luccilabs/loop-measurement-supplement

Loop measurement tables and analysis code

Version 1.0.0 of the numerical tables and analysis procedures for Loop. The accompanying research report and reproduction guide are published by Lucci Labs. This package contains no original media or raw sensor streams.

Natalie Hill, Lucci Labs. ORCID. Project index: Loop collection.

This supplement accompanies Measuring fine-grained manual activity from the wrist, a single-participant research report. It contains five numerical tables (72 rows total), a standard-library table verifier and the two preserved exploratory signal-analysis procedures. It is sufficient to recompute the event-pairing, synchronization-fit and control-summary arithmetic. Raw sensor signals are withheld, so it cannot reproduce signal extraction or establish video semantics from independent inputs.

Scope and provenance

Eight recordings comprise six legacy software-timestamped sessions (H01–H06), one earlier hardware-timestamped session (H07) and one synchronized ten-stitch pilot (V01). All concern the named author, who was also the crocheter and human annotator. These identifiers remove operational filenames; they do not make the study anonymous. H01–H07 retain chronological source order.

Tables were extracted from preserved originals and checked against the October 18 evidence reproduction. Source analysis revision: eeb169543cf617dbb43ec761524ba5dcc0d35468 in the restricted Loop repository. The table verifier was added for this numerical supplement. The two historical scripts are byte-identical to their preserved source at that revision. Randomized-control values retain full precision from the original procedure rerun with seed 7. No new capture, model training, annotation or adjudication was performed.

Excluded: video, audio, still images, frame sheets, raw inertial channels, wall-clock times, original recording filenames, participant context records and private AI traces. Event and tap times are relative elapsed seconds. No public original-media release is proposed.

Files and fields

data/session_timing.csv — eight rows

• session_id: H01–H07 or V01, as defined above.
• format_version: original Watch export schema version (1 or 2); not a participant or class label.
• samples: count of saved motion samples; interval count is samples minus one.
• sampleSpanSeconds: last saved timestamp minus first, seconds.
• effectiveHz: (samples minus one) divided by sample span.
• observedMedianHz: inverse median adjacent timestamp interval, Hz.
• intervalJitterCvPercent: 100 times population interval SD divided by mean interval.
• shortIntervalsBelowHalfExpected: intervals less than 0.5 times the declared interval (50 Hz for H01–H06; 200 Hz for H07/V01).
• longIntervalsAboveOnePointFiveExpected: intervals greater than 1.5 times that declared interval. These are not physically missing-sample counts.
• captureMarks: original capture-time mark count, before exploratory filtering.
• captureMarksOutsideSampleRange: marks before the first or after the last saved sample.
• secondsFromLastCaptureMarkToLastSample: last sample minus last capture mark, seconds; negative indicates the stream ended first. Empty when there is no mark (V01).

data/review_events.csv — thirty rows

• session_id: V01 throughout.
• source: human_A and human_B are two passes by the same maker; AI is a separate visual-only candidate review.
• event_index: one-based chronological index within that review; not an independently verified stitch identity.
• video_seconds: elapsed source-video event time.
• status: clear for the twenty human events; uncertain for all ten AI proposals.
• earliest_plausible_seconds, latest_plausible_seconds: AI-proposed uncertainty interval; empty for human events because no interval was recorded. Empty does not mean zero uncertainty.

The completion definition was the first visible frame after the final pull-through leaves one loop on the hook. Human passes knew the intended count. AI confidence and viewing conditions differed; candidate timestamps are not ground truth. The report documents instruction-based blinding and its limits. No inter-rater accuracy or detector-training label set is supplied.

data/sync_anchors.csv — six rows

anchor_index is one-based; video_seconds and watch_seconds are elapsed tap correspondences used in the affine fit: Watch time = offset + scale times video time. Fit residuals do not establish absolute event-timing accuracy.

data/autocorrelation_summary.csv — seven rows

session_id, retained_marks, median_mark_interval_seconds, selected_period_seconds, autocorrelation, meets_source_criterion (1/0). Marks within two seconds of the last retained mark were merged only for this analysis. Periods and strengths retain the historical printed precision. The flag uses the original full-precision strict 20% criterion, not a new calculation from rounded table values. Four of seven meet it. There is no null test or independently adjudicated period reference.

data/randomized_control.csv — twenty-one rows

condition is actual or randomized; trial is 0 for actual and 1–20 for the seeded random draws; mean_correlation is the original in-sample template score. Template fitting and scoring share windows. Population SD across random draws is not a confidence interval or held-out accuracy.

Reproduction

Use Python 3.9 or newer, standard library only:

python3 code/verify_tables.py
sha256sum -c SHA256SUMS

On macOS, the checksum command is shasum -a 256 -c SHA256SUMS. The verifier recomputes ordered event pairing with the original distance cap, affine synchronization and numerical aggregates; it checks the report's rounded results. It reads only these five CSVs and writes results to stdout.

code/historical_periodicity.py and code/historical_mark_control.py preserve the exact exploratory algorithms. They require the original Watch JSON schemas and run from a directory containing the relevant exports. Those exports are deliberately absent here. They are included for methodological inspection, not a claim that public tables can reproduce raw-signal analyses. Both resample preceding values at 25 Hz without an anti-aliasing filter. The template procedure is in-sample, with eight refits and approximately ±1.5 seconds of adjustment; seed 7 controls twenty random-center draws.

There are no training/test splits. Each table is exposed as a separate all configuration to avoid pooling incompatible schemas. Adjacent events and repeated reviews are not independent observations.

Citation and reuse

Report citation: Hill, N. (2026). Measuring fine-grained manual activity from the wrist [Research report]. Lucci Labs. See report.bib; supplement metadata is in CITATION.cff. No DOI is claimed. The reproduction guide explains the commands, schemas and measurement limits. The original version 1.0.0 package is frozen at revision d83d8777bb11a17a2bf348bf90b5c85e627de140; subsequent publication-link metadata changes do not alter the five tables or three scripts. Public code is included in the code folder.

The five numerical tables and documentation are licensed CC BY 4.0; the three Python scripts are licensed MIT. See license scopes, the CC BY 4.0 legal terms and the MIT notice. Original media, application code outside this package and private traces are excluded from both the upload and these license grants.

The study supports a limited measurement case, not population generalization, a validated stitch detector or an independently adjudicated training benchmark. Redistribution or reinterpretation must preserve the distinction between human repeatability, uncertain AI proposals and semantic accuracy.