used prompts are recycled with maximum temporal spacing.
In the first 12 day-pairs (April 13-24, 2026), journal sessions occur in the morning or midday (mean hour 9.7, SD 5.9) and calibration sessions occur in the evening (mean hour 19.9, SD 3.6). Journal precedes calibration in 83% of pairs (10/12), with a mean within-pair time gap of 10.3 hours (SD 7.7). The ordering is not randomized; it reflects the participant's natural daily routine.
68 signals are computed per session across four families: dynamical (35 signals, Rust-computed from IKI series: permutation entropy, DFA, MF-DFA, RQA, transfer entropy, PID, DMD, causal emergence, criticality), motor (13 signals, Rust-computed from keystroke timing: sample entropy, ex-Gaussian distribution, motor jerk, tempo drift, compression ratio, digraph latency, hold-flight correlation), process (9 signals, Rust-computed from event log replay: pause location, burst classification, vocabulary expansion, phase transition, strategy shifts), and semantic (11 signals, Rust-computed from response text: idea density, lexical sophistication, epistemic stance, integrative complexity, cohesion measures, discourse coherence, text compression ratio). All signals are classified by ergodicity (Mangalam et al. 2022): signals derived from fractal and entropy measures are safe for longitudinal trending in N=1 designs, while IKI-derived means and motor timing means are valid per-session but unreliable as trend estimators for non-ergodic multiplicative processes. A fifth family (cross-session, 11 signals) is structurally excluded from the delta analysis. Cross-session signals (self-perplexity, NCD, vocab recurrence, digraph stability, text network metrics) are computed by comparing the current session against the participant's prior journal history; calibration sessions lack a comparable calibration-only history to reference, so the comparison is undefined and the family is excluded by design.
Journal sessions are substantially longer than calibration sessions across every structural metric. Mean duration: 240s vs 97s. Mean IKI count: 1152 vs 470. Mean word count: 146 vs 56. Mean total characters typed: 972 vs 340. The ratio is approximately 2.5:1 across all production measures.
This asymmetry is critical for dynamical signals that depend on IKI series length. Many require minimum series lengths (RQA: 100+, MF-DFA: 256+, DMD: 100+), and their statistical properties change with series length even when the underlying process is identical. Normalized measures (LZC, PID) are particularly sensitive because their normalization constants depend on series length.
Journal sessions cluster in the morning (mean hour 9.7); calibration sessions cluster in the evening (mean hour 19.9). The gap is systematic, not random: it reflects the participant's daily routine of journaling in the morning and calibrating before bed. This introduces a circadian confound for any signal sensitive to time-of-day effects (fatigue, motor slowing, cognitive depletion).
Journal precedes calibration in 83% of pairs. The calibration session occurs after a full day of cognitive activity. This confounds any motor signal comparison: calibration motor output reflects end-of-day motor state, not neutral-condition motor state.
Despite the structural differences above, keystroke-level pacing is comparable. IKI mean: 169ms vs 179ms (p=0.51). IKI CV: 1.5 vs 1.5 (p=0.48). Chars per minute: 273 vs 252 (p=0.42). The person types at similar speed in both conditions; the difference is in how long and how much they type, not how fast.
For each signal, compute across all available day-pairs:
Sign consistency. Proportion of days where the delta (journal minus calibration) has the same sign. Binomial test against 0.5 for significance. Threshold: > 0.75.
Hedges' g_z with BCa bootstrap CI. Paired effect size with small-sample correction (Hedges, 1981). 1000-resample bias-corrected accelerated bootstrap confidence interval. Threshold: 95% CI excludes zero.
Calibration stationarity. Mann-Kendall trend test on the calibration-session time series for each signal. Signals where the calibration baseline is drifting (p <= 0.05) are flagged: the provocation delta on those signals is confounded with baseline drift. Threshold: p > 0.05 (stationary).
A signal passes Phase 1 if all three criteria are met simultaneously.
Length-matched recomputation. For each pair with raw keystroke streams, truncate the journal stream to the calibration stream's keystroke count (first N keystrokes of journal where N = calibration count). Recompute the signal via the Rust engine on the truncated stream. Rerun Phase 1 criteria on the truncated deltas. A signal passes if its sign consistency, effect size, and CI survive truncation.
Time-of-day regression. Regress the signal delta against the within-pair time-of-day difference (journal hour minus calibration hour). If the regression is significant (p < 0.05), the provocation effect covaries with circadian timing and cannot be attributed to the question alone.
Pause and burst structure check. Compare pause duration distribution (median, IQR, 90th percentile), burst length distribution, and inter-burst interval distribution between length-matched journal and calibration streams to verify structural comparability at equal keystroke counts.
The number of available pairs varies across signals because each signal has minimum data requirements that not every session clears (e.g., MF-DFA requires 256+ IKIs, ex-Gaussian fitting requires 50+ flight times, discourse coherence requires 5+ sentences with embeddings), and because keystroke stream capture was not available for the earliest sessions (April 13-17).
4 of 68 signals met all three criteria:
| Signal | Family | g_z | Direction | Sign consistency | Ergodicity |
|---|---|---|---|---|---|
| Integrative complexity | Semantic | +1.211 | journal > | 6/6 (100%) | safe |
| Lempel-Ziv complexity | Dynamical | -0.800 | calib > | 6/7 (86%) | safe |
| PID synergy | Dynamical | -0.575 | calib > | 6/7 (86%) | safe |
| Temporal irreversibility | Dynamical | -0.490 | calib > | 6/6 (100%) | safe |
All four are ergodicity-safe. Zero ergodicity-unsafe signals passed. Zero LLM-derived signals passed. Two calibration baselines were drifting: DMD dominant frequency (falling, p=0.04) and lexical sophistication (rising, p=0.02).
Length-matched recomputation was applied to the three dynamical signals (integrative complexity is text-based and not affected by IKI series length). All three collapsed:
| Signal | Original sign | Truncated sign | Original CI | Truncated CI | Verdict |
|---|---|---|---|---|---|
| Lempel-Ziv complexity | 0.86 calib > | 0.57 calib > | excludes zero | spans zero | artifact |
| PID synergy | 0.86 calib > | 0.57 calib > | excludes zero | spans zero | artifact |
| Temporal irreversibility | 1.00 calib > | 0.67 journal > | excludes zero | spans zero | sign flipped |
Temporal irreversibility is the clearest artifact: 100% sign consistency reverses direction entirely under length-matching. LZC and PID synergy lose sign consistency and CI significance.
Time-of-day regression on integrative complexity: r=0.315, R^2=0.099, p=0.458. Not significant. The IC provocation effect does not covary with the 10-hour circadian gap. (Note: the regression has low power at n=7; a weak TOD effect cannot be excluded. See Limitations.)
When journal streams are truncated to calibration length, the two conditions produce comparable writing process signatures. Pause count: 4.3 vs 3.7 per session. Median pause duration: 2929ms vs 2695ms. Pause IQR: 912ms vs 1309ms. Median burst length (IKIs between pauses >= 2000ms): 67 vs 56. Burst IQR: 89 vs 117. Mean inter-burst interval: 2968ms vs 2908ms.
This is a non-trivial finding. The 2.5x length asymmetry in total production coexists with equivalent keystroke-level pacing (IKI mean 169ms vs 179ms, p=0.51; IKI CV 1.5 vs 1.5, p=0.48; chars per minute 273 vs 252, p=0.42) and equivalent within-session microstructure (pause frequency, pause duration, burst length, inter-burst gaps). The reflective question does not change the motor-cognitive process of typing. It changes how long the process runs and what linguistic content it produces. The structural production machinery is the same; the semantic output is different. This dissociation between process equivalence and output divergence is what makes the integrative complexity finding interpretable: the IC difference is not driven by a different writing process but by a different cognitive engagement with the material.
Integrative complexity is the sole signal surviving the full pipeline. It measures contrastive and integrative connective density per sentence (Suedfeld & Tetlock). The reflective question reliably produces more multi-perspective, nuanced writing than the neutral prompt. Effect size g_z = +1.211 (large). Sign consistency 100% across 6 pairs. Not confounded with time of day, series length, or calibration drift.
Small N. 6-12 pairs depending on signal availability. Even the surviving signal's 100% sign consistency at n=6 yields a minimum binomial p of 0.016, which is suggestive but not definitive. The bootstrap CI is wide ([+0.60, +4.60]).
Single participant. All findings are idiographic. The provocation effect of the reflective question on integrative complexity may be specific to this person's writing style, cognitive profile, or relationship to the instrument.
Fixed ordering. Journal always precedes calibration. The time-of-day regression did not find a significant relationship for integrative complexity, but the test has low power at n=7. An order confound cannot be ruled out with certainty.
Fatigue confound. Calibration occurs after a full day. Motor signals in calibration may reflect end-of-day motor state. This is partially addressed by the finding that keystroke-level pacing (IKI mean, CV, chars per minute) is comparable between conditions, but subtle fatigue effects below the resolution of summary statistics cannot be excluded.
Series-length sensitivity. The confound battery revealed that three of four Phase 1 survivors were series-length artifacts. This underscores that any future analysis of dynamical signal deltas must control for series length. The 2.5x length asymmetry is a structural feature of the current design, not an anomaly.
Integrative complexity is a computational proxy. The IC signal used here measures contrastive and integrative connective density per sentence, computed deterministically from word lists. This approximates but does not reproduce the trained-rater integrative complexity construct from Suedfeld and Tetlock (1977), which involves holistic paragraph-level scoring of differentiation and integration by trained coders. The word-list proxy captures connective usage (a necessary component of IC) but not the conceptual structure that human raters assess. A high connective density score can occur without genuine multi-perspective integration if the writer uses connectives as filler or in formulaic constructions. The finding that the reflective question increases connective density relative to neutral writing is robust; the claim that this constitutes increased integrative complexity in the Suedfeld-Tetlock sense is weaker.
No comments yet.