# Ethics This measures recordings of identifiable people at work, in order to say something about their working conditions. That is a different act from measuring a dataset, and it needs its own basis rather than an inherited one. ## Why `vernier`'s basis does not extend to here `../vernier/docs/ETHICS.md` states: *"No claim about consent, compensation, or labour practice. vernier has no evidence either way and will not infer from absence."* Its entire posture is that it measures the dataset, not the people. **cyclegraph measures the people.** A statement about hand-activity exposure is a statement about labour practice — precisely the category `vernier` disclaims. Quietly widening that document to cover this work would damage the credibility of both. This is a new basis, stated here, and it is narrower in what it permits than what it replaces. ## What is knowable, and what is not From the public release: 10,000 hours from 2,153 workers as published — 2,144 as shipped (`../vernier/docs/UPSTREAM-FINDINGS.md` F12) — across 85 factories, captured on the vendor's head-mounted devices in production environments, released under Apache-2.0, gated on contact details. Not stated, and therefore not knowable: the consent instrument or its language; whether workers were compensated for downstream commercial use; whether declining carried consequence; retention, deletion or withdrawal policy; how the 85 sites were selected. ## The basis **Consent to be recorded for AI-training data collection is not consent to occupational health assessment.** Even a fully informed consent to the first does not reach the second, because the second produces a claim about the person's body and their employer's process that the first gives no reason to anticipate. Apache-2.0 is the vendor's licence to grant. It is not a worker's consent, and section 2 records that the consent instrument is unknown. **Therefore: nothing is reported at a unit any person or site could be identified by.** ## What that forbids, concretely - **No per-worker result, ever.** `worker_id` is load-bearing for every interval — clips from one person are not independent observations — and appears in no published number. - **No per-factory result either, and no stratum below the floor.** This goes further than `vernier`. Factory-level aggregation protects individuals while exposing *sites*: "this plant shows exposure above the action limit" is a finding about an identifiable workplace whose operators never participated, and the people most exposed to the consequences are the workers there. Eighty-five sites are not meaningfully anonymous, and the corpus metadata carries no sector field that would let a sector be named either. The only unit between corpus and nothing is a **factory-size tercile** that clears ≥5 factories, ≥50 workers and no factory above 40% of its workers or clips, with one `corpus_rev` per release so no two published aggregates can be differenced (`docs/DECISIONS.md` D019). A stratum that fails the floor is printed as suppressed, not omitted. - **No pilot value.** The pilot factory is nameable by construction; its gates publish pass or fail only. - **No ranking of anything.** Not workers, not sites, not shifts. - **No re-identification of any kind.** No face recognition, no cross-clip linkage, no linkage to any external source. `CONTRACTS.md`'s `ExposureAggregate` has no field that can carry an identifier, and `docs/ARCHITECTURE.md` names the seam where this would otherwise leak. The rule is enforced in the schema because a rule enforced by intention survives exactly as long as attention does. ## The dual-use problem, named The pipeline that computes hand-activity exposure computes, with no modification, a per-worker productivity signal. Duty cycle is how much of the time someone's hands are working. An employer wanting to rank workers by that number needs nothing this project would not hand them. The aggregation floor is the mitigation and it is the only one. It is worth being explicit that it is a *design* mitigation, not a technical impossibility: anyone with the corpus and this repository's method can compute the per-worker number. What this project controls is what it publishes and what its own code will emit, and it controls both. For the same reason this project does **not** ship a "run it on your own site" tool. An instrument a site runs on its own footage is worker surveillance unless the workers, not the site, consented to that use; nothing here can inspect that consent, so nothing here offers the path. The finding is aggregate, and the way to make it actionable is a study run with the participation of the people concerned — a different study with a different document. ## The argument against not doing this Suppression is not the neutral option and should not be treated as one. Occupational health research uses observational data about workers precisely because the alternative is that injury patterns stay invisible until they present as injuries. If a corpus of factory hand work shows exposure above recognised action limits, that is information with a claim on being known — by the people exposed most of all. The position this project takes is that the finding should exist and the individuals should not be identifiable in it. That is a real trade and it is not costless: aggregate-only reporting means a site with a genuine problem cannot be told it has one. That cost is accepted here, with the reason, rather than left implicit. ## What this project does not say It takes no position on Build AI's collection practices. It draws no conclusion the public artifacts support. Reporting on egocentric data collection elsewhere in the industry exists and is not evidence about this vendor. ## Conflict of interest The author has an interest in this work being noticed. That interest is disclosed, and structurally contained: hypotheses, thresholds and stopping rules are frozen in `docs/PRE-REGISTRATION.md` before any clip is decoded, so no finding can be shaped by how it would be received. `docs/RED-TEAM.md` states the containment and its limits.