Representing Organisational Judgement for Artificial Reasoners: An Engineering Essay on Epistemic Engineering

Engineering work depends upon standards for choosing among technically admissible implementations that are often carried through review, correction, apprenticeship and conversation rather than through durable writing. Artificial reasoners do not participate in those channels. The resulting tension is practical even for a solo developer: if local judgement has not been expressed in a form a model can read, the model cannot take account of it when proposing a solution. This essay examines the engineering work of making selected judgement durable and reconstructable, provisionally calling that work epistemic engineering. Evaluation is framed around expressibility, accessibility, fidelity, selectivity, economy and transfer rather than a privileged document genre. Controlled studies on pinned Qwen systems establish a bounded mechanism: withholding local judgement produced selective under-performance; several accessible representations restored it; an incompatible local practice induced predictable errors. Proposition-matched checklists performed as well as the tested stem forms. A joint powerset instrument under contrasting practices then asked which sets of written components are sufficient: on one small open-weight stratum, participation-like episodes alone formed a unique minimal sufficient set with held-out agreement, while compact doctrine alone did not; a larger stratum shifted and failed the frozen criterion under a constructor letter prior on negative controls. These results make written expressibility, accessibility and fidelity engineering concerns and show that combinatorial economy and held-out transfer can be measured without inventing a privileged document genre. They do not establish a general organisational theory or mature new discipline.