A reserve you can't reproduce
is an opinion.
Insurance runs on numbers that have to survive an actuary, an auditor and a supervisor — often years apart, looking at the same quarter. Most of the machinery producing them is a chain of spreadsheets, vendor black boxes and reconciliations that nearly close. We serve this sector with software, not cranes : deterministic compute on Heisen where a triangle, a rating factor or a cession re-runs to the identical number, and where the model's assumptions are recorded next to the result. The actuarial judgment stays with the actuary. We make the compute under it reproducible.
Reproducible
Same quarter, same inputs, same number
Traceable
Every assumption recorded with the result
Firewalled
Maxor Ground on AI-assisted claims workflows
Canada
Hosted here — policyholder data stays in jurisdiction
Six exposures in the compute behind the balance sheet.
Every exposure below is a place where the number reported and the number reproducible have drifted apart. None of them are actuarial judgment problems — they are engineering problems wearing an actuarial coat.
A reserve that cannot be re-derived
The triangle closed, the number went in the statement, and eighteen months later nobody can re-run it — the spreadsheet moved, an input changed, a macro was fixed. The reserve is defensible only as far as its derivation is reproducible.
Vendor catastrophe models as black boxes
Cat model output drives capital, reinsurance purchase and board conversations, yet the internals are a licensed opaque box. Without an independent deterministic check on the exposure feed and the output shape, model risk is unmeasured rather than accepted.
Bordereaux and cession reconciliation
Treaty accounting where the cedant's view and the reinsurer's view differ by an amount nobody can decompose. The difference is usually deterministic and findable — it just is not computed the same way twice.
Rating factors that drift silently
A rating engine amended repeatedly, with the effective-dated version of each factor no longer recoverable. When a regulator asks what rate applied to a policy written on a given day, the answer must be computed, not reconstructed.
Hallucinated claims decisions
A language model summarizing a claim file, drafting a denial rationale, or extracting a coverage position will occasionally invent something coherent and wrong. In a regulated claims process that is not an inconvenience — it is a market-conduct exposure.
Policyholder data residency
Policyholder and claimant records carry Loi 25, PIPEDA and supervisory expectations about outsourcing and residency. A US-hosted platform carries CLOUD Act exposure regardless of contract terms. What we host stays in Canada.
Six workflows where reproducibility is the deliverable.
These are the computations that get re-examined by someone who was not in the room. Each one we can make deterministic, versioned, and re-runnable exactly as it stood on any past valuation date.
Reproducible triangles and roll-forward
Development triangles as deterministic compute with versioned assumptions, so a prior valuation re-runs to the identical number and the movement between quarters decomposes cleanly.
Effective-dated rating compute
Rating factors and algorithms held as versioned, effective-dated deterministic engines — so the rate that applied to a policy on any given day is computed on demand rather than reconstructed from a changelog.
Independent verification around the cat model
We do not replace the vendor model. We put deterministic checks around it : exposure-feed integrity, output-shape verification against prior runs, and a recorded trail of what went in — so model risk is measured instead of assumed.
Bordereaux and cession reconciliation
Cedant and reinsurer views reconciled by deterministic compute that decomposes the difference to its source records, rather than producing a residual that gets carried forward.
Grounded claims-file automation
Maxor Ground between any language model and the claim file : an extracted coverage position or a drafted rationale carries its support from the document, or it does not reach an adjuster.
Audit-ready evidence for the reporting cycle
The inputs, the versions and the run trail attached to each reported figure — so the external audit and the supervisory review read a complete file rather than requesting reconstructions.
The standards we work to.
The compute and its evidence are our scope, so the standards we cite are model-risk, security and reporting-context standards — executed inside the insurer's own governance. The reserve opinion, the capital position and the underwriting decision belong to the appointed actuary and the officers, not to us.
Federal supervisory expectations
Referenced as the governance context for model risk, outsourcing and technology risk at a federally regulated insurer. The institution's own compliance posture remains its own.
Quebec supervisory expectations
The Quebec counterpart, referenced for insurers supervised in the province — including expectations on outsourcing and where data resides.
Insurance contracts — as reporting context
Referenced as the reporting context our compute has to feed : the standard's disclosure demands are exactly the ones that expose an unreproducible derivation.
SR 11-7 / OCC 2011-12 posture
The model-risk-management shape we apply — development, validation and use documented separately, with independent verification treated as a distinct activity rather than a review pass.
Security and confidentiality of what we host
Referenced as the control framing for policyholder and claimant data on our side of the engagement. The insurer's own attestation remains its own.
Our own non-negotiable
Every engine we ship here is pure, bounded and documented : same input, same output, no hidden state. It is the property that turns a reported number back into a re-derivable one, and we treat it as a standard rather than a feature.
One capability line of four. We say so plainly.
Three of our four pillars are lifting and rigging capabilities. An insurer has no use for them, and we do not stretch them to fit — although the physical risks our lift engineering reduces are exactly the ones our clients' policies cover. The software line carries this sector.
Sealed plans + emergency response
Not applicableNot applicable to your operations. There is no crane in an insurance engagement. Worth noting the symmetry, though : the sealed lift plans we produce for heavy industry reduce the very losses underwritten on the other side of this sector.
Distribution + training + implementation
Not applicableNot applicable. CRANEbee® simulates multi-crane operations. No crane, no simulation.
Distribution + advisory + training
Not applicableNot applicable. Murlink® is Dyneema® lifting chain for heavy rigging. Nothing here gets lifted.
Deterministic engineering platform
This is the whole engagement. Deterministic compute on Heisen for reserving, rating, cat-model verification and bordereaux reconciliation ; Maxor Ground as the hallucination firewall on AI-assisted claims work ; Maxor Audit for the signed evidence trail ; Maxor Connect where the operation needs multi-entity project and portfolio operations. Hosted in Canada.
Beyond the four pillars — software built for your operation.
Your triangles, treaties and rating structures are shaped by your book — no off-the-shelf tool holds that shape without a layer of spreadsheets around it. We build the deterministic core instead, in Canada.
Heisen — our deterministic intelligence layer — is optional on any build: embed it or not, your call. Either way it plugs into a fresh custom app or your existing third-party software via API.
Reserving engine with versioned assumptions
Triangles and roll-forward as deterministic compute, every assumption effective-dated, so any prior valuation re-runs to the identical number on demand.
Cession reconciliation that decomposes
Cedant and reinsurer views reconciled to the source record, with the difference explained by line rather than carried as a residual.
Model-risk evidence register
Every model run with its inputs, its version and its independent-verification result in one register — the file a supervisory review asks for, assembled as the work happens rather than afterward.
Make the number re-derivable.
Tell us the computation that has to survive an audit and a supervisor. We will scope the deterministic engine under it — and tell you plainly where our scope stops and the actuary's begins.