Quality Gates for Cooperation-Mode Labels: Detecting When a Behavioral Log Cannot Support Supervised Learning

Platforms that mediate cooperation (mutual-aid networks, referral systems, community request boards) log the raw material for supervised models of how people cooperate: pass by, broker a request onward, originate a sharing chain, or help directly. These operational logs are also the data most likely to contain defects that make such labels invalid while leaving predictive performance intact: administrative accounts mixed with genuine users, features recorded after the outcome, outcomes driven by sheer activity level, and too few positives. We define four cooperation-mode labels over four common log tables and a four-gate certification procedure (exclusion, minimum support, leakage, and an activity-baseline test) that a label must pass before supervised use, reported through a fixed-schema data card. Because no cooperative-platform dataset comes with ground truth about its own defects, we evaluate the gates by defect injection into a synthetic log with a known generative process. Over 500 replications per condition, a rule that certifies whenever cross-validated AUC exceeds 0.70 accepts data carrying administrative contamination, test accounts, leakage or activity confounding in 97.6%–100.0% of cases, and a stronger practitioner baseline still accepts all but the leakage case at 97.6%–100.0%; the gates accept the same data in 0.0%–3.0% of cases while certifying clean data in 100.0%. Ablations show each gate is necessary for at least one defect. We also report where the gates fail: leakage that is not declared in feature metadata and is weaker than the univariate tripwire passes undetected. On a public help-request log whose post-outcome fields are documented by its creators (Random Acts of Pizza, 5,671 requests), the naive rule and the practitioner baseline certify models that rely on 15 and 14 post-outcome features (AUC 1.00 and 0.89); with the documented recording times the gates remove all thirteen documented ones and certify a model at AUC 0.67, which still carries two undocumented edit-related leaks worth 0.02 AUC, the failure mode our stress test predicts. Applied to eight months of logs from an operating mutualaid platform, the gates decline to certify all four labels and separate a sample-size shortfall from a missing log field that no amount of further traffic would fix. This version is a preprint and has not been peer reviewed. Code and data card: https://github.com/kazutakahayashiyorosix/quality-gates-cooperation-labels