Institutionalized disagreement

Abstract

A language model can produce a plan. Nothing about producing a plan establishes that it is sound, and nothing about it being sound establishes the right to execute it. This paper describes Organism, the reasoning layer that sits between human intent and the governed engine described in our previous report, and the six mandatory stages an intent passes through before any of it reaches the commit boundary: admission across four dimensions, decomposition into an intent tree, a huddle in which several reasoners plan in parallel and failures are dropped rather than hidden, adversarial review by five kinds of skeptic, simulation across five dimensions, and commit. Two invariants carry the design. Authority may only narrow down the intent tree, so no subtask can outrank its parent; and learning flows backward into planning priors while authority flows forward, so that no learned quantity can ever appear in an authority computation. We show why the second is not a stylistic preference but a precondition for the determinism proved in the substrate: an engine whose authority depends on learned state has a fixed point that moves. We also take the adversarial stage further than our implementation goes, and argue that aggregating skeptic verdicts is a judgment aggregation problem in the sense of List and Pettit — with a discursive dilemma that a majority vote cannot escape and a veto can. Whether the review loop terminates without a budget is stated as an open problem, not a result.