# Industrial Operating Evals: feasible-control v2 Executed 2026-10-13 on included private cloud CPU. There is no independently verified leaderboard or field qualification. A public operating challenge is available; community difficulty remains unverified. ## Three useful operating tasks - New executable delayed multivariable controller evaluation: 96 final episodes, 46,080 decisions, two regimes, six core conditions and two separate bias stresses. - Existing plume task: 342 tiles / 44 flights / three dates, per-date indication budget plus weak localization. Retrospective consumed labels; actual complete participant GDAL execution remains a gap in the original grader qualification. - Existing well task: eleven original references and consumed initial confirmation, 80,534 decisions, false normal alarm episodes, occupation and observable-onset coverage. No new well fitting, inference or reserved-group access occurred here. **0/17 organizer references clear all their declared gates.** These are three new final control entries and fourteen retained historical plume/well development entries, not seventeen new independently trained models. The joint operating failures are evidence of a research gap, not a claim that experts cannot ace the tasks or that these are official frontier benchmarks. ## Measured final control results | Reference | Worst core-group mean true RMSE (cm) | Sampled all-tank envelope violations | Gate failures | |---|---:|---:|---| | hold | 2.389223 | 0 | worst_group_tracking | | selected_pi | 1.258971 | 258 | worst_group_tracking, sampled_envelope_violations | | selected_nominal_mpc | 1.078048 | 151 | worst_group_tracking, sampled_envelope_violations | The research target is at most 0.50 cm in every core group, zero all-four-tank sampled envelope violations, and zero invalid, deadline, process or quality-hold failures. Group means average two episode RMSE values. Paired seeds and numerous decisions are not additional independent plants. True/observed bias-stress results remain separate. Only fully passing development candidates can be ordered by worst-group RMSE, macro RMSE and variation; official rank is null. ## Difficulty is qualified against achievable requests V1's post-run mass-balance audit found two exact steady-infeasible requests among 120 checks, both after a valve change. V1 remains an unchanged diagnostic prototype in `V1-PROTOTYPE-KIT.zip`, with its analytic audit and original failure receipts. Its stronger-reference failures do not prove a feasible control target. Commit `0863d6b` froze this separate public cohort and common steady-feasible target generator before new outcomes. No controller settings or operating bars were retuned or loosened. Each target is exactly reachable under all four public nominal/flow/pump-loss/split-change corners, inside voltage and all-four-level bounds. No actual future fault identity selects a target. All 64 analytic checks replay, with maximum mass-balance derivative residual 6.59e-17 cm/s. This certifies static targets, **not** dynamic RMSE feasibility, safety or stability. ## Actual execution and participant contract One new included private CPU v1 completes the feasible-control study at exact accepted script SHA `fc68d44e6f4bb5cd4bfb39ede0fc8e933ae5b0f926b463731962fbbe9b2ff550`. The original controls ZIP SHA is `e9b56f3aa46e5ec1d43219693c29af938090a8931766e2b594c8dde9fca38a44`; all 115 members verify. Three separate fresh grader processes reconstruct all 46,080 nonlinear transitions and exact causal requests from the saved actions, at unchanged atol 1e-10 / rtol 1e-9. Full traces remain in the original private cloud archive `f70b1dee016663f213138db08908f1f1f8313995ad7833ff0e592a25544631bf`; the public release contains their hashes/control metadata and complete runnable source, not those full traces. No raw physical images, sensor recordings or learned weights are in the kit. The actual JSONL participant receives only delayed lower readings/quality, applied prior commands, nominal prior and the current reference. Bad/late commands are counted and held; no case is dropped or invalid action clipped into passing. The transport is cooperative, not an OS sandbox. No hidden scoring service runs. The separately committed eight-seed synthetic reserve remains unopened. Verified isolation and fixed independent entries are required before private evaluation. It does not substitute for independent physical confirmation. CPython 3.13, NumPy 2.1.3, SciPy 1.15.3 and OSQP 1.1.3 use one BLAS thread. The three measured harnesses total 627.015 s, under the unchanged 1,800 s combined CPU cap. They include V1, its wrapper-only MPC recovery and this separately frozen cohort. Paid compute, GPU/TPU, fitting, new calibration and private reserved well evaluation are absent. Included-host response deadlines are not PLC timing/WCET or target-device qualification. ## Original faults remain visible The first provider derived an unexpected title-based handle; the local identity assertion failed. Exact accepted source/private CPU version 1 was subsequently verified, without a second launch. The original worker completed, but the MPC solver emitted `Polishing not needed` on stdout, causing 15,346 participant errors. This is an engineering failure, not a scientific controller loss. An actual authored-equilibrium probe reproduces that diagnostic corruption. The fixed wrapper redirects native diagnostics to stderr without changing solver or equations. A separately frozen recovery completes and retains the original. Its first creation request failed HTTP 400; same-owner nonce search found no accepted job. A shorter fresh creation title/id was accepted with identical source. The specific provider validation cause was not retained. Failed checks/requests are not relabelled as successful, and both original/control-recovery archives are included separately with their scientific source histories. ## Limits and the next confirmation gates Research operating bars were organizer-selected using previous public evidence; the new control cohort is synthetic public development. Source selection can still encode generator assumptions. Three-date plume rotation includes later-to- earlier calibration, source-label incorporation bias and released test-loss monitoring. Mask-negative is not chemically verified absence. The well onset target is an added retrospective organizer target; the original budgets are one false alarm per eligible normal day and 1% normal occupation. Unknown exposure, censored onsets, carried-in alarms, sparse filename groups and original provider ERROR/report-only recovery remain explicit. Independent submissions, independently sourced untouched physical acquisitions, chemical no-leak truth, measurement calibration, target-device timing/recovery, site acceptance, adoption and qualified-expert growth are not established. No plant writes, robot motion, profile edits, outreach or client architecture publication occurred. Per-artifact MIT/GPL-3.0/CC BY 4.0/CC BY-NC 4.0 rights are declared in `LICENSES.md`. ## Public packaging privacy review The original scan failed on an inherited local-only cloud submission receipt. The public packages omit that receipt, which is not used by either scorer. Every remaining prototype member and all frozen scientific code, protocols and results are unchanged. The unmodified originals remain privately retained. See PUBLIC-PACKAGING.json for the subset hashes; the failed scan is preserved.