A fourth pack is running on the Athanor SDK: realestate-income — underwriting for income-producing property, built for the case where the marketed return and the defensible return are not the same number.
It is the first pack we have built where the most important design decision was about what the software refuses to do.
1. The Problem: Headline Yield vs. Real Yield
Buy a rental villa through a broker and you are handed a number. Twenty percent, all-in. It arrives inside a glossy pro-forma, and it is not exactly a lie — it is a number that would be true under a stack of assumptions nobody wrote down.
Underneath it, usually, sits some combination of:
- Occupancy taken at the operator's forecast, not at the operator's history.
- Opex that quietly omits building insurance, common-area electricity, security, waste, site management, or envelope repairs.
- Revenue that never touched the entity you are buying — booked through a management company, a personal account, a related party.
- A lease clock nobody discounted. On leasehold, a yield quoted as if the asset were freehold is arithmetic performed on the wrong asset.
None of this requires fraud. It requires a seller, a spreadsheet, and no adversarial reader. The documents that would settle it — bank statements, tenant ledgers, supplier invoices, the lease itself — are usually available, usually scanned, usually 130 pages, and usually nobody reads them line by line.
That is a pipeline problem.
2. What We Built
realestate-income v1.0.0 takes a property, its deal terms, and its data room, and runs:
An extraction pass. A fact extractor pulls structured PropertyFacts out of the data room — lease term, permits, entity structure, rent roll, opex lines, operator agreements. Scanned and mixed-format PDFs are OCR-normalised at ingest, so a 130-page photographed lease bundle is readable text rather than a wall the pipeline bounces off.
Five deterministic analysts, each with a narrow mandate:
- Legal & Title — leasehold term and remaining clock, permits, ownership structure, title chain.
- Financial Underwriter — NOI, cap rate, yield, and a direct reconciliation against the seller's own forecast.
- Operator — management-company track record, fee alignment, and dual-mandate conflicts where the operator both books the revenue and reports it.
- Physical — construction warranty, permits, insurance adequacy.
- Market Comps — ADR and occupancy against local benchmarks, supply pipeline, exit liquidity.
Then deterministic stages: financial projections, sensitivity, and a rubric layer. The IRR is leasehold-aware and modelled twice — with and without lease extension — because on a depreciating tenure those are different investments, and quoting one number hides which one you are being sold.
A Deal Synthesizer resolves the analysts into one of five positions: Strong Buy, Buy, Hold / Renegotiate, Pass, Walk Away.
The output is a Property Screening Report whose spine is a section called Their Numbers vs. Normalized — the seller's pro-forma and the defensible restatement side by side, with the delta itemised and attributed.
3. The Line We Do Not Cross
Here is the design decision that mattered most.
Athanor does not value property. We are not a licensed valuer, we do not issue an opinion of value, and we are not trying to become a cheaper substitute for one. In every market worth operating in, the opinion of value is signed by a licensed professional who carries the liability for it — MRICS, KJPP, MAPPI, or the local equivalent.
So the pack ships a second audience: the valuer edition, a document called an Underwriting Pack. It contains the asset, the preparer's context, the normalized-vs-seller reconciliation, any independent valuations already on file, the claims register, the data room reviewed, and the open gaps.
It contains no verdict. No score, no grade, no scorecard, no red-flag list, no projection, no sensitivity, no recommendation. Not hidden — absent by composition. The section list for that audience never includes them, and a contract test (ValuerAudienceContractTest) fails the build if one ever appears.
The reason is simple. If you send a valuer an evidence pack with your number already printed on it, you have not obtained an independent opinion — you have obtained a countersignature, and anyone relying on it downstream should discount it accordingly. A fund's auditor certainly will.
So the promise we make to valuers in the invitation — Athanor does not set, suggest, or influence value — is enforced in code rather than in a policy document. The valuer gets the evidence base, assembled and cross-referenced, and forms their own opinion under their own license.
The division of labour we are building toward: we do the document archaeology; the licensed valuer does the valuation.
4. Evidence Discipline
Both editions rest on the same machinery, most of which exists because a pilot broke something first.
- A Claims Register. Every material claim carries its source and status. A figure from a bank statement, a figure from the seller's deck, and a figure the model inferred are three different kinds of thing, and the report says which is which.
- Preparer's Context. A human preparer can attach what they know that documents do not capture. It is rendered into the report for disclosure and never fed to the model — context should inform the reader, not quietly steer the analysis.
- Independent Valuations. Where a formal valuation already exists, it is recorded as its own block and the report reconciles against it rather than competing with it.
- A verified-fact freeze. Once a human has verified the facts, they can be pinned. Re-runs then skip non-deterministic extraction entirely and use the verified set. This exists because a pilot re-run silently drifted a corrected figure back to a wrong one — a manual correction that does not survive the next run is not a correction.
- An Asset Passport. A verdict-free one-pager the owner can share, with token-gated access to the underlying data room, so a counterparty can inspect the evidence without being handed a conclusion.
5. Where It Runs Today
First pilots are running in Bali and Estonia — two markets we picked because they stress opposite parts of the same pipeline. Indonesian leasehold puts the tenure clock, permit chain, and PMA structure under load. Estonia puts institutional reporting standards and international LP expectations under load. An underwriting pipeline that survives both is one we would trust in a third market.
The pack is not self-serve yet. It runs on real deals with pilot partners, and we are onboarding funds, valuers, and operators by conversation — deliberately, while the calibration is still improving on live deal flow.
6. What's Next
High level, roughly in order:
- A valuer panel. Recruiting licensed, independent valuers market by market, so a screening can hand off cleanly into a signed opinion of value where the transaction needs one. This is the piece we are most actively working on, and the one where we most want to hear from practitioners.
- Re-underwriting on a cadence. A property is not a one-time judgment. Assets that were underwritten once should be restated periodically against fresh operating data, so a fund can see the delta rather than a stale PDF.
- Verified data feeds. Extending Athanor's Oracle layer — independently verified metrics that outrank self-reported ones in the evidence chain — into property operating data.
- More asset classes. v1 is deliberately narrow: leasehold, managed-rental, income-producing. The shape generalises to other income property; the benchmarks and the rubric do not, and we would rather ship one honest vertical than four vague ones.
7. Who We Want to Hear From
If you run a fund or a family office holding or acquiring income property — we would like to run the pack against a deal you have already underwritten yourself. That is the useful test: not whether it produces a plausible report, but whether it finds what your own diligence found, and whether it flags anything you missed.
If you are a licensed valuer — MRICS, KJPP, MAPPI, or your local equivalent — we want you on the panel, and we want you to tear into the Underwriting Pack before you agree to anything. We will send you a full specimen pack built on a synthetic asset, together with the source documents behind it, so you can retrace every figure by hand and tell us where the evidence base is thin. Independence is the product; if the pack compromises yours, it has failed.
If you operate or broker rental property and think your numbers hold up under this kind of reading — that is worth demonstrating rather than asserting, and a normalized reconciliation is a stronger sales document than a pro-forma.
Write to [email protected], or browse the pack in the pipeline catalog.
We built four packs on the same SDK now — startups, Roblox studios, M&A targets, and income property. Different experts, different rubrics, different failure modes. Same manifest, same stages, same evidence chain.
If you hold deep domain knowledge in something that needs assessing at scale, tell us about it. That conversation is how this pack started.
Questions? Find us on Discord or contact support.
