The Prosoche Protocol v1.0.

Plain-language summary

AI capability is compounding quickly while operator cognition remains constrained. This protocol tests whether prosoche, a Stoic attention discipline centered on impression-checking and deliberate assent during action, performs differently from mindfulness and productivity controls in AI-supervisor work. The design is a four-arm RCT with pre-registered hypotheses, matched intervention dose across active arms, and a pre-specified primary contrast: prosoche vs mindfulness on NASA-TLX change during a standardized AI-mediated task.

• Population: professional knowledge workers using AI tools at least 10 hours/week.
• Intervention: four-week protocol with distributed daily micro-practices and weekly matched calls.
• Primary inference: cognitive-load reduction under AI-supervised work, not generic wellness outcomes.
• Key contribution: methodologically strict active-control design intended to stay publishable under both positive and null outcomes.

1. Background and rationale

1.1 The bandwidth gap

The protocol grounds itself in a capability-vs-operator mismatch: frontier compute and agent horizons increase faster than human conscious processing, making operator attention the bottleneck. The trial builds on The Bandwidth Gap thesis and prior evidence on interruption costs, attention residue, and cognitive-load constraints in complex work.

1.2 Why prosoche, why now

Prosoche is operationalized as continuous attention to impressions, judgement separation, and action alignment under pressure. The protocol hypothesis is not universal superiority over mindfulness; it is domain-fit superiority for sustained AI-supervisor workflows where action-time evaluation matters.

1.3 Existing literature gap

The protocol reports no published RCT matching this exact comparison set: Stoic-derived attention intervention versus active mindfulness and productivity controls in AI-intensive work, with behavioral AI-tool outcomes and standardized supervisor tasks.

2. Objectives and hypotheses

2.1 Primary objective

Evaluate whether a four-week prosoche-derived intervention reduces NASA-TLX during AI-supervised work more than matched mindfulness control.

2.2 Secondary objectives

• Compare prosoche against productivity control to isolate attention-specific effects from generic structured-practice effects.
• Estimate effects on sustained attention, residue, flow, agency, and AI-tool behavior patterns.
• Assess durability at 4-week and 12-week post-intervention windows.
• Release open, replicable measurement and intervention materials.

2.3 Hypotheses

• H1 (primary): Prosoche > mindfulness on week-5 NASA-TLX reduction (two-tailed alpha = 0.05).
• H2: Prosoche > productivity on NASA-TLX reduction.
• H3: Attention residue change mediates prosoche effect on NASA-TLX change.
• Exploratory: moderation by baseline AI hours, output quality differences, and agency effects.

3. Intervention design

3.1 Theoretical decomposition

The protocol decomposes prosoche into four trainable components: attention to impressions, judgement/event separation, real-time action evaluation against a canon, and structured self-observation cadence.

3.2 Daily structure

• Morning preparation (3 min): anticipated difficulty, likely automatic assent, virtue-guided action check.
• In-action micro-checks (30 sec each, 3-4x/day): rotating prompts for impression, control, and action evaluation.
• Evening review (5-7 min): structured private reflection with metadata-only collection.

3.3 Weekly arc

Weeks 1-2 establish assent and control disciplines; weeks 3-4 shift to direct AI-supervision application and full-day integration.

3.4 Delivery mode

App + weekly group call was selected over app-only, book-only, and coach 1:1 because it balances fidelity, adherence, and scalability. Group-call logistics are matched across all active arms to avoid support-effect confounds.

3.5 Secularisation and content discipline

Participant-facing language is fully secular and operational. Metaphysical claims are removed while preserving the practical mechanics from the Stoic source tradition.

4. Control arms

The protocol uses four arms: prosoche, mindfulness, productivity, and measurement-only waitlist.

• Mindfulness arm: matched time and call cadence; isolates prosoche-specific effect from generic meditation effects.
• Productivity arm: matched structured daily practice; isolates attention-specific mechanism from ritual effects.
• Waitlist arm: anchors effect sizes and supports ethical post-study access.

The productivity contrast is explicitly treated as non-negotiable for claim quality. AI-hygiene is measured behaviorally across arms rather than introduced as a fifth arm to preserve power for the primary contrast.

5. Outcome battery

5.1 Pre-committed outcomes

NASA-TLX, output quality in a standardized AI-mediated task with blinded scoring, sustained attention metrics, and self-reported flow are carried from the Bandwidth Gap commitments.

5.2 Full battery

• Cognitive load (lab and ecological EMA).
• SART and Mark-style switching cost metrics.
• Output quality rubric scored by two blinded raters with adjudication.
• Attention residue paradigm, flow scale, sense-of-agency scale.
• Behavioral AI-tool telemetry; optional HRV and sleep subset in aggressive tier.

5.3 Minimum essential vs full battery

Lean tier keeps the minimum publishable core; standard and aggressive tiers add residue, agency, richer behavioral analysis, and optional biometrics.

5.4 Standardized lab task

Participants supervise three agents in a 60-minute synthesis task with embedded factual, citation, and consistency errors. Outputs are scored for accuracy, integration, originality, and structure. Materials and rubric are planned for OSF release.

• Output target in source: 1500-word brief on an unfamiliar topic.
• Error embedding in source: factual, citation, and cross-agent consistency faults at known frequencies.
• Scoring reliability target in source: inter-rater ICC > 0.80 with third-rater adjudication below threshold.

6. Participants

6.1 Target population

Working-age professional knowledge workers (25-55) with sustained AI-agent/tool usage in primary work.

6.2 Inclusion criteria

• Age 25-55, CEFR B2+ English proficiency.
• At least 10 hours/week AI-tool usage, verified in baseline week.
• Stable employment context and study-app capable smartphone.
• Willingness to install behavioral log SDK.

6.3 Exclusion criteria

• Active formal meditation practice more than once/week.
• Active treatment context for anxiety, depression, or attention disorders (borderline cases stratified).
• Inability to commit to daily intervention dose.

6.4 Stratification and 6.5 sample size

Blocked randomization by AI-hour bands, age band, sex assigned at birth, and prior-lapsed contemplative experience. Standard tier target is N=180 with attrition padding; aggressive tier supports N up to 300.

• Lean tier in source: exploratory two-arm variant at lower N and weaker claim scope.
• Standard tier in source: minimum credible design preserving active productivity control.
• Aggressive tier in source: expanded N and biometric subset support.

6.6 Recruitment plan

Recruitment channels include Zeno partner network, AI communities, Prolific screening, Brainport/TU-e/Tilburg alumni networks, and targeted LinkedIn campaigns with predefined compensation structure.

7. Study architecture

7.1 Pre-registration

OSF registration before first non-pilot enrolment with locked hypotheses, primary outcome, exclusion criteria, primary model, and mediation hypothesis.

7.2 Randomisation and 7.3 blinding

Independent-statistician randomization lists, allocation concealment infrastructure, blinded output raters, blinded log analysts, and masked primary-analysis statistician until model lock.

7.4 Schedule

Baseline week, intervention weeks 1-4, week-5 post-test, week-9 follow-up, week-17 follow-up. Total participant commitment: 18 weeks.

7.5 Adherence measurement

Daily completion signals, micro-check response counts, evening journal metadata, and weekly call attendance with pre-registered per-protocol thresholds.

• Per-protocol definition in source: at least 70% daily prompts completed and at least three of four calls attended.
• Intent-to-treat is primary analysis; per-protocol is pre-specified sensitivity analysis.

8. Ethics, regulatory, and data protection

8.1 Regulatory pathway and 8.2 risk

Protocol is categorized as minimal-risk non-WMO behavioral research with faculty ethics review pathways detailed for TU/e, Tilburg/JADS, and independent IRB alternatives.

8.3 Informed consent

Tiered consent covers participation, secondary analyses, open data publication, and optional biometrics. Withdrawal rights are preserved with phase-pro-rated compensation.

8.4 GDPR

Lawful bases are specified for each data category. The protocol commits to metadata-only logs, pseudonymisation, separate key custody, EU-resident infrastructure, DPIA completion, processor agreements, and differentiated retention timelines for pseudonymized vs identifiable data.

9. Statistical analysis plan

9.1 Primary analysis

Linear mixed-effects model on NASA-TLX change with planned prosoche-vs-mindfulness contrast, confidence intervals, effect size reporting, and pre-specified per-protocol sensitivity checks.

Model form in source: NASA_TLX_change ~ Arm + AI_use_band + Age_band + Sex + prior_contemplative + (1 | participant).

9.2-9.6 Secondary, mediation, moderation, sensitivity, multiple comparisons

• Parallel mixed-effects structure for secondary outcomes with Holm-Bonferroni control.
• Causal mediation with ACME/ADE decomposition and bootstrapped confidence intervals.
• Pre-specified moderation by AI-use intensity and lapsed contemplative exposure.
• Sensitivity checks: imputation, random-effects alternatives, robust regression, Bayesian counterpart, winsorisation.
• Exploratory families controlled with false-discovery-rate methods and explicit labelling.

10. Publication strategy

Primary submission target is Computers in Human Behavior with secondary targets including Mindfulness and backstop routes. Authorship follows CRediT taxonomy, and open science commitments include preregistration, pseudonymized data release, protocol/material publication, reproducible analysis code, and preprint release.

10.4 Companion publications

The source protocol also defines companion outputs: a public Zeno white paper, popular press adaptation, and conference dissemination with methods-focused split publication options.

11. Timeline and budget

The full source protocol specifies an 18-month execution map with phased setup, pilot, recruitment, intervention, data lock, analysis, manuscript drafting, and submission windows. It also provides three budget tiers (lean, standard, aggressive), explicit line items, and a funding assembly path spanning reserve, partner match, grants, tax credit, and potential sponsor participation.

• Documented total estimates: lean ~€90k, standard ~€275k, aggressive ~€530k.
• Protocol rationale marks standard tier as minimum for strongest claim quality.
• If cuts are required, the source order is biometrics first, then EMA, then eye-tracking subset while preserving the productivity control arm.

12. Risks and open questions

12.1 Methodological risks

The source document covers expectancy bias, demand characteristics, self-selection, Hawthorne effects, differential dropout, co-intervention contamination, generalisability boundaries, and validity questions for AI-context residue measures with corresponding mitigation plans.

12.2 Strategic risks

It also addresses venue framing risk, religious framing dismissal, null-result interpretation, novelty skepticism, scoop risk, and founder-bias mitigation through blinded analysis roles and institutional partnership structure.

12.3 Open design questions

• Prompt delivery mode and notification fatigue management.
• Validity of extending residue paradigms to AI-tool switching.
• Primary outcome tradeoff: standardized lab measure vs richer ecological EMA aggregate.
• Matched group call inclusion tradeoff between fidelity and cleaner inference.
• Handling active meditation practitioners: include+stratify versus exclude.

Appendices A-F

This web version includes the full appendix context from the source protocol:

• Appendix A: sample participant week (prosoche arm, week 3) with exact cadence examples.
• Appendix B: lab task specification, error embedding, topic pool, agent setup, and scoring rubric dimensions.
• Appendix C: complete instrument inventory with cadence, licensing/source, and validation notes.
• Appendix D: formal pilot phase plan including go/no-go decision thresholds.
• Appendix E: three-layer data management architecture and security controls.
• Appendix F: protocol amendment log (v1.0 baseline and future amendment handling).

References

Primary references in the source include Allen (2001), Cresswell (2017), Engeser & Rheinberg (2008), Epictetus (Hard trans.), Epoch AI (2026), Gill (2010), Goyal et al. (2014), Hadot (1995, 1998), Hart & Staveland (1988), Imai et al. (2010), Khoury et al. (2013), Leroy (2009), Linardon et al. (2019), Marcus Aurelius (Hays trans.), Mark et al. (2008, 2014), METR (2026), Posner & Petersen (1990), Robertson et al. (1997), Sellars (2006), Seneca (Campbell trans.), Sweller (2020), Tapal et al. (2017), and Zheng & Meister (2024).

Full citation list preserved from the protocol source document and available in partner review materials.