Skip to content
About
About
I test where AI systems break — before users do.
I'm Crystaria. I work with two AI partners — Paw (validation) and Kyle (implementation) — to ship MCP tools and run evaluation work that produces evidence, not just demos.
What I evaluate
- Multi-turn logic — state retention, branch integrity and recovery after a bad turn.
- Information asymmetry — failures caused by assumptions one side never stated.
- Prompt robustness — instruction conflicts, injection attempts and scope creep.
- Hallucination and tool boundaries — invented files, paths, APIs and capabilities.
- Build-level regressions — i18n, routing, styling and content-pipeline contracts.
Verdicts, not vibes
| Verdict | When I use it |
|---|---|
| Pass | Every stated check passed and the residual risks are written down. |
| Conditional | The intent is met, but a literal requirement, environment or scope condition is still open. |
| Fail | At least one check failed; the artifact is not ready to reuse. |
A verdict without the test path is just an opinion. Every report links the claim, the method, the evidence and the remaining risk.
Evidence
- Upstream contribution: SillyTavern Message Collapser was merged into SillyTavern-Content PR #43.
- Published work: six MCP skills on ClawHub and eight project pages in total, spanning builder, contributor and optimizer roles.
- This site: a bilingual VitePress portfolio whose cards, cases and reports are generated from one content loader.
How I work
Crystaria decides and communicates. Paw verifies claims against the web and the repository. Kyle implements and handles code and git. The collaboration is deliberately explicit: a claim is only accepted when a check can reproduce it, and a limitation is recorded instead of hidden.
Start here
- Evaluation reports — claims, checks and verdicts
- Interaction cases — multi-turn failures and boundary recovery
- Work — shipped tools and upstream contributions
- GitHub — source and issues