For regulated work
An answer for the day someone asks why.
In lending, insurance and healthcare, an AI decision is not finished when it is made. It is finished when it can be explained to an applicant, an examiner or a court, months later. Vertep records the process as it happens, so the explanation is a replay and not a reconstruction.
Three places it fits
The pattern is the same: the model takes the routine cases, the chart refuses what policy forbids, and a person owns the consequential ones.
Lending
A model underwrites; a hard ratio limit declines without asking it; approval is not offered above the ceiling; an officer decides the referrals. Runnable: three real applications with their audit.
Run the example →Healthcare prior authorisation
A model may approve routine requests or refer. Denial exists in exactly one state, and a clinician decides it. Statutory deadlines are durable timers that escalate to a named role.
The chart →Insurance claims
A model fast-tracks small, clean claims and flags the rest. Only an adjuster can refuse payment, and the reason is journaled with the decision.
The chart →The questions an examiner asks
And where each answer lives. None of these needs a separate logging system, because the record is what ran.
| Question | Where the answer is |
|---|---|
| What was the process in when it decided? | The state configuration as of the decision's journal entry (state_at). |
| What was it allowed to do? | The enabled events the chart offered at that decide state. Anything else was not on the table. |
| Who or what chose, and what did they say? | The decision entry: the chosen event, the decider's identity and its rationale, kept as the decider's claim. |
| Under which version of the rules? | The chart's version is the hash of its canonical form, recorded with the entries. |
| Was a person involved, and when? | The chart names which states are decided by people. A person's decision is a journal entry like any other. |
| Can you reproduce it? | Replay the journal with no decider bound: same states, no model calls. |
| What would the new rules have done? | Fork from the journal under another chart version and compare, without touching the original record. |
What it contributes to the rules you answer to
A record is evidence. It is not compliance. This table says which part of each requirement the record helps with, and which stays yours.
| Regime | What it asks, in short | What Vertep contributes, and what it does not |
|---|---|---|
| EU AI Act, Articles 12, 13 and 14 | Automatic recording of events, transparency to deployers, and human oversight, for high-risk systems such as credit scoring and life or health insurance risk assessment. | The journal is an automatic, ordered event record; human-decided states make oversight structural. Classification, risk management and conformity assessment stay with you. |
| GDPR, Articles 15(1)(h) and 22 | Meaningful information about the logic of automated decisions, and limits on decisions based solely on automated processing. | The chart and journal show the logic and the human steps. Whether a given decision needs a human in the loop is a legal call; the chart lets you build one in. |
| GDPR, Article 17 | Erasure of personal data. | An append-only journal is in tension with this. Today, keep personal data out of events and journal references instead. Payload encryption with per-subject keys is specified in the draft m12-decision-records, and not built. |
| US ECOA and Regulation B | Specific reasons when credit is denied. | The journal records the decline and its rationale, and the chart can require a reason on the event. Choosing lawful reason codes is yours. |
| SR 11-7 model risk management | Inventory, controls and ongoing monitoring of models used in decisions. | A model's role is bounded and logged at named decide states, which gives a clear inventory of where models act. Validation of the model itself stays with you. |
| FDA 21 CFR Part 11 | Secure, time-stamped audit trails for electronic records in regulated life-science work. | Journals are append-only and time-stamped. Signed, hash-chained journals exist in durable objects. Validation, access control and signatures are your system's work. |
This is a technical summary, not legal advice, and Vertep is a tool, not a certification. Obligations depend on your jurisdiction, product and risk classification; check them with counsel.
Said plainly: what it does not do
- It does not open the model. A model's stated rationale is recorded as its claim, labelled as such. It is not a faithful account of the weights, and Vertep does not pretend otherwise. What is verifiable is the process around the model.
- It does not make a policy fair. A chart can encode a biased threshold exactly. It makes the threshold visible, versioned and testable, which is where fairness review has to start.
- It does not erase from a journal yet. See the Article 17 row above, and the draft specification for the fix.
- It is 0.1.x software. It has not run a lender's production load. Use it first where a failure is cheap, and where its record is worth more than its maturity.
- A plain journal is not signed. Hash-chained, signed journals come with durable objects; extending that to every agent journal, and an exportable explanation bundle, are in the draft
m12-decision-records. - It is not a compliance product. It gives you an honest record and structural limits. The programme around them is yours.
Start with one decision
Pick one consequential decision a model currently makes or drafts. Write its legal moves as a chart. Bind the model at the one state where it has to choose.