Run a support-feasible held-out real provider and BayesianPhysTwin gate

[FlorianPfaff]

Objective

Test one genuinely independent real provider on complete physical object/session groups, separating provider competence from downstream guarded BayesianPhysTwin value. Controlled synthetic graph and Prob4D-to-BayesianPhysTwin results are mechanism evidence only; they do not establish real-provider competence or unseen-object benefit.

Stage 0: support feasibility before residual outcomes

Before opening source residuals or any target outcome, publish a content-addressed ProviderSupportFeasibilityV1 decision for the exact provider/model/runtime identity. It must bind:

• complete development, calibration, target, and excluded object/session rosters;
• representation, units, metric-frame prior, causal cutoff, windows, seeds, and immutable model set;
• required streams, identities, masks, timestamps, camera geometry, and physical-query mapping;
• minimum complete-stream support and technical-failure policy; and
• exact unsupported-case fallback.

If complete-stream support fails, stop that provider version and report the negative feasibility result. Do not fit covariance, inspect target outcomes, substitute partial streams, or rescue it with a different graph/anchor/model under the same identity.

Frozen comparison arms

Select at most one Prob4D candidate using development/calibration units only. Score exactly these main arms on the same held-out units:

1. unchanged physical fallback;
2. simple direct visual observation or last_residual comparator;
3. one source-selected Prob4D candidate with explicit complete joint gauge uncertainty;
4. exact physical fallback for every unsupported, invalid, or rejected candidate.

Metric-assisted variants must be labelled sensor-assisted and may not be compared as though they used the same information as an unanchored arm. Additional graph, association, semantic, or covariance variants are out of scope unless this frozen result localizes a specific failure that requires a new protocol.

Freeze before target access

Bind exact Prob4D, BayesianPhysTwin, observation-provider, model, environment, and numerical identities; source/calibration-only reliability and covariance calibration; the BayesianPhysTwin guard; horizons; physical queries; group weighting; bootstrap/randomization seeds; technical failures; and exact fallback. Frames, points, tracks, views, and taxels are nested observations, not independent groups.

Separate endpoints

Provider competence

Report by independent object/session and horizon:

• support and technical-failure accounting;
• point/endpoint or valid identity-based track error;
• seam/drift diagnostics where applicable;
• 50/90/95% coverage, proper score, normalized NEES, and full covariance width;
• identity retention/precision and selective risk; and
• worst-group coverage shortfall.

Downstream guarded query

Report separately:

• deployed physical-query proper score and RMSE;
• object/session-clustered paired interval versus both comparators;
• accepted, rejected, and exact-fallback counts;
• harmful accepted updates and worst-group regret;
• accepted-update coverage and width; and
• provider competence versus downstream benefit without allowing either to rescue the other.

Completion rule

Close with one complete frozen outcome:

1. provider competence and guarded BayesianPhysTwin criteria both pass on the independent cohort; or
2. a valid negative/bounded result localizes failure to support, means/identities, gauge/dependence, calibration shift, object/session transfer, or query identifiability.

A tie retains the simple comparator and physical fallback. Do not retune on the opened target cohort. A downstream result cannot rescue failed support or provider competence, and optional Causal4D results cannot rescue either upstream claim.

[FlorianPfaff]

Mechanistic evidence update from PR #75

A calibration-separated pilot and an independently seeded replication have now been completed for the tree, sequential marginal CI, full-joint graph, and source-cycle-guarded graph.

Replication design:

• 48 clean calibration trials;
• 128 target trials in each of six regimes, 768 target trials total;
• separate calibration and target seed ranges;
• frozen 0.95 higher-quantile source-cycle threshold;
• precise-looking biased skip edges whose reported covariance and overlap residual were deliberately left unchanged;
• 2,000 paired trial-bootstrap resamples.

Clean graph result

The full-joint graph showed a small clean endpoint benefit in correlated regimes:

• correlation 0.75: delta versus tree −0.000320, paired 95% interval [−0.000601, −0.000037], approximately 3.1%;
• correlation 0.95: −0.000192, interval [−0.000323, −0.000057], approximately 4.1%;
• independent clean: −0.000438, interval [−0.001066, +0.000214], not decisive.

Coverage remained conservative in the clean scenarios. Sequential marginal CI showed no clean advantage.

Biased-edge failure and guard result

The unguarded graph was severely harmed by one precise-looking biased skip edge. In the strong regimes, aggregate full-graph endpoint error was about 4.1× and 7.8× the tree, with normalized NEES 69 and 284 and coverage below 0.83.

The source-only guard detected all 99/99 injected inconsistencies and reproduced the exact production tree on every injected trial. However, the current global raw-displacement threshold caused 41/128 false fallbacks in independent-clean trials while causing 0/256 in correlated/highly-correlated clean trials.

Provisional decision

• keep the joint spanning tree as production default;
• retain full-joint graph and exact guarded fallback as experimental;
• do not promote sequential marginal CI;
• next evaluate source-side uncertainty-normalized cycle consistency in issue Normalize the causal cycle guard by source-side uncertainty scale #77;
• keep this issue open because its held-out physical-object/session and separately sealed BayesianPhysTwin criteria have not yet been executed.

Durable evidence is recorded in docs/gauge-graph-science-results-v1.md and evidence/gauge-graph-science-replication-v1/summary.json on PR #75. Replication report ID: a4c9a5294dbb0f6c12040494900c14f39f4c8e42a7499fc2fc5db12f4d789460; artifact digest: sha256:890664f63ca872b156fc44b9687924cf577c28d7ecc03064aa2df1316fd903a3.

[FlorianPfaff]

Synthetic admission gate passed; downstream gates now active

Merged PR #87 resolves the source-cycle admission bottleneck with a simpler strict finite-sample rule rather than the abandoned subgroup implementation in #86.

Fresh registered result:

• 96 clean calibration trials and 768 disjoint target trials;
• split-conformal upper threshold 0.8220838356516649, rank 93/96;
• finite-sample marginal miscoverage bound 4/97 = 0.041237113402061855;
• 104/104 injected biased edges detected;
• worst registered clean false fallback 7.03%;
• exact complete production-tree fallback on every rejection;
• all 10/10 locked criteria passed;
• clean coverage at least 98.96% and largest covariance-trace ratio versus tree approximately 0.9534.

The production decision remains conservative:

1. keep the single-parent joint spanning tree as the claim-bearing default;
2. retain the full-joint graph plus finite-sample normalized guard as experimental;
3. do not promote sequential marginal CI;
4. do not infer physical-object or BayesianPhysTwin benefit from the synthetic pass.

The next registered gate is running in Prob4D PR #89 / BayesianPhysTwin PR #109. It compares physical fallback, naive visual state, marginalized persistent factors, explicit-gauge framewise factors, persistent explicit-gauge factors, and an independently anchored control across 48 calibration groups and 384 disjoint target groups. Workflow run 30901907830 is queued on [self-hosted, Linux, X64, nvidia-smi].

This issue should remain open until both conditions are resolved:

• the decisive controlled BayesianPhysTwin guard/benefit study has a verified pass or valid negative result; and
• a declared released representation and genuinely independent physical object/session or acquisition cohort is evaluated with object/session-level inference. Generic file discovery or an already-open Deform360 cohort is not sufficient.

[FlorianPfaff]

Status update after the completed gauge-graph studies:

• the production single-parent joint spanning tree remains unchanged;
• merged PR Add finite-sample cycle guard calibration #87 supplies the passing controlled-synthetic finite-sample conformal cycle guard, but held-out provider and BayesianPhysTwin gates are still required;
• draft PR Reproduce positive controlled Prob4D-to-BPT study #89 is the authoritative frozen Prob4D-to-BayesianPhysTwin controlled decision study;
• draft PR Add source-only cross-window tracklet association #91 adds a source-only cross-window tracklet-association diagnostic for the specific material-identity gap between windows.

PR #91 is intentionally excluded from the frozen PR #89 methods and must not be used to reinterpret that result. It should advance only if the frozen study or later independent provider evaluation shows that within-window persistent identities have useful headroom but identity resets remain limiting. Production promotion under this issue still requires independent physical-object/session provider competence and the separately sealed downstream guarded physical-prediction gate.

[FlorianPfaff]

Downstream controlled gate resolved; independent provider gate remains

The first of the two open conditions is now complete.

The authoritative Prob4D-to-BayesianPhysTwin controlled study finished successfully in merged IPS-Stuttgart/BayesianPhysTwin#109:

• exact Prob4D revision: aa8ffc6541011d044561e09870569a14ab3f586f;
• exact executed BayesianPhysTwin revision: 07e6f2fe85eb3f1dc9d454ecaf1cb2f83a447131;
• protocol SHA-256: 921da8a6f14f9430b3f4861d68326d904f61b922e3aedd2b35882ea97bc63111;
• workflow run: 30907109833;
• artifact ID: 8891529145;
• artifact digest: sha256:aa71e7dd68b96a8ee45c284441a1c12766ddcb02fe275129d9d6a42d5c7cb94e;
• fresh report ID: 0b94125189ee8c7cdccd6ca799bb576bf344c8983a420dc7b128ac9b291f0707.

All six registered downstream criteria passed. The persistent explicit-joint-gauge arm reduced deployed RMSE from 6.165625 mm to 0.534311 mm, accepted 373/384 target groups, retained zero harmful accepted updates, and reproduced the physical fallback exactly for every rejection. The rowwise-marginalized persistent arm was worse and retained eight harmful accepted updates.

This resolves the controlled BayesianPhysTwin representation/guard gate, but it does not resolve this issue's independent provider-competence condition. The experiment used a controlled synthetic generator rather than a declared released real representation on genuinely independent physical objects or acquisition sessions. The production single-parent tree therefore remains the claim-bearing default; the full-joint graph remains experimental.

Keep this issue open for the remaining gate: a frozen real provider evaluated at the physical object/session level, with graph-versus-tree seam/drift, calibration, uncertainty-width, and guarded downstream outcomes reported without reusing an opened public cohort as confirmation.

[FlorianPfaff]

Controlled BayesianPhysTwin gate passed authoritatively

The preregistered controlled Prob4D-to-BayesianPhysTwin study has now completed from reachable exact sources and merged as IPS-Stuttgart/BayesianPhysTwin#109.

Authoritative execution:

• BayesianPhysTwin revision: 07e6f2fe85eb3f1dc9d454ecaf1cb2f83a447131;
• Prob4D revision: aa8ffc6541011d044561e09870569a14ab3f586f;
• workflow run: 30907109833;
• evidence artifact ID: 8891529145;
• artifact ZIP SHA-256: aa71e7dd68b96a8ee45c284441a1c12766ddcb02fe275129d9d6a42d5c7cb94e;
• fresh report ID: 0b94125189ee8c7cdccd6ca799bb576bf344c8983a420dc7b128ac9b291f0707;
• protocol SHA-256: 921da8a6f14f9430b3f4861d68326d904f61b922e3aedd2b35882ea97bc63111.

Primary controlled result over 384 disjoint synthetic object/session groups:

• physical fallback: 6.166 mm deployed RMSE;
• explicit-gauge persistent Prob4D factors: 0.534 mm;
• improvement: 91.33%;
• accepted groups: 373/384;
• harmful accepted updates: 0;
• paired deployed-minus-physical 95% interval: [-5.880, -5.365] mm;
• accepted nominal-90% coverage: 98.68%;
• exact physical fallback on every rejection;
• all six registered criteria passed;
• method decision: advance-to-fresh-physical-object/session-gate.

Mechanistic controls support the intended interpretation within the generator: the persistent rowwise-marginalized gauge arm reached 1.231 mm and retained eight harmful accepted updates, whereas the persistent explicit-joint-gauge arm reached 0.534 mm with none. The framewise explicit-gauge arm reached 0.719 mm, and the independent metric-anchor control (0.551 mm) did not improve on the unanchored primary.

Evidence limitation: the artifact retains method-level calibration summaries and thresholds but not method-by-calibration-group candidate rows, so the threshold search cannot be independently reconstructed from the bundle alone. The exact-source execution and target decision are confirmed; a future evidence version should retain calibration_trials.csv and independently reapply every threshold candidate and tie-break. Paper PR FlorianPfaff/BayesianPhysTwin-Paper#24 records this boundary.

Issue #49 should remain open. The synthetic source-admission and controlled BPT gates now pass, but production promotion still requires a frozen real provider and genuinely fresh physical objects or acquisition sessions with prefix-only persistent identities, explicit joint gauge covariance, source-calibrated guarding, exact fallback, and object/session-clustered inference. Previously opened Deform360 or PokeFlex targets cannot serve as that confirmation.

[FlorianPfaff]

Cross-window identity evidence update

Draft PR #92 adds the next source-only tranche for the material-identity path:

• complete content-addressed CausalTrackletArtifactV1 inputs, including every array, nested metadata, prediction-manifest identity, exact source revision, and builder configuration;
• cross-window candidate covariance using the full joint Sim(3) gauge prior, including the P_lr/P_rl cross terms rather than an independence approximation;
• provenance-bound association evidence with deterministic result identity;
• fail-closed tests for noncanonical lineage, covariance indefiniteness, dependence sensitivity, and execution-chunk invariance.

This remains an additive diagnostic and does not rewrite provider-v2 point IDs. It is not part of the frozen controlled decision study and does not satisfy this issue's independent physical-object/session or downstream guarded BayesianPhysTwin gates. The issue should remain open pending those prospective evaluations.

[FlorianPfaff]

Controlled downstream gate resolved; independent physical provider gate remains

The frozen Prob4D-to-BayesianPhysTwin controlled study has now completed and passed all six registered criteria in merged BayesianPhysTwin PR #109.

Primary explicit-joint-gauge persistent-identity result:

• deployed RMSE 0.534 mm versus 6.166 mm physical fallback (91.33% improvement);
• paired 95% bootstrap interval for deployed-minus-physical RMSE [-5.880, -5.365] mm;
• 373/384 accepted groups;
• 0 harmful accepted updates;
• exact physical fallback for every rejection;
• all six scenarios improved, with reductions from 86.51% to 95.95%;
• fresh confirmation workflow run 30907109833, report ID 0b94125189ee8c7cdccd6ca799bb576bf344c8983a420dc7b128ac9b291f0707.

This resolves the separately sealed controlled downstream benefit condition. It does not resolve the provider-competence condition for this issue: the executed study uses the real Prob4D/BayesianPhysTwin types and solvers inside a controlled generator, not a fresh physical camera/provider cohort.

The names-only Deform360 preflight also established that the mounted release currently contains no eligible untouched official-release trajectory cohort for a fresh confirmation. Generated experiment paths must not be treated as independent release data. The active Deform360 work is therefore explicitly retrospective and non-fresh and cannot close this issue.

The production decision remains unchanged:

1. retain the single-parent joint spanning tree as the claim-bearing default;
2. retain the full-joint graph plus finite-sample source guard as experimental;
3. do not promote sequential marginal CI;
4. keep this issue open only for a genuinely independent physical object/session or acquisition cohort with declared representation, object/session-level inference, provider competence metrics, and the already demonstrated guarded downstream decision boundary.

[FlorianPfaff]

Implemented the prospective identity-marginalized arm in #100.

The new contract preserves window-local (window_id, track_id) alternatives, requires an exact newest-window null/fallback hypothesis, binds source calibration and association lineage, and provides exact logsumexp likelihood marginalization plus law-of-total-covariance moment matching. It therefore supports the issue's required comparison between within-window persistence, hard cross-window hypotheses, identity marginalization, oracle identity, and exact fallback without rewriting provider-v2 observation IDs.

This is still infrastructure for the frozen object/session-held-out gate; no target outcome or physical claim is opened by the PR.

[FlorianPfaff]

Implementation update for the real Prob4D → BayesianPhysTwin gate:

• Harden causal tracklet and cross-window association contracts #94 hardens within-window material tracklets: deform-supported seeds, explicit target-mask accounting, ambiguity/termination diagnostics, and a length-neutral comparison score without changing cumulative survival semantics.
• Add content-addressed joint-gauge tracklet evidence with covariance-safe ranking #92 adds covariance-safe cross-window material evidence with joint-gauge provenance, conservative ranking, and deterministic fail-closed replay.
• Add append-only material-identity hypothesis streams #95 adds an append-only material-identity hypothesis stream with immutable prefix lineage and canonical provenance validation.
• Add replayable target-blind selection evidence #96 adds evidence-v2 for target-blind candidate selection and exact fallback replay. Its focused Ruff/mypy/pytest contract and independent fail-closed quality workflow are green; the focused suite reports 14 passing regressions.

These contracts prepare the prospective object/session-held-out experiment but do not establish physical benefit by themselves. The claim still requires the frozen target analysis defined in this issue, with one common BayesianPhysTwin guard, exact fallback, object/session-level statistics, and only the accepted BayesianPhysTwin belief plus lineage passed to Causal4D. PRs #92/#94/#95 remain draft while their source-publication jobs and ordinary CI complete; #96 remains draft while the repository-wide test matrix is queued.