# SSE-2026-006 · The bust it couldn't build alone

**Version** 1.0.0 · **Type** case record (inspection level) · **Researcher** Sai Bharath Siliveri (Roi)
**Case study** https://siliveri.systems/research/the-bust-it-couldnt-build-alone/

## The evaluation question

When a model's first attempt fails, how much of that failure is the model's limit, and how much is waiting for the right diagnosis?

**System:** Frontier models under identical instructions; the successful build by a Gemini Flash model with human diagnosis.
**Task:** Build a convincing 3D human bust out of particles in one self-contained HTML file: no 3D engines, no assets, only code and projection math. Pass condition: the form, including an open mouth, holds under rotation.

## What was observed

- Unaided attempts failed three ways: a front-only silhouette, a geometric fallback (triangular orb), and collapse into a blob or diffuse field.
- Diagnosis: viewpoint-dependent depth sorting, a mouth painted as texture rather than carved as a void, and shape checked after generation instead of enforced during placement.
- With one targeted change per diagnosis, a Gemini Flash model produced a bust that holds volume and the open mouth through rotation.

## Evidence and its release state

| ID | Evidence | State | Note |
|---|---|---|---|
| F-14 | Rotation film of the final bust, recorded live (published on the case page) | public | https://siliveri.systems/research/the-bust-it-couldnt-build-alone/ |
| SRC | The final self-contained HTML file | located, not released | five candidate files found on the researcher's machine; the owner must confirm which one is the filmed build before release |

## Corrective action

The fix sequence itself: constraint-field shape enforcement, carved mouth void, re-projection, rotation check before each change.

## Reproduction boundary

This record lets a reviewer **inspect** the case: what was asked, what happened, what it rests on and what is still private. It does not yet support reproducing the analysis or re-running the evaluation, because the primary evidence listed above has not been redacted and cleared for release.

## Limitations

- Does not show no model can do this unaided; it shows how much a correct diagnosis recovers.
- Model versions for the failed attempts are not itemized in the public record.

## Licence and availability

© 2026 Sai Bharath Siliveri. Published for review, evaluation and citation with attribution. No other licence is granted yet. Contact: hello@siliveri.com.