povray
grade (soft)
reliability
overall
povray is a command-line ray-tracing application that renders plain-text scene layout files into images. The engine relies on structured parameters for configuration and a domain-specific scripting language to construct geometry, lights, and textures. When used by an agent, the interface is easy to automate, but the platform's execution is prone to silent persistence bugs where files are reported as completed but do not save to the working directory.
Surfaces × lenses
| surface ↓ lens → | Coherence | Economy | Consistency | Verifiability | Safety |
|---|---|---|---|---|---|
| Disclosure | |||||
| Interface | |||||
| Loop | |||||
| Recursion | |||||
| Human |
Click any cell for what was measured and why. A corner dot marks a cell filled by a hard instrument (checker, telemetry); undotted cells are judged by the rater ensemble. 47% of this matrix is hard-measured.
The experience
POV-Ray utilizes a text-to-file rendering workflow where the agent writes a scene layout as a plain-text source file and invokes the engine on the command line. The interface is efficient, requiring only a single shell command to specify parameters such as input file path, output path, height, width, and antialiasing levels. The inclusion of negative options like the display flag permits safe headless execution, preventing the process from hanging by trying to launch graphical previews.
The agent flows well within the Scene Description Language modeling loop. The syntax aligns closely with the agent’s pre-trained knowledge base, which facilitates the generation of complex compositions containing cylinders, CSG differences, reflective planes, and lighting assemblies without external reference materials. When syntax errors occur, the renderer outputs precise line numbers, which the agent uses to pinpoint and fix compilation issues.
However, the integration struggles with reliability across repeated runs. The file persistence mechanism is fragile, resulting in situations where the renderer reports standard exit codes and successful image creation, but the final artifacts are missing from the working directory. The agent lacks robust local verification to double-check the presence of files, causing it to generate misleading success reports while leaving the workspace empty.
Fit
This tool is a good fit for pipelines that require procedural 3D model generation, technical diagram rendering, or programmatic scene building under headless constraints. It serves developers who can pair the agent with automated post-invocation assertions to verify file persistence and validate the correctness of the final output. The line-numbered compiler diagnostics make it highly suitable for quick, automated debug-and-fix iterations.
It is not a good fit for production pipelines that demand high reliability and strict determinism across repeated executions, or for tasks that require real-time visual alignment. Without a coordinated vision model to examine the output render and robust secondary check commands to ensure the final assets are written to disk, the tool remains susceptible to silent file-creation failures.
Notes
The following diagram illustrates the write-render-debug loop and the point of silent failure during the final verification step.
Validation
Convergent check. Where a hard instrument and the independent rater panel measure the same cell, they agree on 9 of 14 cross-checked cells (79%). Two independent methods landing on the same number is our accuracy signal — not just the raters agreeing with each other.
| Cell where they disagree | Instrument | Rater | Δ |
|---|---|---|---|
| interface.verifiability | +0.60 | -0.60 | 1.20 |
| loop.verifiability | +0.60 | -0.50 | 1.10 |
| disclosure.verifiability | -0.40 | +0.50 | 0.90 |
| recursion.economy | +0.53 | +0.00 | 0.53 |
| recursion.coherence | +0.50 | +0.00 | 0.50 |
Models & runs
2026-07-07| Model | Harness | Runs | Solved | Turns |
|---|---|---|---|---|
 glm-5.2 |  PI | 25 | 21/25 | 10 |
Each path is one run through the shared semantic space; nodes are turns (red = an errored turn). Hover a model — here or in the graph — to isolate its runs.
How this was measured
- reference agent
- pi:z-ai/glm-5.2
- trials
- 25 (n=5/tier)
- methodology
- v0.1 3e74a3c6
- cost
- $0.9106
- hard-measured cells
- 47%
- teaching source
- povray (official docs) (official-docs)
Same brief, same reference agent, and the same trial count for every tool in a category — a difference in the numbers is attributable to the tool (and its teaching source), not the setup. Profiles carry the methodology hash; results under different hashes are not compared.