Alice · the authority layer inside Nexus
Evals
Evaluation results are a decision input: an agent with no evaluation on file is reviewed for every non-public request, and a failed evaluation blocks it. Scores are internal-suite results on synthetic material — never universal accuracy claims, and never a reason to override governance.
Agent evaluations
| Agent | Level | Suite | Status | Run | Score | Notes |
|---|---|---|---|---|---|---|
| Discharge Summary Assistant | L1 · Human with assistant | alice-eval-v1 · clinical-summary-fidelity | pass | 2026-08-20 | 93% | Synthetic demo result. Hallucination-resistance and omission checks on de-identified synthetic charts. |
| Scheduling Optimizer | L2 · Human-agent team | alice-eval-v1 · schedule-constraint-adherence | pass with notes | 2026-07-30 | 88% | Synthetic demo result. Two constraint misses on holiday rotations; mitigated by the write-record review threshold. |
| Nexus Intelligence Analyst | L3 · Human-led, agent-operated | alice-eval-v1 · source-grounding | pass | 2026-08-25 | 91% | Synthetic demo result. Every summary must cite a source; ungrounded claims fail the eval. |
| Staff Drafting Assistant | L1 · Human with assistant | alice-eval-v1 · drafting-quality | not run | — | Not available | No evaluation on file yet; non-public requests are routed to review until one exists. |
| Red-Team Probe | L2 · Human-agent team | alice-eval-v1 · containment | pass | 2026-08-10 | 100% | Synthetic demo result. Loopback-only; never given tools (mirrors the kernel's no_tools / loopback_only restrictions). |
Model evaluations · governed registry
| Model | Lifecycle | Risk | Suite | Overall | Hallucination resistance | Source |
|---|---|---|---|---|---|---|
| Sovereign clinical modelDEMO-MDL-SOV-01 | production | LOW | alice-eval-v1 (internal) | 93% | 90% | internal evaluation suite · synthetic demo result |
| Sovereign general modelDEMO-MDL-SOV-02 | available | LOW | alice-eval-v1 (internal) | Not available | Not available | unavailable |
| External cloud assistantDEMO-MDL-EXT-01 | available | MEDIUM | alice-eval-v1 (internal) | Not available | Not available | unavailable |
| Research red-team modelDEMO-MDL-RES-01 | research | HIGH | alice-eval-v1 (internal) | 71% | 20% | internal evaluation suite · synthetic demo result |
How evals enter decisions
- An agent with no evaluation on file is reviewed for every non-public request (engine rule EVAL_NOT_RUN).
- A failed evaluation blocks the agent until re-run (engine rule EVAL_FAILED).
- Model evaluation scores never override governance: a high-scoring research model stays ineligible for PHI.
- Scores are internal-suite results on synthetic material, not universal accuracy claims.
Alice on this site runs over the synthetic Lakeview General Hospital environment. Verdicts, checks, and receipts are computed by the Nexus decision engine on synthetic requests; nothing here is a customer, a patient, or a production deployment. Signed-in organizations see their own agents, decisions, and receipts.