Core Missing-Physics Support Recovery without a Supplied Scientific Feature Dictionary on the Lorenz-Stenflo ANI Benchmark

This study evaluates a restored archived OrganOS interaction-first pipeline on the Lorenz-Stenflo missing-physics benchmark used by Alternating Neural Integrators (ANI). The central question is whether the core variables required by the missing physical field can be recovered without supplying the legacy 14-feature scientific dictionary or the analytic missing-law expression. Candidate generation used 2,000 TRAIN state-transition pairs, the callable incomplete ANI prior, and the observed time interval. Anonymous state variables q1–q4 and all unordered pair products were generated before dependency compression. The blind frozen candidate was {q1, q4, q2*q4}. After unblinding, q1=x and q4=w corresponded to the complete core variable support required by the true missing field (1.5w, 0, 0, -x-w). The additional q2*q4=y*w interaction was not supported by remove-one ablation and is therefore not interpreted as an independently supported missing-physics term. Selection stability was evaluated using 50 pre-specified 80% TRAIN subsamples. The exact frozen candidate recurred in all 50/50 runs. Under 1%, 5%, and 10% synthetic observation noise, x and w were jointly recovered in all 50/50 repeats at every tested level, although auxiliary-set sparsity degraded as noise increased. After fitting the frozen support from trajectory pairs using continuous shooting, the OrganOS model achieved a mean relative L2 rollout error of 1.76575×10^-4 on the reproduced 1,284-window Lorenz-Stenflo evaluation, compared with 5.73833×10^-3 for reproduced ANI-4 symbolic write-back and 4.62480×10^-1 for the incomplete ANI prior. The contribution is benchmark-specific: OrganOS selection recovered the essential missing variables without the legacy scientific feature dictionary, while trajectory-pair continuous shooting accurately fitted the correction on that support. Universal missing-physics discovery and superiority over the full ANI neural framework are not claimed.