Meridian North
PRICE ORACLE
Automated valuation model (AVM) engine
What it is
Meridian North converts rent and transaction evidence into a defensible value range for a property - never a single number presented as certainty. The engine is designed around effective rent - what a tenant actually pays after concessions, computed at the lease level - rather than the asking rent a listing advertises. Comparables are retrieved by learned similarity rather than crude size-and-postcode filters, and every valuation carries a split-conformal prediction interval with statistically valid coverage; when there are not enough genuinely comparable properties, the engine refuses to produce a number rather than guessing. The language model routes each question and writes the explanation, but it never computes a figure - and if its written explanation ever contains a number the engine did not produce, that explanation is discarded and replaced automatically. Sensitive borrower data is never held in the clear: names and documents are sealed with authenticated field-level encryption and Social Security numbers are replaced by one-way tokens, so records can still be matched and de-duplicated without any system - or any operator - decrypting them. Every sensitive action is sealed into a keyed, append-only chain that a database administrator cannot silently re-seal.
What is built, and what is ahead
Proven at the bytes and re-run: 163 automated tests, zero failed, zero skipped; the encryption, tokenization, and ledger mechanisms; the dual numeric core; the refusal paths. Stated plainly rather than implied. The effective-rent lead is the design thesis, demonstrated on modelled regimes and not yet against a real external index. Accuracy figures are asserted on a fixed dataset. Billing runs in fixture mode, and the desktop installer is built but not yet code-signed. The ledger mechanism, while real and tested, has not yet been operated in production.
The numbers behind it
What it does
Monte Carlo Comp Correlation
Stochastic simulation over correlated market factors yields confidence-banded value estimates rather than a single point figure - with rent growth, vacancy, and exit cap moving together the way they do in a real downturn instead of independently.
The AI Never Produces a Number
The language model routes the question and writes the prose; every figure it quotes is computed by a deterministic engine and unit-tested. If an explanation ever contains a number the engine did not produce, that explanation is discarded and replaced automatically - a hallucinated figure cannot reach a user.
Learned Comparable Retrieval
Comparables are retrieved by learned similarity rather than size-and-postcode filters, so the model fits the properties that actually resemble the subject.
Confidence, and a Refusal
Every valuation carries a statistically valid confidence range - and when the evidence is too thin, the engine refuses to produce a number rather than guessing. The scope disclosure travels inside the data itself, so it cannot be stripped by whoever renders it.
Numeric Core Implemented Twice
The valuation mathematics is implemented twice, independently, in two different languages, and continuously proved to agree to the last decimal - excluding a class of silent numerical error by construction.
A Ledger an Administrator Cannot Re-Seal
Even an administrator with full database access cannot rewrite history or quietly delete the most recent records - the ledger detects both, and reports exactly where the break is. Integrity can be re-verified on demand through an authenticated endpoint.
Compliance as Architecture
The engine is built so it structurally cannot emit a recommended rent, a competitor's rent, or any property-identified rent, and it refuses to publish a statistic drawn from too few properties. Sources that forbid extraction are blocked in code, and the type system will not let anyone accidentally build a live connector to one.
A Public Track Record
Every call the engine makes is logged the moment it is made, graded when the horizon arrives, and published - hit rate and calibration score included. Being visibly wrong sometimes is the price of being believed when right.
Who it serves
A value range with a stated confidence, instead of a black-box online estimate.
If there are not enough genuinely comparable sales nearby, the system says so instead of inventing a number - and the answer is labelled an analytical estimate, not an appraisal. A licensed appraiser remains the decision-maker.
A single-point projection becomes a distribution, and loan-officer work becomes a workflow.
Full P10/P50/P90 outcome distributions with the probability of clearing a hurdle, alongside pricing scenarios, income analysis, byte-level document tamper screening, verification-call compliance, and the transcript lifecycle - every figure computed by the deterministic engines.
Model governance is structural: the language model is architecturally barred from producing numbers.
To that, add a keyed, append-only ledger with an out-of-band head anchor; field-level encryption with cryptographically separated purposes; hybrid post-quantum primitives for long-lived material; and deterministic simulation proved identical across two independent implementations.
What it decides, and what you decide
Each gate below is a control enforced in the software itself — not a policy statement about how it should be used.
- Routing a natural-language question to the right engine and extracting its arguments
- Writing the prose explanation - which is discarded automatically if it contains an ungrounded number
- Benchmarking a data source against a published index before its numbers become publishable
- Thin-evidence refusalThe engine, on the operator's behalf
Below the minimum number of genuine comparables the valuation route returns an error instead of a number. There is no 'best guess' path.
- Estimate disclosureOperator, enforced in the payload
Analytical estimate, not a formal appraisal; not a rate lock, not a loan offer, not a credit decision. The disclosure travels inside the response, not in a footer someone can drop.
- Audit-verify gateOperator
The tamper check requires a bearer credential and fails closed when unconfigured - no credential, no access, rather than no credential, no check.
- Role and tenant gateCompany owner / administrator
Four role levels across a multi-tenant workspace, where a role can never manage a role above it, above which sits an operator-only console.
- Publish gateOperator
A data source must clear a benchmark against a published index before its numbers are publishable at all; anything that fails is downgraded to directional, never quietly presented as fact.
- Go-to-market gatesOperator
Live billing keys and the code-signing certificate are named, open operator gates. Nothing is fabricated to look armed.
How it is built
- Monte Carlo comp correlation
- Learned comparable retrieval
- Mortgage-desk workflow
- 69-route API
- Rust to WebAssembly numeric core
- Keyed, truncation-detecting audit chain
- Model-boundary PII redaction
- Hybrid post-quantum cryptography
- Native desktop application
- Field-level envelope encryption (AES-256-GCM)
- One-way tokenization of national identifiers
- Keyed, truncation-detecting audit chain
- PII redaction before any model call
- Role-scoped, multi-tenant desk access
What is running inside it
Every row below is attributed to this system alone and carries the state it is actually in — installed, present in source, or an honest seam.
- Installed
- present and in use in this system's own tree.
- In source
- present in source; not installed, or behind a try/except. Not a shipped capability.
- Seam
- an honest, fail-closed seam. No implementation is wired behind it.
Hybrid post-quantum key encapsulation (X25519 + ML-KEM-768)
InstalledLong-lived material is sealed with a classical and a post-quantum half together, so recorded traffic is not decryptable later by a machine that does not exist yet. These are standardised, published primitives composed in the standard hybrid pattern, not invented algorithms. This is the only system in the ecosystem where post-quantum cryptography is installed and used, and it is claimed nowhere else.
Hybrid post-quantum signature (Ed25519 + ML-DSA-65)
InstalledBoth halves of the signature must verify. A test corrupts the classical half alone and requires the whole verification to fail, so the hybrid is proven to be an AND rather than an OR.
Self-certifying account root
In sourceThe account itself is a hybrid keypair rather than a row that points at one, so the identity and its cryptographic root cannot drift apart.
WebAuthn passkeys bound to platform authenticators
InstalledSign-in is bound to the device's own authenticator, with each challenge consumed exactly once and atomically, so a replayed challenge is not a race to lose.
Shamir 2-of-3 break-glass recovery
InstalledAccount recovery is split across three shares of which any two reconstruct, using a vetted library rather than hand-rolled field arithmetic - and a standing test proves that one share short recovers nothing at all.
Rust numeric core compiled to WebAssembly, proved byte-parity
In sourceThe valuation mathematics is implemented twice in two different languages and continuously proved to agree, which excludes a class of silent numerical error by construction rather than by review.
Correlated-factor Monte Carlo, seeded and reproducible
In sourceRent growth, vacancy and exit cap move together the way they do in a real downturn instead of independently, and the same inputs reproduce the same run exactly.
Split-conformal valuation intervals with a thin-evidence refusal
In sourceEvery valuation carries an interval with statistically valid coverage, and below the minimum number of genuine comparables the engine returns a refusal instead of a number rather than widening the range to cover the gap.
scrypt credential hashing
In sourceCredentials are protected with a memory-hard function, verified by its own test rather than assumed from the library's presence.
Constant-time webhook verification
In sourceIncoming payment webhooks are compared in constant time, so the comparison itself does not leak how nearly a forged signature matched.
Version-lockstep enforcement, drift is a hard failure
In sourceFour files must agree on the version or the build exits non-zero. This is the only place in the ecosystem where version drift is an error rather than a warning, and it is what makes this system's own release state readable from outside.
How to reach it
Meridian North
Releasing 8.31.2026 - join the waiting list for early access.
Visit meridiannorth.aisystemsbymarjerla.com