Promote query-aware observability from analytic control to held-out evidence

[FlorianPfaff]

Scientific objective

Build the evidence required for a Prob4D-centered paper claim:

Partial gauge information is admitted only when it supports a declared physical query; complete-prior uncertainty is retained in unobserved directions; unsupported precision is rejected; otherwise the exact physical fallback is preserved.

The implementation is on science/query-aware-observability-v1. This issue tracks promotion beyond controlled mechanism evidence.

Stage 1 — implementation and analytic control

• Query-aware implementation PR is green under focused and repository CI.
• Analytic collinear control artifact is frozen with source revision.
• On-support query is factor-supported.
• Broad-prior off-axis query is fallback-unresolved.
• Tight-prior off-axis query is prior-bounded.
• Bogus full-rank completion is fallback-invalid-update despite sub-tolerance reported uncertainty.

Stage 2 — DLO-like controlled sweep

Freeze before running:

• effective rank and geometry-thickness grid;
• overlap/noise/prior-strength grid;
• on-support, near-support, and off-support query families;
• component-wise query tolerances;
• rank, uncertainty, gain, information-mismatch, and unsupported-information thresholds;
• sequence/trial seeds and artifact schema.

Required baselines:

• reject all rank-deficient factors;
• point transform with no covariance;
• pseudoinverse covariance without query admission;
• ridge/full-rank completion;
• exact observable-subspace fusion without query admission;
• query admission without provenance audit;
• full proposed method.

Required outputs:

• query RMSE/NLL/coverage/calibration;
• false-admission and false-rejection rates;
• fallback and harmful-update rates;
• risk/coverage curves;
• results conditional on rank and query-nullspace sensitivity;
• runtime and route accounting.

Stage 3 — fresh-provider qualification

Coordinate with #333 and #49.

• Reconcile the overlapping PointWorld persistent-point draft PRs into one minimal source contract; do not merge duplicates.
• Freeze provider checkpoint/runtime/dataset hashes.
• Pass camera synchronization and calibration gates.
• Pass metric-coordinate and persistent point-identity gates.
• Pass source-only overlap/support feasibility gates.
• Preserve every terminal failure reason.

No target access is authorized merely because the adapter executes.

Stage 4 — held-out registration

Before opening target outputs:

• freeze physical query definitions and tolerances;
• freeze query-admission thresholds and whether prior-bounded is permitted;
• freeze exact fallback and BayesianPhysTwin finite-group harm guard;
• freeze grouping/bootstrap unit and primary endpoints;
• freeze stopping and route-accounting rules;
• commit registration in Prob4D and FlorianPfaff/BayesianPhysTwin-Paper.

Stage 5 — end-to-end evidence

Evaluate on identical held-out factors/queries:

• exact physical fallback;
• unguarded provider candidate;
• exact observable-subspace fusion;
• query-aware admission;
• query-aware admission plus BayesianPhysTwin harm guard.

Primary evidence bar:

• lower harmful-update rate than ridge completion and unguarded exact-factor fusion;
• materially higher supported-query coverage than rejecting all rank-deficient factors;
• calibrated complete posterior;
• nontrivial fresh-provider factor yield;
• sequence-level uncertainty intervals;
• complete reporting of positive and terminal negative routes.

Claim boundary

Until Stages 3–5 pass, this work supports only a controlled method claim. It must not be described as fresh real-provider competence, a new benchmark result, or established downstream BayesianPhysTwin improvement.

[FlorianPfaff]

Related but orthogonal controlled-method lane is now under review in #349, with paper-side evidence in FlorianPfaff/BayesianPhysTwin-Paper#124.

#349 addresses nonlinear tangent-null query uncertainty: even after preserving the complete first-order joint gauge covariance, a physical query can have zero first-order sensitivity while having positive leading curvature variance; an axis spherical-radial reference can miss the same mixed term. The PR adds shared quadratic moment factors and controlled comparisons.

This does not advance this issue's Stages 3–5, establish fresh-provider competence, or authorize any target access. Query-aware observability and curvature are complementary: observability asks whether available information supports the declared query; curvature asks whether the uncertainty representation itself contains the leading nonlinear query uncertainty. A future end-to-end study should combine them only under a separately frozen protocol.

[FlorianPfaff]

DOT rope source real-geometry feasibility update

The public DOT rope installation was sufficient for a source-only controlled-gauge pilot. Actions run 33302466278 opened only R01-R03 publisher 3-D coordinate members, retained R04-R70 as reserved, and produced artifact bf6c1439a3f61416c47c01dc441f0ae9215b5ca2eb8c8ba07aff3b17ed4df885.

Across 114 frozen near-linear/off-axis real-geometry cases, local first-order query uncertainty passed the scalar screen while the finite posterior screen failed in 114/114. This is direct real-geometry support for a query-level nonlinear observability failure. It is not Stage 3 provider qualification because the angular posterior was controlled and no learned visual provider ran.

The result also localizes the next method requirement: under the broad 1 rad source prior, finite-orbit/circular moments outperform local second order. A next provider-qualified protocol should therefore compare both the small-uncertainty shared-curvature closure and an exact/controlled finite-orbit arm, without allowing either to alter provider means or target selection.

Chronological data boundary: R01-R03 opened for source development; R04-R70 remain reserved. The subsequently created inventory draft #352 was closed rather than redefining that split.

[FlorianPfaff]

DOT marker-free provider experiment dispatched but awaiting gpuserver6000

PR #356 is merged and the immutable source request is committed. Actions run 33305718723 passed its merged-main authorization gate. The native CUT3R/RoPE job 99241835661 is still queued for [self-hosted, Linux, X64, gpuserver6000] and has not received a runner.

This is an operational dispatch block. Four older Prob4D runs targeting the same runner have also remained queued, whereas hosted validation is functioning. No self-hosted step has started, so the current experiment has read no normal-view images or marker coordinates; R04-R70 remain unopened. The implementation, protocol, staged provider/evaluation custody, and all hosted/full-suite gates are complete. Bringing the matching runner online is the remaining prerequisite; the existing immutable request should remain the execution record.

[github-actions]

DOT marker-free CUT3R source experiment

• request ID: 353d22d234c57fd7c35803dcdcc7e9035ba71235721c0667bd186f0ba7e41872
• native runtime: skipped / unavailable
• sealed provider: skipped / unavailable
• source evaluation: skipped / unavailable
• lowest registered source NLL: unavailable
• workflow run: https://github.com/IPS-Stuttgart/Prob4D/actions/runs/33305718723

The provider stage opened only R01-R03 normal-view images. Marker payloads were available only to the later job after the provider bundle was sealed. R04-R70 remained unopened. No BayesianPhysTwin or Causal4D outcome was run.

[github-actions]

DOT marker-free CUT3R source experiment on gpuserver4090

• request ID: 93fa0abd39bb275bc6c31502d26fb348e20bf2c6494de16046c0baa7b3fb48b2
• native runtime ID: unavailable
• sealed provider: technical-failure / unavailable
• source evaluation: skipped / unavailable
• lowest registered source NLL: unavailable
• workflow run: https://github.com/IPS-Stuttgart/Prob4D/actions/runs/33318657383

The provider stage could open only R01-R03 normal-view images from the official DOT V29 archive. Marker payloads were available only to the later job after the provider bundle was sealed. R04-R70 remained unopened. No BayesianPhysTwin or Causal4D outcome was run.

[github-actions]

DOT marker-free CUT3R source experiment on gpuserver4090

• request ID: a5c6c7432dfae917e283134aca48e7d4a3a4c242b7488d66286af11cdebd924b
• native runtime ID: unavailable
• sealed provider: technical-failure / unavailable
• source evaluation: skipped / unavailable
• lowest registered source NLL: unavailable
• workflow run: https://github.com/IPS-Stuttgart/Prob4D/actions/runs/33322175895

The provider stage could open only R01-R03 normal-view images from the official DOT V29 archive. Marker payloads were available only to the later job after the provider bundle was sealed. R04-R70 remained unopened. No BayesianPhysTwin or Causal4D outcome was run.

[github-actions]

DOT marker-free CUT3R source experiment on gpuserver4090

• request ID: 744a697c0f4f19ee29450130a82443417d544a9b46ced8c99ab27db81ce6ac85
• native runtime ID: unavailable
• sealed provider: technical-failure / unavailable
• source evaluation: skipped / unavailable
• lowest registered source NLL: unavailable
• workflow run: https://github.com/IPS-Stuttgart/Prob4D/actions/runs/33322457042

The provider stage could open only R01-R03 normal-view images from the official DOT V29 archive. Marker payloads were available only to the later job after the provider bundle was sealed. R04-R70 remained unopened. No BayesianPhysTwin or Causal4D outcome was run.

[github-actions]

DOT marker-free CUT3R source experiment on gpuserver4090

• request ID: 938ce6ccf74b63bb0ed75b58941b9d90a247a8d3a189024d29f85e3b2055c29e
• native runtime ID: unavailable
• sealed provider: technical-failure / unavailable
• source evaluation: skipped / unavailable
• lowest registered source NLL: unavailable
• workflow run: https://github.com/IPS-Stuttgart/Prob4D/actions/runs/33322810529

The provider stage could open only R01-R03 normal-view images from the official DOT V29 archive. Marker payloads were available only to the later job after the provider bundle was sealed. R04-R70 remained unopened. No BayesianPhysTwin or Causal4D outcome was run.