Product
From your rulebook to a signed record.
Six surfaces and three workflows. Everything starts from documents you already have: filed rate manuals, policies and regulations.
What you see
Six surfaces
Decision map
Lists every model, rule and human step in one decision, and which feeds which. Drafted automatically.
Used by: Model risk, underwriting or credit owner
Filing reader
Reads the filed rate manual, regulation or credit policy and writes down the right answer for every case.
Used by: Pricing actuary, product analyst
Decision re-check
Re-runs those cases, and your real past decisions, through a copy of your system or alongside the live one. Reports which rule set each outcome.
Used by: Actuary, IT
Data feed check
For each source of incoming data: fields that arrive missing or filled with a default, re-pricing with one field dropped, dollars lost per field.
Used by: Integration lead, chief underwriting officer
AI approval workspace
The nine-part sign-off pack for generative AI: fill it in, check it, sign in order. Cases that pass are re-checked on every update.
Used by: Model owner, model risk, compliance
Signed decision record
A record for each decision, signed at each hand-off: data received, how it was translated, defaults filled in, model and rate-plan versions.
Used by: Compliance, regulators, partners
How it’s used
Three workflows
Check the price against the filing
- A pricing analyst uploads the filed rate manual and its effective date.
- Tensyl extracts the rate tables and pricing formula, then builds one sample policy per table combination.
- Each sample policy is priced in a copy of your rating system.
- Every combination where the system differs from the filing is reported, with the policies affected.
- The checks are saved and rerun on every rate change, catching a missed filing on day one instead of month 21.
Audit each data hand-off
- Your integration lead shares records of what quote sites sent you.
- Tensyl compares what the shopper entered, what the translation step sent, and what defaults were filled in.
- Each quote is re-priced with each field corrected or dropped.
- The chief underwriting officer sees money lost by field and source, such as prior insurance sent as unknown.
- After the fix, the check reruns on every new batch of quotes.
Approve a generative AI change
- A vendor ships a new model version, or you add a generative AI step.
- Tensyl drafts the checklist and pass marks from the rule text.
- Two of your people check a representative sample by hand.
- The new version is checked, with error margins, and errors are traced through the rest of the decision.
- Owners sign in order: model owner, model risk, business owner, compliance. The next update reopens the checklist.
Inputs and outputs
What you bring, what you keep
| You provide | You get and keep |
|---|---|
| Official documents: filed rate manuals, credit policy, regulations | The right answer for every combination in the rate tables |
| Access to a copy of the decision system, or past records only | Mismatch report: filed value vs. applied value, policies affected, since when |
| Records of real decisions: quote-site submissions, worksheets, claim files | Which rule set each decision |
| The model and vendor versions in use | Money lost per missing or defaulted field, by data source |
| Answers your staff check by hand, and thresholds signed in advance | Signed approval pack, a re-check on every change, and a signed record per decision |
Getting started
Start with a file. Add access only when it pays.
Audit mode
One de-identified export of past decisions. No system access. You get a findings report.
System copy
Checks run against a copy of your rating or decision system, inside your cloud.
Alongside live
Checks run next to production, and every vendor release or rate change is re-checked before it ships.
The pilot: 60 to 90 days, one use case
Weeks 1 to 2
Scope, security review, and your risk thresholds in writing.
Weeks 3 to 6
Checks built from your files. Your team plants 25 to 40 known errors.
Weeks 7 to 10
Run, fix, re-run. We simulate one vendor update end to end.
Weeks 11 to 12
Your internal audit or outside validator reviews and re-runs the evidence.
Success means your planted errors are found at the rate you set, with no correct decision flagged, and your reviewer re-runs the checks without us.
No regulator has endorsed Tensyl’s evidence format. Pilots are scoped in writing before they start.
How it differs
Built for the people who sign off, not the people who build.
Monitoring, experimentation and testing tools serve engineers and data scientists: they tell you the software changed or a model improved. None of them knows what you filed, so none can tell you a price is wrong. Tensyl takes the right answer from the filing or the approved rule, and produces evidence an owner signs.
| Engineering tool | The question it answers | What it misses in a regulated decision | What Tensyl adds |
|---|---|---|---|
Monitoring Datadog, Arize, Fiddler | Did behavior or inputs change? | No notion of the correct outcome. A stable but wrong rule looks healthy. | Expected answers from your rulebook, so it flags wrong, not just different. |
Experimentation Optimizely, Statsig | Which variant performs better? | Prices can't legally vary outside the filing, and true outcomes take months to show up. | Checks that decisions match what was approved. |
Software testing Cypress, Selenium, pytest | Does the software do what the engineers' tests say? | Hand-written cases, software only. Skips models, people and hand-offs between companies. | Cases generated from the official documents, covering models, rules and people, with signed evidence. |
Model evaluation Braintrust, Patronus | Is this one model good enough? | Checks one piece. Every model can pass while the decision goes wrong at a hand-off. | Re-runs the whole decision and reports which rule set the outcome. |
Why not the vendors themselves?
A vendor whose system makes the decision can’t credibly certify its own outputs. Firms use many vendors and need one neutral check across all of them.
Works alongside your stack
Your engineers keep their tools. A vendor release or an alert from their monitoring can trigger a Tensyl re-check before the change reaches a customer.
Prototype results
A banking rule and an insurance filing
Checks built from the source documents found every planted error. A typical hand-built check found one or two.
| Reg E: bank card disputes | Small-business insurance pricing | |
|---|---|---|
| Where the right answers come from | The federal rule (12 CFR 1005.6 and 1005.11) and the regulator's 6 worked examples | MGT Insurance's filed Washington rate manual (public SERFF filing, effective Jul 1, 2025) |
| Answer key matches the source | 6 of 6 official examples reproduced | A sample business priced to the dollar: $2,513 |
| Checks built from the source | 3,672 | 1,630 |
| Correct decisions flagged as wrong | 0 | 0 |
| Planted errors found: Tensyl vs. a typical hand-built check | 9 of 9 vs. 1 of 9 | 7 of 7 vs. 2 of 7 |
| One-change errors found: Tensyl vs. hand-built | 57 of 58 vs. 34 | 23 of 23 vs. 10 |
| Example error caught | A late-report limit computed without its $50 sub-limit: $500 charged where the regulator's example says $150 | A surcharge the filing removed, still applied: $2,856 charged instead of $2,513 |
Errors were planted in mock systems, modeled on real enforcement cases. They are not errors in any company’s live system. A separate AI working only from the regulation extracted the rules with every number right. Not proven yet: runs on live systems, and a compliance reviewer’s sign-off on six open questions about how Reg E should be read.
Free audit
Send one export. Get a findings report.
Start where there’s no security review and no system access: one de-identified file of past decisions. A spreadsheet is fine. We check it against your own rulebook and send back what we found, with dollars attached.
Email only here. We’ll reply with a secure way to send the file, so please don’t include data in this form.
- 01
What we need
One de-identified export of past decisions, such as past quotes for one state and line with the inputs, rate version and premium charged. No access to your systems.
- 02
What you get
A findings report: every decision that doesn't match your filing or policy, the rule that caused it, and the dollars involved.
- 03
If it's useful
A 60 to 90 day pilot on one use case. Your team plants known errors, and your validator re-runs the evidence without us.