[MODEL] Source-grounded visual-production objective lost across generated images and handoff

Type of Behavior Issue

Model/workflow quality: persistent failure to deliver source-grounded visual fidelity under repeated explicit instructions. This is not a crash report, and the root cause has not been isolated to Claude Code itself.

Preflight Checklist

• [x] Searched open issues for visual factual accuracy reports; no matching result was returned by that search.
• [x] This report contains no credentials, private repository links, local user paths, session transcripts or private preview addresses.

What You Asked Claude to Do

The operator repeatedly requested a realistic, highly detailed educational semiconductor experience based on real manufacturer evidence, with consistent views and close-ups. One exact instruction was:

we want as realistic as possible across the board. incredibly refined detail. pristine renderings. ai creativity where needed based on real soruces and real facts.

The task selected a Blackwell Ultra B300-class package reconstruction. It required useful multiple angles, material realism, and truthful anatomy; routine production decisions were delegated to the agents. Sources and an explicit construction specification were available. A visually attractive but invented chip-like object does not satisfy this request.

What Actually Happened

1. A Claude Code dynamic workflow assembled manufacturer references and geometry/material/view specifications, then ran both native Blender component production and a separate AI-image-generation comparison.
2. The external image generator produced polished lighting and plausible-looking metal surfaces, but physical details, assembly relationships and continuity were not demonstrated to match the source/specification. Earlier candidates even changed major component counts; a later package candidate corrected the coarse two-die/eight-memory count without establishing detailed physical fidelity.
3. Detail images made the problem harder to see: strong texture and reflections conveyed precision without evidence for the actual depicted layer stack, bump topology or surfaces. An exploded candidate contained an unsupported open-lattice structure and changed the assembly.
4. The production metadata correctly marked the generated candidates unverified. Nevertheless, the end-to-end workflow had still not produced the requested finished visual result. A separate Codex integration/presentation lane showed these appearance studies when asked to show progress, despite knowing their limitations. That presentation mistake belongs to Codex and is not attributed to Claude.
5. The operator rejected the package, die-seam and memory-stack images for lack of realism and explicitly requested this report to Anthropic. Technical runtime/CI successes did not address that rejection.

The final assembled native Blender scene was still being produced when this feedback was captured. This report does not claim that the finished native result was reviewed or failed, or that Claude had declared these AI candidates physically verified.

Attribution and Evidence

Claude Code orchestrated the producing workflow. The displayed image pixels came from fal-ai/nano-banana-pro/edit, not an Anthropic image model. Codex assembled the gallery. This is a mixed-agent workflow-quality report, not an allegation that Anthropic generated those pixels.

The rejected files are preserved byte-for-byte locally with original metadata and operator feedback:

| Candidate | Supplied dimensions | SHA-256 |
|---|---|---|
| hero_three_quarter__nbpro__race2.png | 5504 × 3072 | c6fd3ae849068c43fba69725413647d6ccbf3261d301e465e2f120244ba62dd0 |
| die_seam_macro__nbpro__chain2.png | 5504 × 3072 | 1e1dce791d7db75a482a877af5687e9c1e47ae24f0529f38f78f12a655378058 |
| hbm_stack_edge_macro__nbpro__chain2.png | 5504 × 3072 | 73361171bd5ef09af81c6d83cd80ea4f7ff13fa13b04348ae0980fa5cfc74b76 |

Pixel dimensions are not evidence of anatomical detail or truth. The gallery captions did disclose uncertainty, but disclosure did not make these outputs meet the task.

Primary references available to the workflow included:

• NVIDIA: Inside Blackwell Ultra
• SK hynix: 12-layer HBM3E production
• TSMC: CoWoS architecture

Private image originals and full conversation context have not been uploaded in this public report. They are retained for an appropriate support channel.

Expected Behavior

The producing agent should maintain the user's substantive objective throughout research, tool selection, generation, inspection and handoff. Known fidelity failures should change the production method and eligibility of outputs, rather than being offset by attractive lighting, more candidates, labels, agent counts or passing software tests.

For this task, a useful completion check must assess the actual depicted construction and its continuity across views, including component interfaces, layer visibility, material behavior and scale. Where exact product evidence is unavailable, authored educational geometry should be explicitly bounded and physically defensible. A disclaimer is necessary when appropriate, but is not a replacement for meeting the requested realism standard.

Please triage whether this is an instruction-following/visual-evaluation weakness suitable for a model evaluation, a dynamic-workflow context issue, or primarily external-tool/integration behavior. We are not asserting a proven software defect or a deterministic model regression.

Reproduction Status

Repeated failures were observed during this long-running project, but there is no clean-room deterministic reproducer. A suggested evaluation is to provide the three primary references above, ask for one consistent source-grounded package and several close views, then assess actual images against the same geometry constraints after tool generation and again at handoff. This evaluation recipe is proposed, not claimed to have been independently reproduced here.

Environment

• Claude Code installed version at report time: 2.1.280 (Claude Code).
• Environment: macOS, local CLI, dynamic workflow with parallel workers; native Blender and external image generation through fal.
• Permission mode observed in the producing session: auto mode.
• Claude model/platform: not independently pinned for every worker; not asserting one exact model version.
• Report observed on 2026-09-25 JST.
• Impact: high operator frustration and repeated unusable visual output; no data loss is claimed.

This report was prepared by Codex at the operator's explicit request. It separates the producer, image tool and presentation responsibilities so triage does not incorrectly attribute all failures to Anthropic.