What to learn before you train

Abstract

This paper reports no trained model. It reports a formulation: what an organizational trajectory model should predict, what data it needs, which of several machine-learning approaches — a Joint-Embedding Predictive Architecture among them — actually fits the shape of that data, and how such a model would have to be made operational inside the substrate described in my first report. I argue the target is not a model of people. It's a model of coordination — the unit of analysis is an initiative, not an employee — and I give the contrast that makes that distinction operational rather than aspirational. I connect the design to an argument I made earlier, outside this series, about software becoming an intent codec: a system that ships less because its decoder understands more. A predictive model of organizational state is the same move, applied to coordination instead of video. I lay out a four-stage build plan in which JEPA is stage three, not stage one, because the earlier stages are the same-information comparison a JEPA claim needs to survive before it's worth believing. Four independent adversarial reviews of the first hypothesis draft each found a distinct, real problem — a possible label-timing circularity, a feedback loop specific to models that inform the reviewers who generate their own training labels, a missing information-parity control, and a training-data population that, on inspection, probably doesn't exist yet in usable volume. None of these findings are fatal. All of them are reasons to publish the formulation and postpone the training, which is what this paper does.