01
Observe
Read the system
We read the system's inputs, state, recent updates, and what it's allowed to do. Read-only. We don't touch anything.
One governance layer. Five stages. One evidence artifact. Sealed forecasts.
Velma sits above control and below intelligence. We don't replace your stack. We watch every decision, anticipate where it is heading, intervene when the trajectory drifts, and sign the record so the regulator can read it.
Read left to right. Policy envelopes enter from above. The evidence artifact exits to the regulator below.
- InputAdaptive AI system
- 01Observe. Context ingestion
- 02Assess. Envelope check
- 03Modulate / Anticipate. Bounded authority, trajectory-aware
- 04Enforce. Runtime intervention
- 05Audit. Hash-chain signature
- OutActuarialReport · evidence artifact
A five-stage pipeline, assured end-to-end.
02
Assess
Check against the rules
Each decision the system wants to make is checked against the active rules. Rate filings, ASIL targets, PCCP limits, fairness constraints, reserve policy. We measure how close it is to the line. We don't guess.
03
Modulate / Anticipate
Read the trajectory, soften before it slips
Inside the envelope, the decision passes. Near the edge, we narrow it. And we project the substrate forward at one, five, ten, and twenty decisions ahead. If the predicted trajectory crosses a drift, cascade, or escalation threshold, the verdict downgrades before the action commits. Anticipatory, not reactive. Graceful degradation, not a circuit breaker.
04
Enforce
Block what breaks
Out-of-envelope decisions are stopped before they ship. Learning updates that would break the safety case roll back to the last good version. The system keeps running. Safely.
05
Audit
Sign the record
Every decision (passed, slowed, or blocked) is signed onto a tamper-evident chain. That chain is the evidence. It's what the regulator, the auditor, and the reinsurer all read.
What the auditor actually sees.
Not a slide. Not a PDF. The live Audit Explorer reads the signed chain and puts every governed decision a single click away. These three panels are pulled from real substrate output. The same records the controller, internal auditor, or regulator opens.
247
193
44
10
Every chain opens with shape. How many decisions, how clean, where the failsafes fired.
LTV 1.10 outside [0.00, 0.80]
0.443
0.278
0.41
Click any record. See what the agent proposed, which rule fired, and what Velma did.
41e97ba546879737…
c343b8d6a23a7270…
cd8bf22df6e8bb98…
Every record's hash binds to the one before it. Tamper with anything and the chain breaks visibly.
What a governed decision actually contains.
One record per governed action. Human-readable, machine-readable, regulator-readable. Cryptographically chained. Exported in whatever filing language the authority on the other side of the table speaks.
Mean time between failure
Availability
Risk grade (A - F)
Constraint pressure
Governed-decline rate
Near-misses intercepted
Trajectory modulations fired
ERI prediction hit rate (h = 20)
Reconciliation receipts at horizon
Learning rollbacks & roll-forwards
Fairness / subgroup drift
Trust trajectories
Audit chain (hash-signed, rotated)
ERI chain (sealed forecasts, parallel)
Non-invasive by design.
We do not replace your underwriting engine, your autonomy stack, or your model catalogue. Velma runs alongside. Reading state, signing decisions, and producing evidence. Removal is a configuration change, not a re-architecture.
- Cloud · on-premise · embedded C++17
- Sidecar to decision path · zero changes to underlying controller or model
- Read-only observation of model state and decision context
- Sub-millisecond decision overhead on governed actions
- Configuration disable · no rewrite · no trapped state
- Your own systems · Velma holds no customer data by default
Built for the people who consume the evidence.
The buyer is the Chief Risk Officer. The daily user is the controller and the internal auditor. The final reader is the regulator. Velma is shaped for all three.
Controllers & internal auditors
The evidence arrives pre-assembled in the format their workpaper tool already uses. AuditBoard, Workiva, OpenPages, ServiceNow GRC. Reconciliation work that took days takes minutes.
Chief Risk & Compliance Officers
Fewer enforcement actions. Lower required capital under model-risk reserves. Faster product launches because the governance proof is built in, not bolted on after.
External auditors & regulators
The evidence speaks NAIC, Solvency II, ISO 26262, FDA PCCP, SR 11-7, and the EU AI Act natively. No translation layer between you and the examiner.
The artifact speaks the regulator's language.
NAIC Annual Statement
Solvency II SFCR
AM Best BCAR
ISO 26262 safety case
DO-178C / DO-254 evidence
FDA PCCP dossier
SR 11-7 model risk package
IFRS 17 reserve attestation
The full spec is a conversation, not a PDF.
Thirty minutes. What plugs in on your side. What the pilot looks like.