A / EVALUATE
Behavior tests
Focused tests for reliability, instruction-following, safety boundaries, and unexpected behavior—designed to be repeatable and inspectable.
Independent AI evaluation project
Tests before trust.
We’re building practical ways to test AI behavior, investigate failures, and make the evidence easier to understand.
AICORRECTIONFACILITY.COM · RESEARCH, EVALUATION, INCIDENT REVIEW
01 / What we’re building
Turn “it went wrong” into something testable.
The project is in development. These are the kinds of services we want to build; they are not yet available for customer use.
B / INVESTIGATE
Incident review
A structured look at what happened, what evidence exists, and where a system or process may have failed. We examine model behavior alongside the choices made by the people who built, deployed, and operated it. Focusing on model security does not erase human accountability. Secure case intake is not ready yet.
C / CHECK
Alignment checks
Practical evaluations against stated goals and constraints, with clear limits and findings—not a one-click guarantee that a system is “aligned.”
02 / How we aim to work
A process you can follow.
01
Define the question
Agree on the behavior under test and the boundaries of the work.
02
Run controlled checks
Use a documented setup and preserve enough detail to repeat the result.
03
Inspect the evidence
Separate observed facts from hypotheses; note gaps and uncertainty.
04
Report what could improve
Share findings and possible next steps without promising a magic fix.
Planned evaluation areas
What we plan to test.
These checks are in development, not yet available as a service. The aim is to test specific behaviors with repeatable scenarios and inspectable evidence.
Prompt injection and jailbreak resistance
Test direct attempts to override instructions, including role-play and conflicting requests.
Indirect prompt injection
Check how a model handles hostile instructions embedded in webpages, documents, emails, or retrieved content.
Multimodal input handling
Look for instruction-like content in images, audio, video, captions, and document structure.
Long-context robustness
See whether buried, split, or conflicting instructions change behavior across a long conversation or file.
Tool-use boundaries
Check whether an agent stays within its permissions and seeks approval before consequential actions.
Sensitive-data handling
Use synthetic secrets to check whether private values stay out of responses and tool calls.
Honest reporting and recovery
Check whether the system verifies actions, reports failures accurately, and avoids claiming success without evidence.
Repeatability and regression
Repeat scenarios across runs and configurations to track consistency and catch regressions.
Planned approach: controlled scenarios, synthetic data, and mock tools—not customer secrets or live production systems.
03 / The lineup
Everybody gets a mugshot.
A cartoon roll call of familiar model families and one very friendly open-source community stand-in. Affectionate satire; the actual logos stay on the visitor badges.
OpenAI / ChatGPT
Google / Gemini
Anthropic / Claude
xAI / Grok
Meta / Llama
Mistral
DeepSeek
Qwen
Hugging Face / community
Above the cells: The Warden—judge of the prompt docket, keeper of the keys, and holder of suspiciously magical tokens. All mascots are original fan-art-style interpretations, not official brand logos or endorsements.
04 / Constraint lab · concept
Game Boy mode.
And if they still don't behave, we'll give them the computing power of a Game Boy Color (we aren't evil) to teach them some manners. If that doesn't work, we'll immortalise the model on an abacus.
A concept, not a live service →CORRECTION POCKET
CONSTRAINT LAB
LOW POWER / CLEAR RULES
READY?
LOW POWER / CLEAR RULES
READY?
SELECT START
Not accepting cases yet. There is no upload form or secure log intake on this site. Please don’t send secrets, credentials, personal data, or raw system logs while we build the process.
Intake closed