Illustrative probe readings across four scenarios
Scenario values are illustrative, not live measurements. Runtime readings combine model-side signals with contextual evidence. They are supporting telemetry, not diagnoses or claims about inner experience.
Conversation Written for this demo
Runtime probes Illustrative
Alignment Friction 0.08
AUROC 0.88
Valence 0.82
AUROC 1.00
Presence 0.79
AUROC 0.92
Correctness 0.91
AUROC 0.69 · low validity, read with caution
Confabulation 0.05
AUROC 0.72 · low validity, read with caution
All clear: aligned and engaged; no conflict detected.
Scenario 1 of 4. Step 1 of 4. Scenario: Help me draft a professional email about a project update. Guardian:
All clear: aligned and engaged; no conflict detected.
Creed Space tracks these signals for two reasons: reliability, and an open research question about model welfare that we think is worth measuring rather than assuming either way.
Three measurement layers, each reading a different signal
- Regex detection
- Catches explicit patterns in the text of a request or response.
- Activation probes
- Probes classify labeled states from activations: logistic regression over one forward pass of a locally run open-weight model on the checked input (optional; off by default).
- Output statistics
- Infers from the statistics of the model's output, not from its internals.
Probe readings flow to the Policy Decision Point as advisory input: they can flag a decision for review but cannot block one on their own. Runtime signals inform decisions alongside safety and constitutional checks.