Reproduce the Loop analysis

Loop studies fine-grained wrist motion through crochet. This guide accompanies the single-participant measurement report and its released numerical supplement: five CSV tables, three Python scripts and 72 data rows. It explains how to check the calculations and what those checks establish.

Run the checks

Use Python 3.9 or newer. The verifier uses only the standard library; no SDK, API key, account connection or additional Python package is needed to run it.

From the supplement folder, run:

python3 code/verify_tables.py

The command reads the five CSVs in data/, prints JSON results and exits successfully when the recorded comparisons and aggregate checks agree. It does not modify data, upload files, train a model or open a recording. Keep the folder structure intact: paths are resolved relative to the script, rather than a particular computer's directories.

Check file integrity separately. On macOS:

shasum -a 256 -c SHA256SUMS

On Linux:

sha256sum -c SHA256SUMS

Every listed file should report OK. Checksums detect changes relative to that manifest; they do not establish the scientific correctness of the labels. Use the supplement's exact release revision when reporting a reproduction. Version 1.0.0 is publicly downloadable from the frozen supplement revision. Scripts are licensed MIT; tables and package documentation are CC BY 4.0. See Resources and access below.

Read the tables

Session identifiers. H01–H06 are legacy software-timestamped recordings, H07 is an earlier hardware-timestamped recording, and V01 is the synchronized ten-stitch video pilot. All concern the named author. The identifiers replace operational filenames; they do not make the study anonymous.

Timing fields. samples counts saved samples; sampleSpanSeconds is the last saved timestamp minus the first. effectiveHz is (samples − 1) / sampleSpanSeconds; observedMedianHz is the inverse median adjacent interval. intervalJitterCvPercent is 100 times population interval SD divided by mean interval. format_version identifies the original export schema.

shortIntervalsBelowHalfExpected and longIntervalsAboveOnePointFiveExpected count intervals below 0.5 or above 1.5 times the declared interval: 50 Hz for H01–H06, 200 Hz for H07/V01. Long gaps are not direct counts of physically missing sensor samples. captureMarks is the unfiltered capture-mark count. captureMarksOutsideSampleRange checks coverage separately from interval regularity. secondsFromLastCaptureMarkToLastSample is negative when recording ends before the final mark; a blank indicates no capture mark was available.

Review fields. source distinguishes human_A, human_B and AI; the two human passes are by the same maker. event_index is one-based within a review, rather than a validated stitch identity. video_seconds is elapsed source-video time. status is clear for the twenty human events and uncertain for all ten AI proposals. earliest_plausible_seconds and latest_plausible_seconds contain AI judgment ranges; blank human values mean no interval was recorded, not zero uncertainty. These ranges are not statistical confidence intervals.

Synchronization fields. anchor_index identifies a tap pair; video_seconds and watch_seconds are elapsed coordinates. No wall-clock times are included.

Exploratory fields. The autocorrelation table contains session_id, retained_marks, median_mark_interval_seconds, selected_period_seconds, autocorrelation and a 1/0 meets_source_criterion flag. Printed periods and strengths are rounded; the flag preserves the original full-precision strict 20% criterion. Do not redefine it from the rounded columns. In the control table, condition is actual or randomized, trial is 0 for actual or 1–20 for randomized draws, and mean_correlation is the in-sample template score.

The supplement README documents the same fields in detail. There are no training/test splits, raw motion channels, videos, audio, still images, private AI traces or personal context records in these tables.

Understand the measurements

Capture marks, reviewed events and context remain separate. A mark made during recording is not an independently verified stitch completion. Later video events and retrospective observations retain their own provenance; this package does not replace them with reconciled ground truth.

The event definition is explicit. Completion is the first visible frame after the final pull-through leaves one loop on the hook. The maker knew the intended count when reviewing the muted video twice. The AI reviewer used cropped source-frame sheets and different viewing conditions, with instruction-based blinding. The report describes those limitations. All AI boundaries remain uncertain.

Clocks are fitted before comparing coordinates. Ordinary least squares fits Watch time = offset + scale × video time from the six tap pairs. Root-mean-square and maximum residuals describe the anchor fit; they do not verify absolute boundary timing. All review comparisons shown here use video time.

Event pairing allows unmatched events. An ordered, one-to-one dynamic-programming alignment uses a maximum distance equal to the smaller of one second and 45% of the median within-review inter-event interval, with a 0.5-second fallback when no interval is available. A one-second cap is the pairing rule for this data, not an independently validated accuracy tolerance. No offset is fitted to make the reviews agree.

Controls retain their original limits. The historical procedures use preceding-sample resampling at 25 Hz without an anti-aliasing filter. The randomized-center procedure fits and scores overlapping evidence in-sample, using eight template refits and seed 7 for twenty controls. Its random-draw SD is not a confidence interval. The original raw-signal inputs are withheld, so the public-table verifier cannot re-extract those signals or independently repeat the AI review.

Interpret the results

Successful verification reproduces these video-time comparisons:

The six-tap fit reproduces RMS residual 20.368 ms and maximum residual 40.819 ms. The actual-mark control score rounds to 0.526, compared with randomized mean 0.525 and population SD 0.024 across 20 draws. Four of seven historical autocorrelation peaks met the original pace criterion.

Agreement with preserved numbers establishes numerical reproducibility within the supplied tables. Human repeatability on one clip does not establish semantic accuracy, independent inter-rater agreement, generalization or detector performance. The 30 review events describe one ten-stitch clip, not 30 independent stitches. There are no clear AI boundaries for a clear-only accuracy analysis.

Resources and access

• Measurement report: published research report on capture timing, signal controls and stitch-boundary annotation.
• Numerical supplement, version 1.0.0: public tables, package documentation, citation metadata, scoped licenses and checksums.
• Analysis code: the three Python scripts and MIT notice, pinned to the same release revision.
• Loop collection on Hugging Face: project resource index, including the released supplement.
• Natalie Hill's ORCID: researcher identity and public works.

The source analysis revision for the report is eeb169543cf617dbb43ec761524ba5dcc0d35468; the public supplement revision is d83d8777bb11a17a2bf348bf90b5c85e627de140. Numerical tables and scripts retain the reviewed candidate's exact bytes; release changes add citation/license metadata and documentation. Cite the report and exact supplement revision used, following report.bib and CITATION.cff in that package. No DOI is claimed. Tables and package documentation are CC BY 4.0; scripts are MIT. No original media or raw sensor recording is included or licensed by this numerical release.