Every response generated by ALEX first passes through an independent governance runtime that decides not only how the model should answer — but whether it should answer at all.
Each chat turn traverses every stage before Claude is called. The Sovereign layer makes the final participation decision.
Scientific claims backed by real governance decisions from production, updated in real time.
Averaged over last 500 decisions.
Anonymous — no user data, no prompt content. Only the governance computation is shown.
The real scientific question is not what the runtime decided — it is whether people agree. These metrics will populate as feedback labels accumulate.
Metrics update as users submit feedback from the Living Lab. No data yet — bridge is 0 days old.
The public observatory validates the governance architecture in production. The same runtime is available as a model-independent API for institutions that need enforceable response policies, audit trails, and human escalation.
"Separating Cognition from Participation: A Governance Runtime for Trustworthy Large Language Models"
ALEX treats participation itself as a computational decision rather than an emergent property of the language model. This distinguishes the approach from constitutional AI (model self-governs) and safety filters (post-generation pattern matching).
We are seeking academic collaborations on:
"Anonymous" on this page is a specific technical claim, not a label. Here is precisely what the governance layer computes over, and what it never touches.
This section is not optional — it is the most important part of this page. We believe honest disclosure of current limitations is more valuable to the field than any claim of achievement.
CRF is a proxy metric. The Cognitive Risk Factor is computed from VIX and regime confidence — it is not a direct measure of epistemic uncertainty. We treat it as a signal, not a ground truth.
Output modes are under active evaluation. The thresholds that determine SPEAK / SPEAK_WITH_CONSTRAINTS / REMAIN_SILENT were set a priori and are being validated against user feedback for the first time. The Living Lab study is that validation.
The governance runtime does not guarantee correctness. It is designed to modulate confidence — not to produce correct answers. A SPEAK decision means the system believes a response is appropriate, not that it is factually accurate.
User labels carry self-selection bias. Participants who choose to provide feedback may differ systematically from those who do not. We are designing controls to quantify this.
Domain is financial advice only. All data comes from financial Q&A contexts. Governance behavior and user calibration may differ substantially in other domains. Generalization claims will require additional studies.