Register independent unseen-object BayesianPhysTwin validation

[FlorianPfaff]

Objective

Test the bounded Bayesian discrepancy/covariance interface on genuinely unseen physical objects or independently acquired object sessions. This is the principal remaining scientific gate after the full-22 tournament retained last_residual and the existing PokeFlex result established same-object new-interaction transfer only.

Frozen statistical design

Use complete physical object-session groups as the independent units. Before target outcomes are opened, freeze:

• disjoint development, calibration, target, and—where feasible—confirmation object/session rosters;
• the unchanged physical fallback;
• last_residual as the principal matched deterministic comparator;
• at most one Bayesian candidate, without reopening the completed full-22 challenger tournament;
• the calibration-domain guard and domain covariance-calibration grid;
• horizons, physical queries, group weighting, bootstrap/permutation procedure, and random seeds;
• exact fallback and unsupported-domain behavior; and
• all repository, package, provider, model-set, and numerical-environment identities.

Frames, points, tracks, views, and taxels are not independent groups.

Endpoints

Primary endpoint: equal-object/session difference in a preregistered proper predictive score relative to last_residual and physical fallback.

Secondary endpoints:

• object-balanced Chamfer distance and track error;
• coverage and interval width by horizon and domain;
• harmful accepted-update frequency;
• exact-fallback frequency and reason;
• worst-domain and worst-object/session regret;
• object/session win count; and
• provider competence reported separately from downstream physical-query benefit.

Decision rule

Advance only if the frozen candidate improves the primary proper score, has no preregistered harmful group regression, passes the domain guard, and preserves exact fallback. A tie retains last_residual. A negative result is complete evidence and must not be retuned on the same target cohort.

Required artifacts

• protocol and roster identities;
• calibration and target access attestations;
• locked numerical environment and dependency resolver input;
• sealed predictions before target scoring;
• group-level score table and exact randomization/bootstrap record;
• content-addressed decision and claim bundle; and
• explicit paper-facing claim boundary.

Boundary

No existing opened MotionCrafter or Deform360 calibration cohort may be relabelled as fresh confirmation. Physical acquisition or access to a genuinely independent cohort is required before this issue can be completed.

[FlorianPfaff]

Frozen candidate from retrospective covariance-only development evidence

The registered full-22 covariance-only analysis in PR #485 completed successfully in workflow run 31461910994 (attempt 1) on analyzer revision 9ac5344036331a40e9029a1aa5814f601c0eaf15.

Evidence identity:

• artifact 9090224528, full22-covariance-only-hybrid-v1-31461910994-1;
• archive digest sha256:945dd5ed5db9b4119d81cf15d6d6c6da304b9421ab5938e71ad65390bbca1676;
• protocol ID 91c15957c8a71c579429c4245ccb2570f22ac6480dd5aebc9ddd73d1edb66b02;
• report ID 5fc777163fd6173c9669b497309d883e2780a5ebe23da5dbe4cdaf682ad8806a;
• result.json digest 31ed0828b0c024a508467904c56c040f4e38374c4688a06167f2e584b13111a0.

The primary cross-fitted covariance-only arm changed Gaussian NLL by -9.136 with a max-t simultaneous 95% interval [-13.961, -4.312] relative to zero-covariance last_residual. It was better in 17/22 object-session cases. The predictive mean was the exact same array object in every case, so Chamfer and track effects were exactly zero. Marginal 90% coverage changed from 0.706 to 0.910, with mean full width changing from 0.01645 m to 0.05094 m (3.10×).

This issue therefore freezes exactly one Bayesian candidate for a genuinely disjoint fresh-object/session study:

• mean: exact last_residual;
• covariance donor: independent_endpoint_v1;
• early/middle/late covariance scales: [8.0, 16.0, 16.0];
• deterministic comparator and fallback: exact last_residual with unchanged physical fallback;
• primary endpoint: the issue's preregistered proper predictive score at the complete object-session unit;
• secondary endpoints: calibration, width, Chamfer, tracking, harmful-update frequency, fallback behavior, and worst-group regret.

No further donor, scale-grid, horizon-bin, observation-noise, endpoint, or action-family tuning is permitted after fresh target access. Pushing was approximately neutral and the single weird_package case was worse in the retrospective audit, so the fresh study must retain the preregistered harmful-group and worst-group checks rather than restrict the cohort to favorable lifting actions.

Information boundary remains explicit: analysis_status=retrospective-cross-fitted-development-only, claim_authorized=false, selection_authorized=false, and promotion_authorized=false. This comment freezes a candidate; it does not establish fresh transfer or authorize a paper/deployment claim.

[FlorianPfaff]

Frozen covariance-only candidate for the independent validation

Merged PR #505 (247bc8b85ec425f4272ba867e3a7c878a7e05d56) completed the registered retrospective full-22 covariance-only study and freezes the following single candidate for a separately registered, genuinely fresh object/session evaluation under this issue:

• predictive mean: exact caller-owned last_residual array;
• covariance donor: independent_endpoint_v1;
• early/middle/late covariance scales: [8.0, 16.0, 16.0];
• common observation standard deviation: 0.005 m;
• covariance eigenvalue floor: 1e-12 m²;
• point-prediction changes: forbidden;
• rejected/unsupported behavior: exact registered comparator/fallback;
• protocol ID: 91c15957c8a71c579429c4245ccb2570f22ac6480dd5aebc9ddd73d1edb66b02.

The retrospective source result was overall Gaussian-NLL effect -9.136 with simultaneous 95% CI [-13.961, -4.312], exact mean identity in 22/22 cases, marginal coverage 0.706 -> 0.910, and width ratio 3.10×. It fixes claim_authorized=false, selection_authorized=false, and promotion_authorized=false.

For the future study, do not reopen the donor roster, scale grid, full-22 outcomes, or current Deform360 target. Before any fresh target outcome is opened, bind the exact merged implementation/distribution, candidate above, numerical environment, disjoint object/session roster, proper-score primary endpoint, harmful-group and width rules, exact fallback, and all source/calibration artifacts. A negative result is complete and must not be retuned on the same cohort.

[FlorianPfaff]

Fresh covariance-only execution prerequisites merged

Two exact protected-main changes now bind the requested unseen-object study:

• 287509dbebba39bc08badf3bde93e43d7f138e22 (Scope Deform360 prepared-inventory validation to its generator command #513) scopes the Deform360 source preflight to the prepared-inventory generator while preserving the already command-scoped runtime replay, frozen source cohort, physical model, candidate roster, suffix barrier, and closed v5/v6 target boundary.
• 247bc8b85ec425f4272ba867e3a7c878a7e05d56 (Test covariance-only uncertainty with exact last-residual means #505) installs the covariance-only hybrid and its sealed full-22 evidence. The only candidate permitted by this issue remains exact last_residual mean identity with independent_endpoint_v1 covariance and the frozen early/middle/late scale schedule [8.0, 16.0, 16.0].

The protected-main runner has been triggered through the reviewed active-runner path. It must first produce the complete target-blind source receipt and 100 sealed source prediction records. Until that barrier is satisfied:

• development suffix access remains forbidden;
• fresh object/session selection remains forbidden;
• v5 confirmation and v6 target payloads remain unopened;
• target outcomes cannot be used; and
• no replacement, donor change, scale retuning, endpoint change, or action-family restriction is authorized.

The fresh study remains the four-arm comparison requested here: unchanged physical fallback, deterministic last_residual, its preregistered prefix-only/exact-fallback routing, and the same exact mean with frozen source-independent covariance. Complete physical object-session units remain the only independent statistical units.

[FlorianPfaff]

Execution status: scientific path ready, GPU runner unavailable

The exact fresh-study prerequisite is now queued as protected-main workflow run 31465387504 at revision 287509dbebba39bc08badf3bde93e43d7f138e22. Both target-closed hosted contract jobs passed. The real workstation2 job is 93697602775 and remains queued with labels self-hosted, Linux, X64, nvidia-smi.

A read-only Actions diagnostic found no in-progress self-hosted job; another later GPU validation was also queued. Thus the current stop is runner availability, not a model, data, covariance, source-contract, or scoring failure.

The covariance-only candidate remains frozen exactly as registered:

• mean: exact last_residual object identity;
• donor: independent_endpoint_v1;
• early/middle/late scales: [8.0, 16.0, 16.0];
• target-side retuning, replacement, donor substitution, and action-family restriction: forbidden.

No fresh target has been selected or opened. The source job remains first in the observed GPU queue and should continue unchanged when a matching runner becomes online.

[FlorianPfaff]

Existing-data route selected; no new robot acquisition

The independent validation will use the already registered public Deform360 split rather than require new robotic acquisition.

• development/calibration: the exact ten previously opened object-session units in deform360_official_hub_visuotactile_v1_selection.json;
• target: the exact twelve separately selected confirmation object-session units in the same content-addressed selection artifact;
• statistical unit: one complete physical object-session;
• source and target rosters remain disjoint by object identity;
• target payloads and outcomes remain sealed until a frozen source decision authorizes one evaluation;
• no object, episode, camera, failed stream, or action family may be replaced after access;
• unsupported or rejected target evidence must use the preregistered exact fallback and remain in the all-target analysis.

This does not reinterpret the earlier complete-324-stream official-Hub result, which remains a valid support-negative result for that provider version. The currently registered v6 route is a separate, already-frozen source-prediction design with metadata-only camera selection, reserved endpoint views, and a complete prediction barrier.

Its first protected-main attempt, run 31465387504, stopped before any source suffix or confirmation access because the four pinned dependency repositories appeared as untracked paths in the parent checkout. Draft PR #536 repairs only that execution-layout defect while retaining every source-file digest, cohort, selector, endpoint, fallback, and information boundary.

No new measurement is required for this route. A source-negative outcome remains complete and does not authorize opening the twelve target object-sessions.

[FlorianPfaff]

Paper-grade independent-validation path advanced

The exact target-closed runtime repair has now merged on protected main as PR #540, commit b0f6b46991a20c54260baf58ddf62fbb6dab7813.

The repair changes only the isolated GPU bootstrap and failure receipt:

• pins and verifies virtualenv==21.7.4 in the inherited CUDA runtime;
• gates the source-prediction science step on complete bootstrap success;
• emits a bounded zero-evidence, target-closed receipt if bootstrap fails before the ordinary receipt exists; and
• preserves every frozen source, model, estimator, covariance, cohort, endpoint, fallback, and target-access rule.

All exact-head checks passed before merge, including the dedicated Deform360 source-prediction contracts, changed-source preflight, security scanning, stable-core branch coverage, full suites on Python 3.10/3.12/3.14, distribution builds, and isolated wheel/sdist smoke tests.

The protected-main source-prediction path is therefore unblocked. Its required next output remains exactly 100 prefix-only source prediction seals over the ten development object-sessions. Development suffixes may be attached only after that complete prediction barrier exists. The twelve disjoint target object-sessions remain unopened.

The paper-grade hypothesis remains unchanged and frozen:

• predictive mean: exact last_residual object identity;
• covariance donor: independent_endpoint_v1;
• early/middle/late scales: [8.0, 16.0, 16.0];
• observation standard deviation: 0.005 m;
• primary endpoint: equal-object Gaussian NLL difference versus deterministic last_residual;
• mandatory companion reporting: coverage, full width, exact point parity, per-stratum effects, worst-object harm, and exact fallback;
• no target-side tuning, replacement, donor substitution, action-family restriction, or denominator change.

A source-negative result remains complete and keeps the target closed. A source-positive result may authorize one preregistered evaluation on all twelve target object-sessions; no result from the target panel may be used to revise this specification.

[FlorianPfaff]

PR #524 now explicitly records this issue as a blocking cohort/information-order conflict. Its source-only residual-history adapter has been corrected to preserve each material's exact last valid causal residual, and its target-lock/acquisition tooling remains draft. No target decode, prediction, scoring, or claim promotion is authorized until #461's independent unseen-object/session design is reconciled with that separate 24-object protocol and one immutable cohort/split definition is selected before outcome access.