render-export-evidence

Purpose

Classify preview, render, .blend, and GLB/export evidence into a conservative, one-row-per-artifact-type evidence ledger.

• Runtime truth states: Not Run | Attempted | Produced | Verified | Failed
• Artifact truth states: Not Run | Not Produced | Produced | Verified | Failed (default Not Produced; never use Attempted for an artifact)
• Direction: Downgrade-only, never upgrade.

Role boundary: render-export-evidence classifies artifact evidence into a ledger; pre-handoff-verification gates handoffs (confirm-or-downgrade); preview-report renders the report Markdown; glb-web-handoff owns web handoff. This skill does not run Blender or create artifacts.

Quick start

1. Identify the runtime state (Not Run | Attempted | Produced | Verified | Failed) and list one ledger row per claimed artifact type.
2. Collect command/tool (plus exact command and exit code for CLI), input/script/scene, output path, current-run provenance, file existence, validation notes, and limitations.
3. Assign the most conservative valid artifact status (Not Run | Not Produced | Produced | Verified | Failed, defaulting to Not Produced, downgrade-only) and label any Path 2 evidence as non-release.
4. Pass the classified ledger to pre-handoff-verification to gate handoffs, preview-report to render report Markdown, glb-web-handoff for web integration, or plain-language-response for plain-language user responses.

When to use

Use this skill when:

• a user asks whether a preview, render, .blend, or GLB/export was actually produced or verified
• an agent needs to classify raw runtime outputs into a structured, one-row-per-artifact-type evidence ledger
• a runtime attempt (runtime-attempt-report) recorded commands, tools, or output files that need artifact status labels
• pre-handoff-verification or preview-report requires classified artifact truth labels as input

When not to use

• the user only needs a scene plan with no runtime or artifact evidence → use ../runtime-readiness/SKILL.md (Delegates out-of-scope requests to runtime-readiness)
• runtime readiness or path selection has not been completed and the user wants setup or routing; use runtime-readiness or runtime-path-picker → use ../runtime-readiness/SKILL.md (Delegates out-of-scope requests to runtime-readiness)
• the task is gating a handoff via confirm-or-downgrade rules; use pre-handoff-verification → use ../pre-handoff-verification/SKILL.md (Delegates out-of-scope requests to pre-handoff-verification)
• the task is rendering the formatted preview/eval report Markdown; use preview-report → use ../preview-report/SKILL.md (Delegates out-of-scope requests to preview-report)
• the task is GLB web integration after evidence is already classified and gated; use glb-web-handoff → use ../glb-web-handoff/SKILL.md (Delegates out-of-scope requests to glb-web-handoff)

Trigger phrases

• "Did the render produce?"
• "Can I claim the GLB exists?"
• "Verify export evidence."
• "What artifact status should this have?"
• "Summarize preview/render/GLB evidence."

Prerequisites / readiness

• Each claimed artifact type is identified (preview, render, blend, or GLB/export) with one ledger row per artifact type.
• Runtime state is explicitly known (Not Run | Attempted | Produced | Verified | Failed).
• Evidence sources are available or explicitly marked missing.
• No artifact claim is accepted from intent, prompt text, transcript, or executed Python code alone.
• Any pre-existing file has verifiable provenance linking it to the current run; otherwise it cannot support Produced or Verified.

Input schema

Required inputs

Optional inputs

Assumptions to confirm

• Runtime status (Not Run | Attempted | Produced | Verified | Failed) and artifact status (Not Run | Not Produced | Produced | Verified | Failed) are tracked separately.
• Every artifact defaults to Not Produced unless concrete output evidence and current-run provenance exist.
• Status transitions are strictly downgrade-only, never upgrade.
• Produced is never upgraded to Verified without passing validation notes.
• Path 2 evidence is always labeled non-release and never used as release evidence.

Output schema

Primary output

A one-row-per-artifact-type evidence ledger separating runtime truth from artifact truth.

Secondary output

• runtime status and stack identity
• per-artifact conservative status (Not Run | Not Produced | Produced | Verified | Failed)
• missing evidence and provenance gaps
• next safe skill handoff (pre-handoff-verification, preview-report, glb-web-handoff, or plain-language-response)

Evidence / caveat output

Runtime status: Not Run | Attempted | Produced | Verified | Failed
Runtime path: <Path 1 host a | Path 1 host b | Path 2 (non-release) | CLI fallback (appendix) | text-only | unknown>
Release eligibility: <release-eligible | non-release | N/A>
Blender version: <output of `blender --version` or unknown>
Blender-side add-on identity: <Lab MCP add-on version | mcp-for-blender addon.py commit | N/A for CLI>
MCP server source: <Lab .mcpb bundle filename or commit | uvx mcp-for-blender version | N/A for CLI>
MCP host product: <Claude Desktop (Connector or manual MCP) | Cursor | Codex | OpenCode | Cline | other | N/A for CLI>

Artifact ledger (one row per artifact type):
| Artifact type | Artifact status (Not Run | Not Produced | Produced | Verified | Failed) | Command / tool (+ CLI exit code) | Output path + current-run provenance | Validation notes | Limitations |
|---|---|---|---|---|---|
| <preview | render | blend | GLB/export> | <status> | <exact command + exit code, or MCP tool> | <path + provenance> | <notes or missing> | <caveats> |

Next safe action: <pre-handoff-verification | preview-report | glb-web-handoff | plain-language-response>

Required laws

• ../../laws/official-runtime-only.md
• ../../laws/evidence-before-done.md
• ../../laws/non-blender-user-language.md
• ../../laws/no-arbitrary-python-interface.md

Official runtime boundary

This skill classifies evidence from the 2-path + CLI appendix runtime model (see ../../docs/runtime-stack-strategy.md):

1. Path 1 — Official Blender Lab MCP in Blender 5.1+ (add-on manifest literal blender_version_min = "5.1.0"), hosted from either (a) Anthropic Connector in Claude Desktop or (b) any manually configured MCP client. The Connector is NOT standalone (Anthropic tutorial step 2 installs the Lab add-on inside Blender; the tutorial prerequisite says "Blender 4.2 or later" but the binding floor is 5.1+). One MCP bridge runs per Blender instance (default port 9876). The 20-tool surface includes get_blendfile_summary_path_info, get_blendfile_summary_datablocks, get_objects_summary, *_for_cli variants, and execute_blender_code.
2. Path 2 — Community mcp-for-blender (renamed from ahujasid/blender-mcp), third-party, Blender 3.0+, nine tools including get_scene_info and execute_blender_code, one bridge per Blender instance (default port 9876). Allowed only with caveats and never release evidence — every Path 2 ledger row must be labeled non-release.
3. CLI fallback (appendix) — direct blender --background --python, appendix only, never a peer path without an eval record. Documented flags: --background, --python, -E/--engine (run blender -E help for exact engine tokens; Eevee token is BLENDER_EEVEE on 4.0–4.1 and 5.x, and BLENDER_EEVEE_NEXT on 4.2–4.x). CLI evidence requires the exact command and exit code.

This skill must not run Blender, render, export GLB, mutate scenes, install runtimes, gate handoffs (pre-handoff-verification), or promote Path 2 evidence as release evidence.

Operating procedure

1. Record the runtime status (Not Run | Attempted | Produced | Verified | Failed) and runtime path/host identity.
2. Create one ledger row per artifact type in scope (preview, render, blend, GLB/export).
3. For each artifact row, collect:
   • exact command + exit code (required for CLI) or MCP tool name
   • input/script/scene
   • output path
   • file existence and provenance linking the file to the current run
   • validation notes
   • limitations, failures, and non-release label if Path 2
4. Assign conservative artifact status (Not Run | Not Produced | Produced | Verified | Failed, default Not Produced) using the decision tree (downgrade-only, never upgrade).
5. Hand the ledger to pre-handoff-verification before any user-facing handoff or to preview-report for Markdown report rendering.

Decision tree

No runtime action recorded?
  → Runtime: Not Run | Artifact: Not Run (or default Not Produced)
Runtime action recorded, but command failed or required artifact is absent/corrupt?
  → Runtime: Attempted or Failed | Artifact: Failed or Not Produced (never Attempted)
Transcript/code exists, or file is pre-existing without current-run provenance, or CLI lacks exact command + exit code?
  → Runtime: Attempted | Artifact: Not Produced (downgrade-only)
Output file/visible evidence exists with current-run provenance, but validation notes are missing or incomplete?
  → Artifact: Produced (label non-release if Path 2)
Output file exists with current-run provenance AND validation notes pass?
  → Artifact: Verified (label non-release if Path 2; never release evidence)

Playbooks

Playbook A: Text-only planning output

• Runtime status: Not Run.
• Artifact status: Not Produced (default).
• Next action: runtime-readiness or runtime-path-picker.

Playbook B: CLI fallback (appendix) output

Require:

• eval record context (CLI is appendix-only, never a peer path without an eval record)
• exact blender --background --python command (and -E/--engine token if specified: BLENDER_EEVEE on 4.0–4.1 and 5.x, BLENDER_EEVEE_NEXT on 4.2–4.x)
• CLI exit code (mandatory) and logs
• script/input path and output path
• generated file existence with current-run provenance
• validation notes

Without the exact command, exit code, and generated file provenance, keep artifact status at Not Produced or Failed.

Playbook C: Path 1 (Official Blender Lab MCP — host a Connector or host b manual MCP) output

Require:

• Blender 5.1+ (blender_version_min = "5.1.0") and non-standalone Lab add-on + server identity
• host identity: (a) Anthropic Connector in Claude Desktop or (b) manually configured MCP client
• recorded Lab MCP tool call (from the 20-tool surface, e.g., get_blendfile_summary_path_info, get_blendfile_summary_datablocks, get_objects_summary, *_for_cli variants, or execute_blender_code)
• generated file paths or visible preview/render evidence with current-run provenance
• validation notes and session/port caveats (single bridge per Blender instance, default port 9876)

Read-only summary tool calls or execute_blender_code transcripts alone are not artifact evidence; keep artifacts Not Produced until output files are verified.

Playbook D: Path 2 (community mcp-for-blender, formerly ahujasid/blender-mcp) output

• Label every ledger row as Path 2 (non-release) — Path 2 is third-party (Blender 3.0+, 9-tool surface including get_scene_info and execute_blender_code) and is never release evidence.
• Record addon.py commit/version, uvx mcp-for-blender version, single-bridge port (default 9876), and execute_blender_code unsandboxed execution caveats.
• Require the same file existence, current-run provenance, and validation notes before marking Produced (non-release) or Verified (non-release).
• See ../../docs/unofficial-runtime-bridges.md for full Path 2 caveats.

Mode handling

Text-only mode

Use to mark artifacts Not Run / Not Produced and prepare evidence checklist.

Runtime-ready mode

Use after the runtime/eval action has happened or when reviewing a captured eval record.

Blocked runtime mode

Use when runtime was unavailable; record blockers and keep artifacts Not Produced.

Validation checklist

• One ledger row is recorded per artifact type (preview, render, blend, GLB/export).
• Runtime status (Not Run | Attempted | Produced | Verified | Failed) is separated from artifact status (Not Run | Not Produced | Produced | Verified | Failed).
• Artifact status defaults to Not Produced, never uses Attempted, and only moves via downgrade-only rules.
• Runtime path/host is named (Path 1 host a/b in Blender 5.1+, Path 2 non-release, CLI appendix, or text-only).
• CLI evidence includes both exact command and exit code.
• Path 2 evidence is explicitly labeled non-release and excluded from release evidence.
• Transcript or execute_blender_code code alone is never accepted as artifact evidence.
• Pre-existing files have provenance to the current run before Produced or Verified.
• Validation notes exist before any artifact is marked Verified.

Pass / Warn / Fail rubric

Failure handling

• Missing output path or missing file/visible evidence → runtime may be Attempted, but artifact status cannot exceed Not Produced (or Failed if the command errored).
• Transcript or executed code only → artifact status remains Not Produced.
• Pre-existing file without current-run provenance → downgrade artifact status to Not Produced.
• CLI run missing exact command or exit code → downgrade artifact status to Not Produced (or Failed).
• Missing validation notes → artifact status cannot exceed Produced (downgrade any Verified claim to Produced).
• Path 2 evidence presented for release → label non-release and block release promotion.

Troubleshooting

Best practices

• Maintain strictly one ledger row per artifact type (preview, render, blend, GLB/export).
• Keep runtime states (Not Run | Attempted | Produced | Verified | Failed) distinct from artifact states (Not Run | Not Produced | Produced | Verified | Failed, default Not Produced).
• Apply downgrade-only status changes; never upgrade an upstream status claim.
• Require exact command + exit code for CLI fallback and label all Path 2 rows non-release.
• Route classified ledgers to pre-handoff-verification to gate handoffs and preview-report to format Markdown reports before calling glb-web-handoff or plain-language-response.

Good examples

• “Render status: Produced. Output path: out/hero.png. Validation: camera framing checked; material fidelity not yet verified.”
• “GLB status: Not Produced. Runtime action was Not Run; no export path exists.”

Bad examples

• “GLB produced” with no file path.
• “Render verified” with no validation notes.
• “Runtime eval passed” from read-only smoke evidence only.

User-facing response template

Artifact evidence ledger
- Runtime stack/mode: <Path 1 host a | Path 1 host b | Path 2 (non-release) | CLI fallback (appendix) | text-only>
- Runtime status: <Not Run | Attempted | Produced | Verified | Failed>
- Release eligibility: <release-eligible | non-release | N/A>

| Artifact type | Artifact status | Command / tool (+ CLI exit code) | Output path + current-run provenance | Validation notes | Limitations |
|---|---|---|---|---|---|
| preview | <Not Run | Not Produced | Produced | Verified | Failed> | <cmd/exit or tool> | <path + provenance> | <notes> | <caveats> |
| render | <Not Run | Not Produced | Produced | Verified | Failed> | <cmd/exit or tool> | <path + provenance> | <notes> | <caveats> |
| blend | <Not Run | Not Produced | Produced | Verified | Failed> | <cmd/exit or tool> | <path + provenance> | <notes> | <caveats> |
| GLB/export | <Not Run | Not Produced | Produced | Verified | Failed> | <cmd/exit or tool> | <path + provenance> | <notes> | <caveats> |

- Next safe action: <pre-handoff-verification | preview-report | glb-web-handoff | plain-language-response>

Anti-patterns

• Assigning the runtime state Attempted to an artifact row instead of Not Produced or Failed.
• Upgrading an artifact or runtime status instead of enforcing downgrade-only transitions.
• Accepting a prompt transcript or execute_blender_code snippet alone as artifact evidence.
• Accepting a pre-existing file without provenance linking it to the current run.
• Accepting CLI fallback evidence without both the exact command and exit code.
• Treating Path 2 (mcp-for-blender) output as official release evidence or omitting the non-release label.
• Merging preview, render, blend, and GLB/export into a single status instead of one row per artifact type.
• Blurring boundaries with pre-handoff-verification (gating handoffs), preview-report (rendering report Markdown), or glb-web-handoff (web handoff).

Cross-skill handoff

• [upstream_reference] Runtime readiness → ../runtime-readiness/SKILL.md
• [upstream_reference] Runtime attempt recording → ../runtime-attempt-report/SKILL.md
• [upstream_reference] Scene validation → ../scene-readiness/SKILL.md
• [progression] Handoff gating (confirm-or-downgrade) → ../pre-handoff-verification/SKILL.md
• [progression] Report Markdown rendering → ../preview-report/SKILL.md
• [progression] GLB/web handoff → ../glb-web-handoff/SKILL.md
• [progression] User-facing explanation → ../plain-language-response/SKILL.md

Non-goals

• Run Blender, render images, export GLB, or modify scene data.
• Install or configure runtime bridges.
• Gate handoffs or perform confirm-or-downgrade release decisions (pre-handoff-verification).
• Render the formatted preview or evaluation report Markdown (preview-report).
• Own GLB web integration handoffs (glb-web-handoff).

References

• ../../docs/skill-system.md
• ../../docs/runtime-stack-strategy.md
• ../../docs/evals/official-runtime-verification-criteria.md
• ../../docs/evals/official-runtime-manual-eval-packet.md
• ../../laws/evidence-before-done.md