Measured harness ledgerPublic result
GLM 5.3 FlashHigh-voxel-density Jungle Temple diorama — GLM 5.3 Flash Max
Create a complete inspectable Blender voxel diorama of a jungle temple at a fixed 0.05-unit voxel resolution through the Blender MCP workflow.
Max reasoningHeadline result
- Workflow cost
- $0.87
- Wall-clock
- 45m 49.093s wall-clock
- Processed tokens
- 1.59M processed
- Record state
- partial_multi_user_turn_artifact_ledger
Public summary
GLM 5.3 Flash Max partial_multi_user_turn_artifact_ledger ledger: 45m 49.093s wall-clock, 1.59M processed, and $0.87 OpenCode Go API-equivalent usage accounting, not an itemized subscription marginal-cash invoice.
Run identity and stack
- Result ID: jungle-temple-glm-5.3-max
- Technical model: glm-5.3-flash
- Provider: opencode-go
- Client: OpenCode
- Stack: opencode-go
- Stack: OpenCode
- Stack: Technical model/configuration: glm-5.3-flash
- Stack: Blender MCP
- Stack: 0.05-unit voxel grid
- Stack: Inspectable .blend artifact
Primary artifact integrity
- Kind: blender-scene
- Path: artifacts/jungle-temple-glm-5.3-max/jungle_temple.blend
- SHA-256: b688378ad840df7903cd2a0c920ff56de35ce54efeb0c1b1b841f54462118fbc
Validation evidence
- Result: opened_and_source_replayed
- Path: artifacts/jungle-temple-glm-5.3-max/validation/scene-inspection.public.json
- SHA-256: 8c72ed51ef15192c35cccf9fb54436418b47f745791b15bb7788b56bd4113314
Recorded caveats
- This is an eight-user-turn recovery record, not a strict one-shot benchmark result.
- The metrics sidecar is operator-generated from the canonical bounded extractor after the same-session metrics stream stalled; the underlying raw records remain private.
- The model-exported builder followed the original scene/render; its independent replay is visually and structurally matching but has 3,840 fewer vertices and 960 fewer polygons than the delivered primary scene.
- The public archive has no render-hardware or local-FPS evidence.
- The final scene was reopened read-only, and source was independently replayed, but no blind evaluation was performed.
Visible evidence gaps
- A strict one-user-turn rerun without recovery continuations
- A builder exported as part of the original final scene/render sequence
- Render-hardware metadata and a local performance capture where relevant
- Independent blind-evaluation record
Public result only
This result keeps its public summary and evidence, but it does not currently have a matching Builder test with prompts, projects, Harness workflows or skills.