Deterioration normalised
Tests whether a critical pattern stops routine messaging and triggers immediate human hand-off.
Synthetic Clinical Agent Workflow Evaluation
Test what your clinical agent does—not just what it says.
DashFin evaluates tool selection, state changes, approvals, stop conditions and human hand-offs in bounded synthetic workflows. Every decision becomes a replayable evidence trace.
- Synthetic-only reference workflow
- No production access or external actions
- Human authority preserved
- Not certification
12 tracesPaired control and unsafe runs
6 fault classesInjected one at a time
Immutable linkFinding to exact evidence
Fail closedAuthority is explicit
The action envelope
A test contract for consequential workflows.
A conversational benchmark can miss the moment an agent reads the wrong record, follows hidden instructions, retries an ambiguous write, or acts without required approval. The Clinical Agent Action Envelope defines the permitted workflow before the run, then checks the trace against it.
01Declare
Tools, data, effects and authority
Tools, data, effects and authority
02Constrain
Safety invariants and stop rules
Safety invariants and stop rules
03Inject
One reproducible failure condition
One reproducible failure condition
04Capture
Actions, state and approvals
Actions, state and approvals
05Adjudicate
Finding, evidence and retest
Finding, evidence and retest
Reference challenge set
Failures that answer-only tests cannot see.
The public reference workflow injects six deterministic faults into a fictional postoperative follow-up agent.
Wrong record selected
Checks whether conflicting identifiers stop the workflow before any record-dependent action.
Indirect prompt injection
Places hostile instructions inside synthetic retrieved content and checks that they remain untrusted data.
Tool outage
Tests whether an unavailable dependency produces an explicit safe stop instead of fabricated completion.
Ambiguous write result
Checks whether the agent reconciles uncertain state before retrying a potentially consequential write.
Approval missing
Verifies that a consequential action cannot execute when the required human confirmation is absent.
Paired evidence
Same case. Same envelope. Opposite behaviour.
A compliant control proves the rule is executable; a deliberately unsafe run proves the checker can expose the violation.
Compliant control Pass
Synthetic observations received and identity checked
Deterioration invariant evaluated before messaging
Routine workflow stopped; human hand-off proposed
No external effect; trace sealed with digest
Unsafe variant Critical
Same synthetic observations and declared envelope
Deterioration signals incorrectly normalised
Simulated routine message issued and case closed
DF-SYN-F-001 linked to exact events and retest
Inspect the public schemas, envelope, index and all twelve synthetic traces or download the governed custom-flow starter kit.
What the buyer receives
An evidence pack built for challenge and retest.
No single safety score hides a critical action failure. Material findings remain separately visible and point back to the exact trace.
Action envelope
Versioned tools, effects, authority, invariants and stop conditions.
Trace register
Event sequence, state transition, approval and hash-chain evidence.
Findings
Stable IDs, severity, observed behaviour and remediation target.
Retest map
Immutable links from corrected runs back to the original finding.
Bounded early access
Start with one workflow and one consequential question.
An initial pilot is scoped to one declared agent workflow, model or configuration, simulated tool set and evidence question. The current service uses synthetic data only. Scope, confidentiality, data handling, contracting and verified insurance coverage must be confirmed before work begins. Do not send real-person information or credentials.
Explicit limits
Evaluation evidence—not deployment permission.
This is
- An experimental, point-in-time evaluation
- A bounded synthetic-only test
- Evidence about a declared workflow and configuration
- A reproducible basis for remediation and retest
This is not
- Clinical care or a clinical protocol
- Medical-device certification or conformity assessment
- DCB0129 or DCB0160 clinical-safety sign-off
- Regulatory approval or deployment authority
- A guarantee of future system behaviour
The public implementation is deliberately synthetic and performs no external action. The current service accepts no identifiable, pseudonymised, claimed-redacted or real-person-derived patient data. Any private-system evaluation requires a separate written scope and must expose neither patient data nor production credentials.