Conflux Laboratory builds the computer-vision models that detect children in monitored agricultural sites. The aim is to give the Child Labour Monitoring and Remediation System (CLMRS) a continuous evidence channel it does not have today — as a research prototype, in which every finding is a lead for a human reviewer and never a conclusion. The models are trained and run by the lab and have completed a seventeen-day unattended field pilot in Zimbabwe with EthicalChain, on-device so that sensitive imagery never leaves the site.
Why detecting children is the hard part
Child labour is overwhelmingly an agricultural and supply-chain problem, and the sites most affected are the hardest to observe. Monitoring has relied on periodic household visits — slow, easy to miss children who are out working, and dependent on self-reporting that tends to understate the issue. Cameras can watch continuously; the difficulty is teaching a model to do the one thing that matters here, which is to recognise that a person on site is a child.
That turns out to be the hard case for computer vision. General-purpose detectors are trained on imagery dominated by adults, so children — smaller, differently proportioned, often partly hidden behind equipment or other workers — are exactly what they miss most. And the decision that protects a child is a narrow one: separating, say, a thirteen-year-old from a young adult. A model tuned for average accuracy across all ages spends none of its capacity where the entire value of the system lies. This is why Conflux Laboratory builds its own models for the task rather than re-using an off-the-shelf detector.
What we built
The lab develops, fine-tunes, and runs two models, deliberately kept separate because they answer two different questions. We describe them by function rather than by architecture:
- A person-and-face detector, fine-tuned for children. We took a modern real-time detector and re-trained it on child-rich imagery so that small, partly occluded, and oddly-posed children are found rather than skipped. It locates both the whole person and the face in every camera frame.
- A two-stage age-estimation cascade, with a child specialist we trained. A general age model runs first; whenever it returns a young estimate, a second model — fine-tuned by the lab specifically on children — re-estimates the age. Concentrating a dedicated model on the under-eighteen boundary is what makes the estimate trustworthy exactly where it counts.
- Tracking, de-duplication, and an evidence record. Detections are tracked across frames and matched across cameras so one child is not counted many times, and each candidate produces a structured record — a best-shot image, the age estimate, a confidence tier, and timestamps — for a human to review.
Everything is optimised to run on a single device at the site, in real time, with no dependence on the cloud.
How well it performs
Two kinds of number matter here, and they are not interchangeable. The first come from annotated benchmarks, where ground truth exists and accuracy can be measured directly. The second come from the field, where the question is not how good a model is on curated images but how a whole system behaves over weeks on a working farm. The benchmarks are below; the field results have their own section, because they are the more informative — and the less flattering — of the two.
Data stays where it is collected. The models run on-premises; video is never sent to the cloud. Only structured detection records — not raw footage — leave the device, and only into the partner's audited workflow. This is a design constraint, not an optimisation: it is what makes the work defensible in settings involving minors.
Where it fits — CLMRS
The Child Labour Monitoring and Remediation System (CLMRS) is an internationally recognised framework, aligned with ILO conventions, for finding children in child labour and responding with remediation rather than punishment — through awareness-raising, monitoring, identification, support, follow-up, third-party verification, and partnership with governments. The lab's models sit at monitoring and identification: they raise well-evidenced candidates, and a person always decides. The model never determines on its own that a child is in child labour.
Partnership with EthicalChain
The detection models are one input into a wider compliance workflow operated by EthicalChain, which digitises the operational side of CLMRS. EthicalChain takes findings from three channels — public reports, auditor observations, and automated device detections — and routes them through a structured review in which a human auditor assesses the evidence. Verified findings feed certification and supply-chain traceability, so a confirmed problem has a defined consequence rather than disappearing into a report. The lab is responsible for the models and their evaluation; EthicalChain provides the platform, the auditor workflow, and the link to certification.
Field pilot — Zimbabwe
The system ran unattended for seventeen days on a working farm, six cameras on a single edge device, and was checked each day against the site's own attendance register — an external ground truth for who was actually present. To our knowledge this is the first published account of a pipeline of this kind running unsupervised at a real site with children present, measured against a register rather than against itself.
The pilot changed where we think quality actually lives. With the models frozen, software tuning alone raised the child-detection rate 36-fold — the untuned system had been missing roughly 98% of visible children. And the dominant limit on the whole deployment turned out not to be model capacity but camera placement: one of the six cameras yields a usable face for 35–39% of the people it tracks, the other five for under 2%. Remounting those five is worth more than any model improvement.
Three findings set the boundary of the method, and we would rather state them here than have them discovered later:
- The youngest children are beyond it. Carried infants are detected as geometry but produce no usable age signal from either the face or the body path — a face pressed against a carrier yields nothing to estimate from. This is not a tuning gap; there is no information to recover. The route forward is to send the carry geometry itself to a human.
- Age cannot be verified from the evidence. Child faces from the best camera are 22-to-38-pixel backlit profiles, on which neighbouring age bands are indistinguishable to a human reviewer. A filed record is therefore a lead for human investigation, never evidence of age.
- Where to set the threshold is not ours to decide. The deployed recall-first setting measures precision 0.64 at recall 0.83; a stricter one reaches 0.90 at 0.60 for the same F1. Which error costs more — a missed child or a wasted review — is a child-protection policy question, and it belongs to the partner running the programme, not to the people who built the model.
One correction is worth recording, because it is the kind that usually goes unpublished. An explanation for the residual over-reporting fitted the first four days of register data at a correlation of 0.90; six more days dropped it to 0.20. Four points had been enough to convince us, and they were not enough. The real driver remains unidentified, and it is reported that way in the paper.
Publication
The method and its evaluation are written up in full and posted as a preprint — the benchmarks, the field pilot, the data-protection and human-in-the-loop safeguards, and the limits above stated as limits. It also documents what did not work: two training approaches that made the model worse, a quantisation toolchain that failed silently, and a sequential A/B test of camera settings that produced a confidently wrong answer until it was re-run interleaved. A methods paper that reports only its successes is not much use to anyone trying to replicate it.
The paper is a preprint: it has not yet been peer-reviewed. The figures on this page are the ones it reports, and both should be read accordingly — as measurements from a research prototype, with the limits and failure cases set out in the paper itself.
- Paper
- Nowak, M. (2026). Artificial Intelligence as a Tool for Combating Child Labour: A Real-Time Edge Vision Pipeline for Child Detection and Age Estimation. Conflux Laboratory.
- Preprint
- arXiv:2608.14770 — submitted 2026-08-14, cs.CV / cs.AI.
- DOI
- 10.48550/arXiv.2608.14770
Ethics
This is work with imagery of children, so the responsible-data principles the lab applies elsewhere are non-negotiable here. Footage is processed locally and not retained beyond what review requires; identifiable images of minors are not published or moved off-site; and the system is positioned to trigger remediation, never enforcement against families. As with every project the lab accepts, the three questions in our ethics statement — whose decisions this changes, whose data is used and on what basis, and what happens if the model is wrong — were answered before the work began.