Systems
AI SystemsMethodWorkOperationIndustriesField Notes
Solutions
EnterpriseRegulated IndustriesProcessesIntake and ServiceDocument ReviewReconciliation
Integration
Systems of RecordModel GovernanceAudit TrailChange ControlBlueprintData Residency
Company
AboutEngagementsSecurityOversightCareersContact
EngagementsNews
Talk to UsScoping Session

Model GovernanceIn every blueprint

Model governance:
every step, its own model.

Each step of a system runs on the model chosen for it, written into the blueprint and changed only under change control. Never trained on your data.

  • Chosen per step
  • Written down
  • 25+ languages
  • Phase 01

How a model is chosen and changed

Chosen per step

Every step of a system runs on the model that suits it: reading, drafting, checking or deciding.

  • Each step matched to the model that does it best
  • Reasoning models only where a step needs judgment
  • Smaller models wherever a smaller one does the job

Written down

Every model choice is written into the blueprint: the step, the model, the reason, and who signed it off. Pick a process to read its entry.

  • One entry per step, in your team's own words
  • The reason for each choice, not just the name
  • Signed off by your team before anything is built

Step 2, claims intake. Model chosen for long policy documents and handwritten forms, because it reads the whole file before it drafts. Signed off by the Head of Claims.

Tested on your own cases

Before a model is chosen or changed, it is tested on your own past cases, inside your environment, and the results go to your team.

  • Tested against cases your team already decided
  • Run inside your environment, nothing copied out
  • Results reviewed by the people who do the work
  • The same test rerun before any change

Change control

A model changes only under change control: the candidate is compared with the current one, reviewed, and signed off before it goes live.

  • Current and candidate compared side by side
  • Version pinned, so nothing changes on its own
  • Every change logged with who approved it

The full governance record

Everything written down about the models a system uses, from the first choice to the last change.

  • Model per stepEach step matched to the model that suits it
  • Written reasonsWhy each model was chosen, in plain words
  • Named ownersEvery choice signed off by a person on your side
  • EvaluationTested on your own past cases before it is chosen
  • Change controlNo change ships without review and sign-off
  • 25+ languagesDocuments read in the language they arrive in
  • Written downOne record of every model choice, kept current
  • Version pinningNothing changes on its own between reviews
  • FallbacksA person takes over when a model is unsure
  • ThresholdsConfidence limits set per step, per process
  • Model familiesChosen across providers, never locked to one
  • Audit trailEvery output logged against the record it touched
  • No training on your dataNever used to train a model, ours or anyone's
  • ResidencyProcessing kept where your regulators require
  • RetirementModels replaced on schedule, under change control

One model per step, in 25+ languages

Review

  • Insurance
  • Documents

Careful, thorough, and always cited.

Phases

Governance in every phase

Model choices are made in the blueprint, reviewed at go-live, and revisited only under change control.

Read the Method
  1. Blueprint

    Each step's model chosen and written down

    Phase 01

  2. Go-live

    Every choice reviewed by your engineers

    Phase 02

  3. Operation

    Changed only under change control

    Phase 03Ongoing(after go-live)

Need a model choice reviewed?

Talk with Lyrion about a model choice, a change request, or a review of what runs today.

Enterprise

Controls that hold up in a review

The controls a security review asks about, in place before go-live.

  • Your EnvironmentAble to run inside your own environment where residency requires it.
  • Named ownersEvery model choice signed off by a person on your side.
  • DPA supportData processing agreements and residency where your regulators require it.
  • Monitored after go-liveOur engineers watch every model in production and act when one drifts.
  • Approvals wherever your team wants themThresholds set per step, so routine work flows and exceptions go to a person.
  • Access and audit loggingSingle sign-on, role-based access, and every action logged against the record it touched.
  • Never used to train a modelYour data never trains a model, and retention follows your policy.

Ready to scope your first process?

Tell us about the process you want handled, or read how each model is chosen and changed.