We present DOORM WorldModel-OS, a hybrid local/cloud operating system for auditable agentic world models, paired with the model-agnostic Finetune Workbench. The design rejects two dominant assumptions in contemporary large-model practice: (i) that capability is monotone in parameter count ("scale is all you need"), and (ii) that alignment can be bolted onto a post-training checkpoint. Instead, we factor the system into seven cognitive increments, grounded in a discrete state space built from an 8-mode base prior, a 64-state composite obtained by Cartesian self-pairing, and a 6-position ordinal extension yielding 384 anchor points. Safety is enforced through a four-level rollback contract embedded in every inference trace. Training uses a five-class grounded-citation data pipeline with three hard gates (citation, consistency, compliance), a 3%-degradation bucketed-evaluation rollback protocol, and an immutable adapterᵥersion handshake between platform and workbench. We motivate the design through medical-adjacent applications — Traditional Chinese Medicine syndrome differentiation and rehabilitation robotics — where per-decision auditability and refusal-by-default beat benchmark-scale gains. No empirical results are claimed in this version; we specify the reproducibility envelope, three hard release gates, an explicit non-coverage list, cross-domain migration cost estimates, and falsification criteria against which our positions can be refuted. Author order reflects relative contribution, not academic seniority. All economic rights are held by DOORM AI PTE. LTD. pursuant to a written Author Agreement.
Tong et al. (Wed,) studied this question.