How an audit-defensible cost segregation study works.

From evidence to sealed study

The study is a pipeline, and each stage only consumes what the previous stage recorded. That ordering is the defensibility: nothing appears downstream that was not established upstream.

1. 01

   Evidence collection

   The study starts from the property's own documents: closing statement, floor plans, photographs, public records, receipts where they exist. Each document is triaged and stored with its provenance. Nothing downstream can cite a fact that does not trace to one of these sources.

2. 02

   Observation extraction

   Extraction proposes structured observations from the evidence: printed dimensions transcribed from a floor plan, rooms identified in photographs, measurements from public records. AI is used only to transcribe and propose. It never computes a measurement, prices a component, or writes a final value.

3. 03

   Deterministic resolution

   Software, not a model, validates and dedupes the observations and resolves them into a spatial model of the property: rooms, walls, levels, and site elements as explicit 3D geometry with shared vertices and explicit connectivity. Every geometry value is labelled by how it was established: measured, public record, derived, or a governed default. Absent evidence produces a labelled derived value or an explicitly unresolved segment, never a fake measurement.

4. 04

   The sealed reconstruction snapshot

   A resolved build is frozen into an immutable, versioned reconstruction snapshot. Corrections create a new version that points at its parent; nothing is overwritten. Each engine run records exactly which snapshot it consumed, so the model behind any study can be re-opened later, unchanged.

5. 05

   The deterministic engine

   Classification, basis allocation, and the depreciation schedule are computed by a pure, deterministic engine: the same inputs always produce byte-identical output, and that property is enforced in continuous integration by replaying reference studies and comparing hashes. Component classifications carry citations to the underlying tax authority, and the method is calibrated to IRS standards, including the Cost Segregation Audit Techniques Guide.

6. 06

   The hash-sealed Workpaper

   Every engine run is stamped with the engine version and a content hash over its inputs and outputs. A delivered study pins its facts, its logic version, and its result under one composite hash, alongside a reproducibility manifest and the Workpaper, the audit-defense record that retraces the calculation line by line back to the source documents.

Why a 3D model, not a percentage

The reconstruction is real geometry: each structure carries its own vertices, rooms and walls reference shared vertex pools with explicit connectivity, and everything is placed in a common parcel frame. That precision matters for tax, not just for pictures.

• Quantities come from the model. Flooring area, wall runs, fixture counts, and site improvements are computed from the resolved geometry, so component quantities are derived from the property rather than assumed from its square footage.
• Provenance is attached to every value. Each geometry value is labelled measured, public record, derived, or governed default. A derived value is usable and clearly labelled, and it is replaceable the moment better evidence arrives. The system never records a fake measurement.
• Rooms anchor the tax reasoning. Components are classified where they exist. A room-by-room allocation that sums exactly to the allocated basis is a by-product of the model, not a formatting exercise.

What this changes about defensibility

Defensibility comes from method, not aggression. The method is calibrated to IRS standards, including the Cost Segregation Audit Techniques Guide, and it differs from a traditional study in three ways.

• Reproducible, not remembered. A traditional study's reasoning lives in an engineer's notes. Here the full chain, evidence through calculation, is recorded and replayable; determinism is enforced in continuous integration by replaying reference studies and requiring byte-identical output.
• Cited, not asserted. Every component classification carries a citation to the underlying tax authority, and the Workpaper retraces the calculation line by line.
• Sealed, not editable. Snapshots are immutable and versioned, engine runs are stamped with the engine version and a content hash, and a delivered study is pinned under one composite hash. What was delivered is exactly what can be reproduced.

Unlevered prepares and signs the engineering study. The property owner's CPA reviews it, exercises independent professional judgment, and remains the sole tax return preparer.

Common questions

How does an audit-defensible cost segregation study work?

It works by making every number reproducible. Unlevered rebuilds the property from its own evidence into a 3D component model, freezes that model as an immutable snapshot, and computes the study with a deterministic engine that always produces the same output from the same inputs. Each component classification carries a citation to the underlying tax authority, and the finished study is sealed under a content hash with the Workpaper, which traces every dollar back to a source document. An examiner can follow the whole chain without asking for anything the file cannot produce.

What does it mean to reconstruct a property before estimating it?

Instead of applying percentage rules of thumb to a purchase price, the platform first builds a model of the actual property: rooms, dimensions, levels, structures, and site elements, assembled from floor plans, photographs, public records, and receipts. Components are then identified, quantified, and priced against that model. The estimate inherits the property's real facts rather than an industry average.

What is a reconstruction snapshot?

It is the frozen, versioned record of the property model a study was computed from. Once a reconstruction is resolved it becomes immutable; corrections produce a new version that references its parent rather than overwriting history. Every engine run records the exact snapshot it consumed, which is what lets a study be reproduced or examined years later.

Does AI decide the numbers in the study?

No. AI is used to transcribe and propose: reading printed dimensions off a floor plan, identifying rooms in photographs, extracting line items from documents. Deterministic software validates every proposal, performs all geometry and tax math, and material AI-proposed facts require human review before they take effect. The engine that computes classifications, basis, and the depreciation schedule contains no AI at all.

How is this different from a traditional cost segregation study?

Traditional studies range from percentage-based estimates to site visits summarized in a PDF. In both cases the reasoning usually lives in the engineer's head or a spreadsheet that cannot be replayed. Here the reasoning is the system: evidence, observations, resolved geometry, snapshot, and calculation are all recorded, labelled with provenance, and sealed by hash, so the study can be reproduced exactly and defended from its own file.