Turn object images into an evidence-bound procedural reconstruction. The output may be a feasibility verdict, a build plan, or working code; every branch follows the same six stages.

Use every supplied view. If the target or intended use is ambiguous, resolve that ambiguity before committing to a representation. Otherwise, assume an interactive browser prop whose authored scale can be revised.

A single view determines only visible evidence. Hidden surfaces, thickness, absolute dimensions, material parameters, and articulation remain assumptions unless symmetry, known construction, extra views, or measurements constrain them. Label each assumption. Exact likeness, manufacturing dimensions, or an unseen mechanism requires the missing view or measurement; a stated stylized approximation is a valid alternative.

For a local image, set to the directory containing this , then run using probe_reference_image.py. It reports format, byte count, dimensions, aspect ratio, and metadata parsing status. That status describes file metadata only; visual inspection in stage 1 alone determines whether the pictured object is readable.

Inspect every view before planning geometry. Record:

Visual inspection alone assigns an evidence readability verdict:

When conditional or unreadable evidence affects an identity-defining feature, request the smallest missing input or state the reduced claim. Inspection is complete when every supplied view is accounted for, the verdict has evidence, and every consequential uncertainty is explicit.

Freeze one object coordinate frame and one authored unit. Partition the object coarse to fine:

Mark identity anchors: silhouette breaks, negative spaces, signature proportions, distinctive joints, repeated rhythms, and local material or geometric features whose loss would change what the object is. Keep each anchor as an explicit component, boundary, field, or verification target; a later average that erases one fails the reconstruction.

Decomposition is complete when every visible identity anchor has one owner, every child part names its parent/contact, and every inferred part is labeled as an assumption rather than image evidence.

Choose a representation per semantic part, using the cheapest form that preserves its identity anchors and future ownership:

After choosing one or more of these branches and before implementation, read the matching sections of procedural object patterns and apply their construction and verification rules.

Keep separate nodes for parts that move, detach, change material independently, need picking/collision identity, or own an attachment. When a child part touches, enters, hinges from, or follows a parent, read attachment and joint correctness before blockout.

Selection is complete when every decomposed part has exactly one representation and owner, every rejected alternative has a concrete failure condition, and every conditional reference required by the selected branches has been applied.

Build only macro masses, dominant openings, coordinate frame, and identity-critical negative space. For feasibility or planning, the blockout is a dimensioned primitive/profile sketch; for implementation, it is renderable coarse geometry in the target repository.

Use the intended comparison camera before adding detail. Match framing and projection separately from object proportions so camera error does not become geometry. Preserve semantic node boundaries needed by later attachments or motion.

Blockout is complete when the target reads from silhouette alone in every required view, dominant proportions and negative spaces are within the declared tolerance, and geometry and layout carry the match without relying on material or micro detail.

Open only branches supported by the request and evidence, in this order:

After each opened branch, compare against its identity anchors before opening the next. Refinement is complete when every applicable branch meets its local criterion, every omitted branch is inapplicable rather than forgotten, and every identity anchor remains visible or structurally represented.

For code, run the target repository's syntax/type/build checks and its existing capture path. Inspect the final render directly. Keep comparison camera, projection, framing, viewport, exposure, and review lighting fixed across iterations.

Compare in this order:

Route each mismatch to one cause:

Verification is complete when the artifact passes the checks appropriate to its branch, every required view has been inspected in the stated order, every remaining mismatch has a routed correction or explicit limitation, and the final report distinguishes observed fidelity from remaining assumptions.

Return the evidence-readability verdict, evidence and assumptions, decomposition and identity anchors, representation decisions, branch results, verification evidence, and remaining limits. For implementation requests, edit the target code and report only checks and images actually inspected.