Pre-Registration Draft — EF vs IF Attentional Encoding in IMU-Driven AR-Mediated FPA Retraining

Status: Draft, for researcher review before formal submission to OSF Registries (Standard Pre-Data Collection Registration template).
Source: Content compiled from 01_Protocol/Protocol_Master_Draft.md (finalized version, incorporating all researcher decisions).
Venue track: This preregistration is intended for a Registered Report submission (Stage 1: in-principle review of the question, design, and analysis plan prior to data collection, followed by Stage 2 result-neutral publication commitment after data collection — see Section 1 for the explicit publish-regardless-of-outcome commitment). Target journal: Sensors (MDPI) is the current default candidate given the sensor-hardware-forward framing, but the venue is not locked — a Registered-Report-track journal offering a better fit (e.g. higher standing in Sport Sciences/Rehabilitation Engineering, see the earlier CAS-tier discussion) will be considered before Stage 1 submission, provided it accepts Registered Reports in this subject area.
Must be completed before submission: Researcher reads through the full text to confirm accuracy → copy/adapt into the official OSF form (and/or the chosen journal's Registered Report Stage 1 template, once the venue is finalized) → submit and lock (cannot be edited after submission; any subsequent change requires an OSF amendment with a timestamped record).
Revision note (post-frontend-update sync): This draft has been updated to reflect the Unity AR client as currently implemented (IMUMoCap/docs/summry/AR-FRONTEND.md), which has diverged from the design frozen in Protocol_Master_Draft.md Section 4.1 in several respects (icon rendering, on-screen dwell time, feedback colour valence, focus-reminder duration — see Section 6 "Newly reintroduced asymmetries" and Section 7 Limitations). This introduces two newly identified confounds (unequal target-icon dwell time; unequal feedback-valence coding between EF and IF) in addition to the confounds already discussed in the prior review round.

Confound-resolution decision (recorded): Per the Round 3 preregistration-review panel's Must-Fix #1 (Preregistration_Review_Report_Round3.md), the authors have resolved the previously-open accept-vs-fix fork in Section 6 by selecting Option (b): a further AR-frontend revision was requested and has been completed. The implementation described throughout Sections 5-7 (EF: stone target visible continuously for the block, pulsed scale/colour, orange-only "missed" signal; IF: footprint hidden by default, flashing briefly on each foot-contact event with bidirectional blue/orange colour) is the confirmed, final build for data collection — it is not a placeholder pending a further engineering decision. The two confounds disclosed below (dwell-time asymmetry, feedback-valence asymmetry) are not eliminated by this revision; they are present in this final build and are retained as deliberately-accepted, disclosed limitations of the chosen EF/IF operationalization, per the directional-bias analysis added to Section 6. Protocol_Master_Draft.md Section 4.1 still needs a mechanical re-sync to match this confirmed build before the Stage 1 freeze timestamp is set (see Companion document freeze policy, Section 6) — that resync is now a bookkeeping task, not an open design decision.

---

1. Study Information

Title: EF vs. IF Attentional Encoding in IMU-Driven AR-Mediated FPA Retraining: A Parallel Comparison of Behavioral and Candidate Mechanism-Adjacent Indicators

Authors: Xiasong Chen, Niall Murray, Thiago Braga Rodrigues (author order consistent with the published IMX'26 paper)

Description / Background:
Foot progression angle (FPA) is a modifiable gait parameter relevant to knee osteoarthritis (KOA) and knee adduction moment (KAM) (Guo, Axe, & Manal, 2007, Gait & Posture, 26(3), 436–441, DOI: 10.1016/j.gaitpost.2006.10.008). Prior AR/IMU-based FPA feedback systems (e.g. Karatsidis, Richards, Konrath, van den Noort, Schepers, Bellusci, Harlaar, & Veltink, 2018, Journal of NeuroEngineering and Rehabilitation, 15:78, DOI: 10.1186/s12984-018-0419-2) default to internal-focus (IF) style feedback without this design choice having been examined against external-focus (EF) alternatives. This study extends a previously published baseline-driven personalized FPA retraining system (Chen, Murray & Rodrigues, 2026, ACM IMX'26, DOI: 10.1145/3788851.3812844) to directly compare EF vs IF attentional encoding, reframing the research question from a Quality-of-Experience (QoE) evaluation to a parallel comparison of behavioral and candidate mechanism-adjacent indicators associated with two theoretically-grounded pathways proposed to underlie any external-focus advantage (this is not a formal mediation/path analysis — see Section 6 for the confirmatory group comparisons and the exploratory indirect-effect estimate reported alongside them):

1. The constrained-action hypothesis (EF → reduced conscious motor control → reduced cognitive load; Wulf, McNevin, & Shea, 2001, The Quarterly Journal of Experimental Psychology Section A, 54(4), 1143–1154, DOI: 10.1080/713756012; reviewed in Wulf, 2013, International Review of Sport and Exercise Psychology, 6(1), 77–104, DOI: 10.1080/1750984X.2012.723728)
2. OPTIMAL theory's enhanced-expectancies pathway (EF → strengthened goal-action coupling → increased perceived competence; Wulf & Lewthwaite, 2016, Psychonomic Bulletin & Review, 23(5), 1382–1414, DOI: 10.3758/s13423-015-0999-9)

This study manipulates one of OPTIMAL theory's three proposed factors (external focus of attention) and measures a candidate indicator of the mechanism it is theoretically predicted to engage (enhanced expectancies, via IMI Perceived Competence); OPTIMAL's third factor (autonomy support) is held constant across conditions and is explicitly not tested, due to sample size constraints (a full factorial test of all three factors would require a design well beyond what N=22 supports).

Literature gap verification: An independent literature check (PubMed, Crossref, Google Scholar; see 00_Background/Literature_Check_Report.md) confirmed no published study has directly compared EF vs IF attentional encoding within an AR/IMU gait retraining system targeting FPA specifically.

Publication commitment: Consistent with this study's anti-p-hacking design philosophy (see Section 6's multiple-comparisons policy and its commitment to reporting exact p-values, effect sizes, and confidence intervals for all primary comparisons regardless of significance), the authors commit to submitting the results for publication regardless of the direction or statistical significance of the outcomes, including null or mixed results across the six primary hypotheses.

Translational context: This study isolates candidate attentional-focus mechanisms in a healthy, AR-naive population using a controlled AR/IMU protocol; it is explicitly a mechanism-isolation study, not a clinical efficacy trial. However, by establishing whether EF or IF encoding yields superior FPA target achievement, retention, and lower cognitive load, this work is intended to directly inform the feedback-encoding design choices of the next-stage AR/IMU gait-retraining systems aimed at knee osteoarthritis (KOA) populations, for whom toe-in/toe-out FPA modification is an established, non-pharmacological intervention for reducing knee adduction moment and pain (Guo, Axe, & Manal, 2007). Identifying an evidence-based default attentional-focus encoding at this healthy-adult proof-of-mechanism stage is intended to de-risk the feedback design of a future clinical (KOA) trial, rather than requiring that later trial to separately test EF vs IF from scratch.

Task-novelty and skill-level caveat: All participants are AR-naive (prior professional AR/VR experience is an exclusion criterion, Section 4), and the sample is exclusively healthy adults. Attentional-focus effects are known in the broader motor-learning literature to be moderated by task complexity and performer skill level (Wulf, 2013); the magnitude and direction of any EF/IF difference observed here should therefore not be assumed to generalize to AR-experienced users, to more/less complex gait-retraining tasks, or to clinical (e.g. KOA) populations without further study.

---

2. Hypotheses

The full rationale for test directionality and the multiple-comparisons strategy is given in 01_Protocol/Protocol_Master_Draft.md Section 7.0; only the final version needed for the preregistration is listed here.

RQ1 (Secondary): Can a personalized baseline-driven AR feedback system induce measurable FPA changes toward individualized targets in healthy participants?

• H1 (two-tailed, non-directional test of a directional target, see note): Mean FPA during the training block will shift toward the individualized target direction (μᵢ ± k·SDᵢ; toe-out training → positive shift, toe-in training → negative shift) relative to baseline FPA (one-sample test against zero shift, per condition). Note: analyzed via two-tailed test against zero (consistent with H1's non-directional status re: EF vs IF), but interpreted with reference to the pre-specified target direction — a significant shift in the direction opposite the individualized target does not count as support for H1, even if statistically significant against zero.

RQ2 (Primary): Do external-focus encoding (EF) and internal-focus encoding (IF) differ in FPA target achievement rate, within-session adaptation trend, and immediate retention effect?

• H2a (one-tailed, directional): EF will yield a higher FPA target achievement rate than IF during the training block.
• H2b (one-tailed, directional): EF will produce a smaller immediate retention decrement than IF (i.e., better retention). Scoping note: "retention" here refers strictly to the ~5-minute within-session retention test (feedback removed, following the NASA-TLX(1)+IMI-PC questionnaire delay — see Section 3 Session structure), not longer-term (day/week-scale) motor retention; no claim of longer-term retention is made or tested.
• H2c (two-tailed, non-directional): EF and IF will show different condition × block interaction patterns in within-session FPA adaptation trend.

RQ3 (Primary, revised — parallel comparison): Do EF and IF differ on candidate cognitive/motivational indicators associated with the external-focus advantage — namely (i) cognitive load, as a candidate indicator of the constrained-action pathway (H3a), and (ii) perceived competence, as a candidate indicator of the OPTIMAL enhanced-expectancies pathway (H3b) — and are these mechanism-adjacent indicators associated with behavioral FPA outcomes (exploratory)? Scope note: this tests one of OPTIMAL theory's three proposed factors (external focus of attention); the theory's other two factors (autonomy support, and the intrinsic-motivation pathway more broadly) are not manipulated or tested here (see Section 1).

• H3a (one-tailed, directional): EF will produce lower NASA-TLX cognitive load scores than IF during the feedback-guided training block (Administration 1).
• H3b (one-tailed, directional): EF will produce higher IMI Perceived Competence scores than IF.

H3c: Formally withdrawn (originally depended on the Custom QoE questionnaire, which has been removed from the study in its entirety; see Section 5 below and CHANGELOG.md items 1–3, 13 for full history).

Exploratory analyses (no H-numbers, not error-corrected, not hypothesis-tested in the confirmatory sense):

• NASA-TLX (Administration 1) correlation with FPA target achievement rate (Spearman + supplementary LME) — required to fulfil RQ3(revised)'s commitment to testing whether mechanism-adjacent indicators associate with behavioral outcomes
• IMI Perceived Competence correlation with FPA target achievement rate (Spearman + supplementary LME) — added to match RQ3(revised)'s reference to "these mechanism-adjacent indicators" (plural); same rationale as above
• NASA-TLX Administration 2 (feedback removed) vs. Administration 1 (feedback present) descriptive comparison — uses data already collected for confirmatory H3a, no added burden
• Manipulation Check item 3 (stability self-rating) descriptive analysis and optional correlation with NASA-TLX/FPA achievement — tied to the pre-approved extension of the Manipulation Check instrument (Option A, see Protocol_Master_Draft.md Section 5.3)

Note on Order × condition interaction (counterbalance check): This is not an exploratory hypothesis and has been moved out of this list. It is a precondition/assumption check (verifying the counterbalancing worked as intended before interpreting the primary EF/IF comparisons), listed separately in Protocol_Master_Draft.md Section 7 as a "precondition check (not a numbered hypothesis)". Standard practice in any crossover design, not a discretionary analysis choice.

Why these 4 items are retained rather than arbitrarily trimmed: each corresponds to a specific rationale already decided during protocol development (RQ3 consistency ×2, zero-cost reuse of already-collected data, an approved instrument extension) — these are not "ad hoc exploratory additions tacked on after data collection"; the number of items alone does not constitute a rigor problem (see the methodological rationale below), so they are retained.

Methodological rationale (justification for preregistering exploratory analyses):

• Simmons, Nelson, & Simonsohn (2011). False-positive psychology: Undisclosed flexibility in data collection and analysis allows presenting anything as significant. Psychological Science. DOI: 10.1177/0956797611417632
• Wagenmakers, Wetzels, Borsboom, van der Maas, & Kievit (2012). An agenda for purely confirmatory research. Perspectives on Psychological Science. DOI: 10.1177/1745691612463078
• Gelman, A., & Loken, E. (2014). The statistical crisis in science. American Scientist. DOI: 10.1511/2014.111.460
• Nosek, Ebersole, DeHaven, & Mellor (2018). The preregistration revolution. PNAS. DOI: 10.1073/pnas.1708274114
• Wicherts, Veldkamp, Augusteijn, Bakker, van Aert, & van Assen (2016). Degrees of freedom in planning, running, analyzing, and reporting psychological studies: A checklist to avoid p-hacking. Frontiers in Psychology. DOI: 10.3389/fpsyg.2016.01832

---

3. Design Plan

Study type: Experiment (within-subjects crossover)

Blinding: Participants cannot be blinded to condition (EF vs IF instructions and AR content necessarily differ). Outcome assessors for the FPA/IMU pipeline are automated (no manual scoring), reducing assessor bias for behavioral outcomes. Manipulation Check open-text coding will use two independent raters (see Section 6).

Study design: Two-visit within-subjects crossover, EF and IF conditions completed on separate days, 48–72 hours apart (one condition per day). This interval is a conservative buffer built on top of the ~24h sleep-dependent motor memory consolidation window documented in Walker & Stickgold (2004, DOI: 10.1016/j.neuron.2004.08.031) and Robertson (2004, DOI: 10.1016/j.cub.2004.11.048) — note: the literature supports the ≥24h principle, not the specific 48–72h figure itself, which is a practical/scheduling choice; see Protocol_Master_Draft.md Section 8 for the full, transparent caveat.

Randomization / Counterbalancing: Participants randomly assigned to one of two order groups (EF-first vs IF-first) to control for practice/order effects. Order × condition interaction checked as a precondition (Mixed ANOVA). Given that this interaction test is itself likely underpowered at N=22 to detect a true order effect (interaction effects typically require substantially larger samples than main effects), all six primary comparisons (H1–H3b) will be reported in two versions regardless of the precondition result: with and without order as a covariate. The precondition test result and both versions of each primary comparison will be reported; the precondition result will inform interpretation but will not be used as a binary gate for whether order is considered in the primary analyses.

Session structure (per visit, ≈43 min total, see Protocol_Master_Draft.md Section 6.4 for the full stage-by-stage breakdown with measured/estimated time annotations):
Pre-session (15 min) → Baseline Walk (100 steps) → Target Generation (<1 min) → Training Block 1 (100 steps/foot) → Rest + Manipulation Check 1 → Training Block 2 (100 steps/foot, no rest/MC after) → Training Block 3 (100 steps/foot) → Rest + Manipulation Check 2 → NASA-TLX Administration 1 + IMI-PC (~5 min, doubles as retention delay) → Retention Test (100 steps, feedback removed) → NASA-TLX Administration 2 (~3 min).

---

4. Sampling Plan

Existing data: None (prospective data collection).

Inclusion criteria: Healthy adults, aged 18–65 years; able to walk independently ≥20 minutes without assistive devices; normal or corrected-to-normal vision; no history of lower-limb orthopaedic surgery or acute lower-limb injury.

Exclusion criteria: Known gait disorder, neurological condition, or vestibular dysfunction; history of motion sickness/cybersickness; lower-limb pain in the past 6 months affecting walking; prior professional AR/VR training or research experience (controls device-familiarity bias); self-reported colour vision deficiency (e.g., red-green colour blindness), screened via an added item on the Participant Health and Fitness Screening Questionnaire — added because the AR feedback's deviation signal is colour-coded (see Section 3 and 01_Protocol/CHANGELOG.md for the blue/orange palette adopted to reduce this risk further).

Sample size rationale: A priori power analysis (G*Power, paired t-test, two-tailed, α=.05, power=.80) using the EF vs IF retention effect size from Chua et al. (2021) meta-analysis (Hedges' g = 0.583, DOI: 10.1037/bul0000335) gives required N=22; this full calculated N is targeted (revised upward from an earlier N=18–20 compromise, so recruitment is not stopped below the study's own calculated requirement). To buffer against anticipated attrition, enrollment will continue slightly beyond 22 as needed to reach N=22 valid completers (both sessions). This study is explicitly framed as an exploratory/pilot investigation; post-hoc statistical power will be reported for all primary comparisons. Note: switching to one-tailed tests (Section 2) for directional hypotheses makes this power estimate conservative for H2a/H2b/H3a/H3b (achieved power ≈0.84 at N=22 for an effect of this magnitude); for the two-tailed hypotheses (H1, H2c), achieved power at N=22 is ≈0.74 for an effect of this magnitude, and lower still if the true effect is smaller than the borrowed retention effect size. This a priori power calculation is anchored specifically to H2b's effect size; a priori power for H1, H2a, H2c, H3a, H3b is not independently established, and effect sizes with confidence intervals — not power/significance claims — will remain the primary basis for interpreting those five hypotheses.

Borrowed effect-size transferability caveat: the Chua et al. (2021) meta-analytic effect size (g=0.583) predominantly reflects studies using distal/classic external-focus instructions (e.g., verbal cues directing attention to an external object or effect); its applicability to this study's proximal EF operationalization (a head-locked, 3DoF AR display, see Section 3 and Protocol_Master_Draft.md Section 8 Limitations) has not been independently established and may overestimate the achievable effect. As a sensitivity check, achieved power at N=22 for smaller, plausible effect sizes is: d=0.5 → 61% (two-tailed) / 73% (one-tailed); d=0.4 → 43% / 57%; d=0.3 → 27% / 39%. Readers should treat the power estimate above as an optimistic upper bound rather than a guarantee, particularly for H1 and H2c.

Stopping rule: Data collection stops once N=22 valid participants (completing both sessions) is reached, or at the end of the funded/allotted recruitment window, whichever comes first. Safety stopping rule (per-participant): session discontinued immediately if participant reports "Moderate" or "Severe" on session-level discomfort screening, or requests to stop.

Missing data / attrition handling policy (added in response to preregistration review, Preregistration_Review_Report.md Methodology-W4):

• Definition of "valid participant": a participant is counted toward N=22 only if they complete the core procedure (baseline walk, all 3 training blocks, retention test) in both the EF and IF visits. Completing only one of the two visits does not count as a valid participant.
• Participant-level missingness (one visit completed, the other not): such participants are excluded from all confirmatory comparisons (H1–H3b), which all require within-person paired data across both conditions. Recruitment continues until N=22 valid participants is reached (per the Stopping rule above), i.e., these participants are effectively replaced, not counted toward the target. Their single-visit data are not discarded from the record: attrition count and stated reason (where available) will be reported descriptively in the eventual manuscript (e.g., a simple enrolled → completed-visit-1 → completed-both-visits flow), and single-visit safety/discomfort data still contribute to the safety-monitoring summary.
• Measurement-level missingness within a completed pair of visits (both visits completed, but some specific data point is missing or corrupted):
  • FPA/step-based outcomes: if a training block or the retention test has a partial step count due to sensor dropout or session interruption, the block's data are still used if at least 70% of the intended steps for that block were validly captured; below that threshold, that specific block is treated as missing for that participant for that outcome only (it does not remove the participant from other blocks/hypotheses).
  • Single-item rating scales (NASA-TLX's six subscales, each a standalone 0–100 slider): a skipped subscale cannot be imputed (there is no sub-item to average) and is treated as missing for that participant on that subscale only; H3a's per-subscale comparison for that subscale is computed on the remaining valid participants for that subscale.
  • Multi-item scales (IMI-PC, 5 items): if exactly 1 of 5 items is missing, the scale score is computed as the mean of the remaining 4 items (standard psychometric practice for short scales); if 2 or more items are missing, the IMI-PC score for that participant/visit is treated as missing.
• Reporting commitment: because the above rules can make the analyzed N differ slightly across H1–H3b, the eventual results will report the exact analyzed N for each of the six hypotheses individually, rather than assuming N=22 applies uniformly to every test.
• No imputation is used for participant-level (whole-visit) missingness or for FPA/step-based outcomes beyond the 70% threshold rule above; this pilot study does not employ multiple imputation, which is judged to add unnecessary researcher degrees of freedom relative to its benefit at this sample size.

---

5. Variables

Manipulated variable: Attentional focus encoding of AR feedback (EF vs IF), operationalized via AR display content + verbal instruction. As-built AR display behaviour is documented in IMUMoCap/docs/summry/AR-FRONTEND.md; this is the confirmed final build for data collection following the Round 3 review's required confound-resolution decision (Option (b): a further frontend revision was requested and completed — see Section 6 for the decision record and directional-bias analysis). Protocol_Master_Draft.md Section 4.1 still describes an earlier design and is pending a mechanical re-sync to this confirmed build (see Section 6, Companion document freeze).

Measured variables (confirmatory):

• FPA (foot progression angle), estimated in real time via 3× Xsens MTw Awinda IMUs (both feet + pelvis, 100 Hz)
• FPA Absolute Shift, FPA Target Achievement Rate, Within-Session Adaptation Trend, Retention Decrement (all derived from the FPA time series; exact formulas in Protocol_Master_Draft.md Section 3.2)
• NASA-TLX (Raw, 6 subscales), Administration 1 and 2. Proxy-choice caveat: NASA-TLX is a self-report subjective workload scale, used here as a pragmatic, indirect proxy for the constrained-action hypothesis's proposed mechanism ("reduced conscious motor control"); it does not directly measure motor automaticity, and more standard direct indicators in that literature (e.g., behavioral dual-task costs, implicit automaticity measures) are not included in this pilot's battery. This is a deliberate scope/burden trade-off for this exploratory pilot, not an oversight, and readers should interpret H3a accordingly.
• IMI Perceived Competence (5 items: IMI Task Evaluation Questionnaire items 4/7/12/16/22, Markland & Hardy 1997 operationalization, DOI: 10.1080/02701367.1997.10608863)

Measured variables (non-confirmatory / exploratory / procedural):

• Manipulation Check (open response + forced choice + stability self-rating; adapted from Porter, Nolan, Ostrowski & Wulf 2010, DOI: 10.3389/fpsyg.2010.00216, with the stability self-rating item as a study-specific addition beyond the original method — see Protocol_Master_Draft.md Section 5.3)
• Device discomfort / safety screening (verbal, pre-session and each session; pilot-based protocol, n=3 informal, ethics-approved under 008-11-2025, pilot completed)

Removed variables (formerly planned, now excluded — see CHANGELOG.md for full history): Custom QoE Questionnaire (4 dimensions), SAM (Valence/Arousal/Dominance), SSQ/ARSQ formal discomfort questionnaire (replaced by verbal screening).

---

6. Analysis Plan

Statistical software: R (version 4.4.2 or later; exact version and package versions to be confirmed and locked at the time of the analysis script being finalized, e.g. via sessionInfo() / renv::snapshot()). Key packages: stats (base R; Wilcoxon signed-rank tests, paired t-tests, Shapiro-Wilk normality checks), afex/ez (repeated-measures ANOVA, condition × block interaction, H2c), lme4 + lmerTest (exploratory linear mixed-effects models), irr (Cohen's kappa, manipulation check inter-rater agreement), effectsize (standardized effect sizes: Hedges' g, rank-biserial r).

Hypothesis	Test	Tails	Correction	Outcome classification rule
H1	One-sample Wilcoxon signed-rank (or paired t-test if normality holds, Shapiro-Wilk) vs. zero	Two-tailed	None across hypotheses (see below)	Three-way rule (explicit): (a) significant + shift toward the individualized target direction → counts as support for H1; (b) significant + shift in the direction opposite the individualized target → does not count as support for H1, even though statistically significant against zero; (c) non-significant → no support for H1. (Full rationale in Section 2.)
H2a	Wilcoxon signed-rank (EF vs IF)	One-tailed (EF > IF)	None across hypotheses	Standard directional test; significant + predicted direction = support, otherwise no support
H2b	Paired t-test (EF vs IF retention decrement)	One-tailed (EF < IF decrement)	None across hypotheses	Standard directional test; significant + predicted direction = support, otherwise no support
H2c	Repeated-measures ANOVA (2 condition × 3 block), condition×block interaction	Two-tailed	Greenhouse-Geisser if Mauchly significant; none across hypotheses	Standard non-directional test; significant interaction (any pattern) = support, otherwise no support
H3a	Wilcoxon signed-rank (EF vs IF), per NASA-TLX subscale	One-tailed (EF < IF)	Bonferroni within the 6 NASA-TLX subscales (local correction only)	Standard directional test; significant + predicted direction = support, otherwise no support
H3b	Wilcoxon signed-rank (EF vs IF), IMI-PC	One-tailed (EF > IF)	None across hypotheses	Standard directional test; significant + predicted direction = support, otherwise no support

Multiple comparisons policy (must be stated explicitly, decided prior to data collection): No correction is applied across the six primary hypotheses (H1, H2a, H2b, H2c, H3a, H3b). Rationale: each hypothesis addresses a distinct scientific question (not repeated measurement of the same construct), and this study is explicitly exploratory/pilot (N=22); an overall correction (e.g., Bonferroni to α≈.0083) would substantially reduce already-limited power. As a transparency safeguard, exact p-values, effect sizes, and confidence intervals will be reported for all primary comparisons (not binary significant/non-significant labels only), allowing readers to apply their own correction if desired. This decision does not affect the existing local Bonferroni correction already applied within the 6 NASA-TLX subscales (a genuinely repeated-construct family). Per the missing-data policy in Section 4, the exact analyzed N for each of the six hypotheses will also be reported alongside its p-value, effect size, and confidence interval, since the missing-data rules can make analyzed N differ slightly across H1–H3b rather than uniformly equal to 22.

Family-wise false-positive rate (quantified): under a global null (i.e., if all six hypotheses were truly false), the probability of at least one spurious p<.05 result across the six uncorrected primary comparisons is 1-(1-.05)^6 ≈ 26.5%. This is explicitly acknowledged rather than left implicit; the authors commit to giving equal narrative prominence to all six results in the eventual manuscript regardless of which one(s) cross the significance threshold, so this elevated false-positive risk cannot be selectively exploited by only foregrounding whichever hypothesis happens to reach p<.05.

Clarification of the "Primary" label: throughout this document, "Primary" denotes a hypothesis's confirmatory-test status (i.e., it is analyzed with a pre-specified, non-exploratory test and is not subject to post-hoc reinterpretation) — it does not denote or imply that the hypothesis is adequately powered. As detailed in Section 4, only H2b has an independent a priori power justification; the other five Primary hypotheses (H1, H2a, H2c, H3a, H3b) are confirmatory in status but of unestablished a priori power.

Companion document freeze: the exact outcome-variable formulas (FPA Absolute Shift, FPA Target Achievement Rate, Within-Session Adaptation Trend, Retention Decrement) are defined in Protocol_Master_Draft.md Section 3.2, referenced rather than reproduced here. Protocol_Master_Draft.md is frozen at the same timestamp as this OSF submission; any subsequent change to Section 3.2's formulas after submission is treated as a deviation from the pre-registered plan under the same disclosure rule stated in Section 7 ("Deviations from this pre-registration"), not as a silent update to a "living" companion document.

Manipulation Check analysis: Pass/failure rate + inter-rater agreement (Cohen's κ, two independent coders classifying open-text responses) + sensitivity analysis restricted to "passers." Not hypothesis-tested (no significance test / no H-number).

Exploratory analyses (see Section 2 for the list; all Spearman correlations + supplementary linear mixed-effects models): reported descriptively, explicitly flagged as underpowered given N=22 for detecting small-to-medium correlations, no confirmatory claims will be drawn from these. Additionally, an exploratory indirect-effect estimate (a×b, using the paired-difference approach appropriate to this two-condition within-subject design: the EF−IF difference in each candidate indicator as path a, already registered as H3a/H3b, combined with its exploratory correlation with FPA target achievement rate as path b) will be computed for each of NASA-TLX and IMI-PC, with a bias-corrected bootstrap confidence interval, reported descriptively as a non-confirmatory, likely underpowered probe of the proposed pathways — not presented as an established mediation finding.

Known confounds / how handled: Order effects (counterbalanced; checked as an underpowered-but-reported precondition, Section 3; all six primary comparisons additionally reported with and without order as a covariate, per Section 3, rather than relying on the precondition test as a binary gate). Practice/familiarity with AR/VR (excluded via exclusion criteria). Within-session carry-over between conditions (mitigated via 48–72h washout, see Section 3; acknowledged as not fully eliminable, see Protocol_Master_Draft.md Section 8 Limitations). EF/IF feedback complexity matching — resolved confounds vs. retained necessary differences (see Protocol_Master_Draft.md Section 4.1 for the full specification):

Partially resolved confounds (unintended asymmetries, not part of the theoretical manipulation; targeted by an explicit design change, but the design change itself has since been superseded by a further implementation revision — see "Newly reintroduced asymmetries" immediately below): the original EF display was static except for a brief 0.2s flash at foot contact, while the original IF display updated continuously every frame throughout the step — an asymmetry in sustained visual engagement independent of the intended attentional-focus manipulation. A first-pass fix made both conditions terminal (flash-on-foot-contact only) and unified their flash timing via one shared configuration constant (FootOutDisplaySec, currently 0.35s for both conditions — see AR-FRONTEND.md, "Known issues": this value is hand-tuned, not derived from a fresh data-driven analysis the way an earlier 0.2s EF-only figure was). Both conditions also draw from one shared colour-blind-safe blue/orange palette (FeedbackColors.cs; see Section 4 exclusion criteria for the accompanying colour-vision screening item) and display an equal, word-count-matched 3-second text reminder of the assigned attentional focus before each training block (corrected from an earlier-stated 5 seconds; focusReminderDurationSec in the current build is 3s — not during the block itself, to avoid introducing a new concurrent-vs-terminal asymmetry and to avoid a reading-while-walking safety hazard).

Newly reintroduced asymmetries (current build, per IMUMoCap/docs/summry/AR-FRONTEND.md, not yet resolved or fully disclosed prior to this revision): since that first-pass fix, EF and IF have diverged onto genuinely different rendering mechanisms rather than a single shared terminal-flash implementation, reopening two asymmetries of the same general kind the first-pass fix was meant to close:

1. Baseline-visibility duty cycle: the EF "stone" graphic (stoneLeft/stoneRight) is shown as soon as a training block starts and stays visible continuously for the whole block (it only ever pulses in scale/colour on a foot-contact event, via PulseStone()); the IF "footprint" graphic (leftSolid/rightSolid) is hidden by default for the entire block and becomes visible only for the ~0.35s flash window on each foot-contact event, via FlashFoot(). In effect, EF now has near-100% on-screen dwell time for its target icon across a block, while IF has near-0% dwell time outside its per-step flashes — the inverse configuration of the original static-EF/continuous-IF problem, not a symmetric fix of it. The previously-reported "equalised proportion of step time during which the display is actively changing" metric (flash duration ÷ average step duration) remains valid for describing the transient pulse/flash event itself, but no longer describes total visible exposure, because EF's target is visible at rest even when not pulsing and IF's is not.
2. Feedback-valence coding asymmetry: EF's pulse uses only FeedbackColors.OffTarget (orange) on a missed step and leaves the stone at its own cached, unmodified texture colour when on-target — it never turns OnTarget (blue). IF's flash uses both OnTarget (blue, hit) and OffTarget (orange, miss) from the same shared palette. EF therefore only ever encodes a negative ("you missed") signal, whereas IF encodes both positive and negative outcomes; this is a difference in feedback informativeness/valence structure independent of the external-vs-internal attentional-focus manipulation itself.

Because both of these are display-implementation choices rather than deliberate instantiations of the EF/IF theoretical contrast (unlike the head-locked-anchoring and fixed-target-vs-measured-rotation differences addressed under "Retained necessary differences" below), they are treated here as acknowledged confounds rather than intended manipulation content, following the same disclosure standard applied to the display-anchoring finding in this section: a positive or null EF-vs-IF result on H2a–c/H3a–b cannot, on the current build, be fully separated from (1) unequal visual-target dwell time or (2) unequal availability of positive (on-target) feedback.

Decision (recorded): the authors selected Option (b) — a further AR-frontend revision was requested and completed — rather than Option (a) (accept the pre-revision build unchanged). The implementation described above (items 1-2) is the result of that revision and is confirmed as the final build used for data collection; it is not an interim state pending a further engineering choice. Choosing to revise the frontend did not eliminate either asymmetry (a frontend change that, for example, added an on-target colour state to the EF stone, or bounded its visible duration to match IF's flash window, was considered but not adopted); rather, the revision converged on this specific implementation and the authors elect to retain both asymmetries as disclosed limitations rather than re-engineer further, given that (i) ethics approval 008-11-2025 is already finalized for the underlying protocol, (ii) neither asymmetry requires a new participant-facing procedure to address it, and (iii) the directional-bias analysis immediately below is judged sufficient to make the resulting interpretive risk explicit and pre-specified rather than requiring elimination. Protocol_Master_Draft.md Section 4.1 as currently written still describes an earlier, pre-divergence, first-pass-fix design (identical solid-fill icon for both conditions, 5-second reminder, single "feedback active time" equality metric) and has not yet been re-synced to this confirmed build; per the Companion document freeze policy below, this re-sync is a required bookkeeping step before the freeze timestamp is set, not an open design decision.

Directional-bias analysis (per hypothesis, added per Round 3 review Must-Fix #2): because the confound-resolution decision above retains both asymmetries rather than eliminating them, their likely direction of influence on each primary hypothesis is pre-specified here so that results can be interpreted against a stated prior rather than post hoc:

• H2a (EF > IF target achievement rate): the dwell-time asymmetry (EF's target continuously visible; IF's measured state visible only ~0.35s/step) plausibly inflates an EF advantage independent of attentional-focus content, since EF participants have a persistent visual reference to aim at for a much larger fraction of each step. A significant H2a result is therefore not cleanly attributable to attentional focus alone; a null or IF-favoring H2a result would be comparatively harder to explain away via this confound (it would hold despite a dwell-time advantage for EF).
• H2b (EF < IF retention decrement) / H2c (condition × block interaction): both asymmetries plausibly act during the training blocks (not the retention test, where feedback is removed for both conditions equally), so their direct influence on the retention-test outcome itself (H2b) is less direct than for H2a; however, if dwell-time or valence differences shape what is learned during training (e.g., differential reliance on the persistent EF target, per the guidance-hypothesis literature on continuous vs. faded feedback, Salmoni, Schmidt, & Walter, 1984, Psychological Bulletin, DOI: 10.1037/0033-2909.95.3.355), this could plausibly produce a larger apparent EF retention decrement than attentional focus alone would predict (EF participants may be more dependent on the persistent visual target), which would work against H2b's predicted direction. H2c's within-session adaptation-trend comparison inherits both the H2a-direction (dwell-time) and H2b-direction (valence/dependency) risks across its three blocks.
• H3a (EF < IF cognitive load) / H3b (EF > IF perceived competence): the feedback-valence asymmetry (EF: orange/"missed" signal only, never a positive on-target colour change; IF: both blue/"hit" and orange/"missed") plausibly biases against both EF-favoring hypotheses. Under OPTIMAL theory's own enhanced-expectancies mechanism (Wulf & Lewthwaite, 2016), receiving explicit positive feedback is itself expected to boost perceived competence; EF participants structurally never receive this signal while IF participants sometimes do, independent of attentional-focus content. This means a null H3b result is ambiguous between "no true EF competence advantage" and "a design artifact suppressing it," while a significant H3b result (EF > IF) would be a comparatively stronger finding, since it would hold despite a feedback-valence structure tilted toward IF. The same logic applies to H3a in the opposite subjective direction: the absence of positive confirmatory feedback is a plausible, independent contributor to subjectively higher workload/uncertainty for EF, so a null H3a result is similarly ambiguous, and a significant H3a result (EF < IF cognitive load) would likewise be the stronger reading.
• Summary for interpretation: the two confounds are not expected to bias all six hypotheses in the same direction. Dwell-time asymmetry plausibly inflates an EF advantage on H2a (and, more speculatively, H2c); valence-coding asymmetry plausibly works against an EF advantage on H3a/H3b (and may contribute to an inflated H2b/H2c retention decrement for EF via feedback-dependency). A pattern of "EF wins H2a but not H3a/H3b" is therefore consistent with the confounds alone and does not by itself distinguish an attentional-focus effect from these implementation artifacts; this pattern-level caveat will be stated explicitly in the eventual manuscript's Discussion regardless of which hypotheses reach significance.

Retained necessary differences (not confounds — these differences are how the EF/IF manipulation itself is operationalised, and are deliberately not equalised): both EF and IF are rendered as head-locked, 3DoF HUD elements fixed in the wearer's field of view (per Protocol_Master_Draft.md Section 4.1's technical clarification: "Ours is head-locked (3DoF): it moves with the head and is not anchored to the physical world") — the two conditions do not differ in display-anchoring mechanism, and the hardware does not support true world-locked/ground-anchored (SLAM) rendering for either condition. Correction to an earlier draft of this section, which incorrectly described EF's rendering as "world-locked 3D... anchored to a real-world ground location": this was a drafting error, not a design feature; EF is properly described (per Protocol_Master_Draft.md Section 8's "3DoF hardware constraint") as proximal EF — a head-locked display whose content is positioned to represent an external, environment-anchored target location (~1.5 m ahead, ground level), rather than a display that is itself world-anchored. What is retained and deliberately not equalised is therefore the represented content/referent of each display, not its anchoring mechanism: EF's content represents an external ground-target location, consistent with EF's theoretical definition of attention directed at an external, environment-anchored effect; IF's content represents an abstracted state of the participant's own foot/body, consistent with IF's theoretical definition of attention directed at the body movement itself. Because both displays share an identical head-locked rendering mechanism, the specific rival-explanation risk of a display-anchoring-format confound (differential depth-cue processing, gaze/head-coordination demand, or vestibular-ocular/postural load arising from one display being world-anchored and the other head-anchored) does not apply here. A narrower, already-acknowledged limitation remains, however: because EF is only a proximal approximation of a true world-locked external target (a hardware constraint, not a chosen equalisation), a positive EF > IF finding cannot be fully attributed to attentional-focus content alone independent of this proximal-EF approximation; this is the same limitation already flagged under "3DoF hardware constraint" in Protocol_Master_Draft.md Section 8, restated here so Section 6 and Section 7 are mutually consistent rather than describing two apparently different confounds.

This is also where the current build's remaining EF/IF content differences belong, now that the icon-shape claim above has been corrected: EF renders as a distinct "stepping stone" sprite that is set once per block to a fixed target angle/position and never rotates to show the participant's actual measured angle; IF renders as a "footprint" sprite that is hidden until a foot-contact event, then briefly rotates to the participant's actual measured error (msg.errorL/msg.errorR = FPA − target) before hiding again. This fixed-target-only-vs-measured-rotation distinction is judged a retained, theoretically-motivated difference (EF communicates only the external goal location, consistent with attention-on-effect; IF communicates the participant's own measured body state, consistent with attention-on-movement) and is listed here for completeness rather than under "Newly reintroduced asymmetries" above, which is reserved for differences not motivated by the EF/IF theoretical contrast itself.

Not yet independently verified: whether the blue and orange hues chosen are subjectively matched in visual salience/contrast against the AR display background (colour-blind-safety and salience-matching are related but distinct properties); this will be checked informally during the next pilot round rather than assumed from palette selection alone.

---

7. Other

Data availability: De-identified processed data (FPA-derived outcome variables, questionnaire scores) and analysis code will be made publicly available via the OSF project associated with this registration upon manuscript acceptance. Raw IMU time-series data (large file volume) will be archived on Zenodo with a persistent DOI, linked from the OSF project. Data will be de-identified prior to deposit; no directly identifying information (name, contact details, raw video) will be shared. Access to any residual identifiable material (e.g., session video used only for internal QA) is available upon reasonable request to the corresponding author, subject to institutional ethics approval (008-11-2025) data-sharing terms.

Ethics: Ethics approval 008-11-2025 obtained for the finalized protocol (including the pilot-based safety monitoring replacing formal SSQ/ARSQ); confirmed by researcher as finalized, no further protocol changes anticipated (see Protocol_Master_Draft.md Section 9(b)).

Acknowledged limitations: See Protocol_Master_Draft.md Section 8 for the full list (no independent no-feedback control group; 3DoF/proximal-EF hardware constraint — both EF and IF are head-locked HUD renderings, so this limitation concerns EF's content being only a proximal approximation of a true world-anchored external target, not a display-anchoring-mechanism difference between conditions, see Section 6 above; two-visit design washout cannot fully eliminate carry-over; exploratory sample size N<20 underpowered for small effects; healthy-participant sample may not generalize to clinical populations). Additionally, in the confirmed final AR frontend build used for data collection (see Section 6, "Newly reintroduced asymmetries" and "Decision (recorded)"): EF's target icon is continuously visible for the whole training block while IF's icon is visible only during its brief (~0.35s) per-step flash (an unequal visual-target dwell-time confound), and EF's feedback pulse encodes only missed-target (orange) outcomes while IF's flash encodes both on-target (blue) and missed-target (orange) outcomes (an unequal feedback-valence confound). The authors selected Option (b) (requested and completed a further frontend revision) rather than leaving these open, but elected to retain both asymmetries in the resulting final build as disclosed, accepted limitations rather than eliminate them; they are distinct from the already-disclosed 3DoF/proximal-EF and head-locked-anchoring limitations above. Per Section 6's directional-bias analysis, these are not expected to bias all six hypotheses in the same direction: dwell-time asymmetry plausibly inflates an EF advantage on H2a (and possibly H2c), while valence-coding asymmetry plausibly works against an EF advantage on H3a/H3b (and may inflate an EF retention decrement on H2b/H2c via feedback-dependency); a positive or null EF-vs-IF finding on H2a–c/H3a–b cannot, on this build, be fully attributed to attentional focus alone independent of these two implementation asymmetries, and the pattern described above (not a uniform bias) will be stated explicitly in the eventual manuscript's Discussion.

Deviations from this pre-registration: Any deviation from the analysis plan above, discovered necessary after data collection begins, will be explicitly reported and justified in the manuscript as a post-hoc/exploratory departure from the pre-registered plan, not presented as originally planned.