High-risk tools in Visionmcp

Queue the exact training run created by an artifact-bound active-learning plan.

Create a correction dataset and offline training run against a fixed benchmark.

Generate, review, and require a verified accepted L3 synthetic calibration project.

Bootstrap an unaccepted, measured DGX Spark reconstruction candidate.

Bootstrap an unaccepted, measured RTX 5090 FE reconstruction candidate.

Queue a recessed DGX foot refinement without changing the 150 mm body envelope.

Queue a material and rear-I/O DGX candidate while retaining exact body dimensions.

Queue an evidence-bound RTX front X-frame refinement at strict dimensions.

Queue a seven-blade RTX visual candidate while retaining strict dimensions.

Queue a non-destructive, audited Blender LOD checkpoint for named mesh objects.

Queue bounded retopo, UV, PBR, bake, rig, animation, LOD, collision, and repair.

Queue safe headless renders for the current evidence-bound camera solution.

Derive source-linked pinhole frames while preserving pose and scale authority.

Advance exactly one validated OBSERVE-to-ROLLBACK controller state.

Resume a paused campaign at its preserved controller state.

Start a persistent evidence-governed reconstruction controller.

Evaluate every mandatory gate atomically and auto-reject any regression.

Queue allowlisted Blender generation into a new audited checkpoint.

Execute a bounded governed search for gaps, or issue precise capture requests.

Queue safe headless Blender rendering for a planned synthetic dataset.

Plan and queue offline feature-model training; no weights are downloaded.

Execute reviewed search queries and register results with unresolved source rights.

Queue independent geometry hypotheses followed by license-aware consensus.

Queue an evidence-bound photogrammetry lane without promoting unresolved scale.

Execute cheap local portfolio lanes and record unavailable workers truthfully.

Build a fresh editable dimension-bound seed with no private starting model.

Queue the evidence-through-render bootstrap without claiming final fidelity.

Queue or resume isolated render-and-compare evaluation for a planned search.

Generate replayable automatic mask proposals with no review authority.

Queue an approved repair, checkpoint audit, topology/ray validation, and rear render.

Advance verified bytes through license, rights, and static scan.

Build approximation-only structure, UI-state, and behaviour graphs.

Run the owned application benchmark; external gates report blocked rather than pass.

Run one bounded headless Ghidra analysis in an isolated worker process.

Build an editable unbranded industrial-design hierarchy with geometric residual authority.

Run requested reconstruction backends and seal a ReconstructionPortfolio.

Compile observed layout/text/paint evidence into deterministic HTML/CSS.

Export browser motion table from a CameraPathGraph; optional Blender bake is honest.

Compile MotionGraph into reusable sticky, scroll, reveal, media, camera, and responsive primitives.

Run a capsule evaluation or mandatory global frontend non-regression gate.

Fit a RANSAC plane/box/cylinder/sphere; result authority is MODEL_DERIVED.

Groom fur in Blender; returns measured groom report or an explicit blocked state.

Generate measured LODs; report blender_used=false when Blender is blocked.

Generate tileable PBR maps from a MaterialHypothesis (not observed textures).

Generate UVs in Blender; return measured quality or a blocked state.

Generate after verified search exhaustion or an explicit original declaration.

Run governed pixel, perceptual, layout, typography, crop, stacking, and distribution metrics.

Retopologize via real Blender; blocked state is reported if Blender cannot run.

Queue normalized geometry evidence generation without promoting unresolved scale.

Run specialist critics and seal a PerceptualCritique with measured findings.

Plan and commit a saccade to a new region (IOR-aware).

Recover metric OpenCV cameras from an authoritative measured chessboard.

Queue camera solving while preserving backend authority and uncertainty labels.

Solve a LightingHypothesisSet capped by derive() over inputs.

Solve an evidence-bound product camera from six or more 3D/2D landmarks.

Associate detections to tracks for one frame (IoU + appearance + Kalman).

Compile and verify a clean-room behavior capsule without copying source assets.

Plan and queue camera-bound offline appearance-oracle training.

Report a structured failure and retry up to the job's configured attempt limit.

Run the first complete audit slice asynchronously from evidence through receipt.

Execute the next safe evidence-derived action or stop at an authority boundary.

Queue a robust fit proposal; named review is required before parameters change.

Launch an explicitly L0 generated-design portfolio without measured-target claims.

Launch the computer-hardware ontology and reconstruction portfolio.

Launch anatomical landmarks and distinguish real reconstruction from generation.

Launch packaging folds, layers, print regions, and reconstruction hypotheses.

Queue the complete staged reconstruction/audit workflow and retain all blockers.

Launch the vehicle ontology, constraints, search plan, and candidate portfolio.