Glossary

Plain-English definitions.

Every term used across the model, the architecture pages, and the agent fleet — the one place each is defined. Search, or link straight to a term with its # anchor.

Capital markets & sourcing

Take-out #
Capital that replaces an earlier piece of financing — e.g. a permanent loan that “takes out” (pays off) the construction loan once the building is built and leased. A “take-out buyer” buys the finished asset.
Off-taker #
The party that ultimately takes the project off your hands at a given stage — the buyer of the stabilized building or its long-term loan. A “pension-fund take-out” means a pension fund is that off-taker.
ESG mandate #
A rule requiring an institution to invest only in assets meeting environmental, social and governance criteria. If a deal can’t prove it qualifies, that capital legally can’t buy it.
LIHTC equity #
Low-Income Housing Tax Credit equity — federal tax credits sold to investors that bring cash into a deal in exchange for setting aside affordable units.
Basis (LIHTC) #
The eligible cost on which LIHTC credits are calculated. “Eligible basis” is the portion of development cost that qualifies for the credit, so how you spend money affects how much credit equity you raise.
CRE CLO #
Commercial Real Estate Collateralized Loan Obligation — a vehicle that bundles transitional / bridge property loans. A common lease-up-stage lender; typically higher-leverage and floating-rate.
Agency / HUD debt #
Government-backed permanent mortgages (Fannie Mae, Freddie Mac, HUD). Long amortization, non-recourse, low-cost — but sized strictly on debt coverage, and HUD carries prevailing-wage rules.
Builders-risk #
Insurance covering a building while it is under construction (fire, theft, weather damage). A construction-stage cost that the capital stack has to fund and time.
OCIP / CCIP #
“Wrap-up” insurance programs — one master policy covering everyone on a job site. OCIP is owner-controlled, CCIP is contractor-controlled. Affects cost and who controls claims.
Pension-fund general account #
The main investment pool of a pension fund — patient, long-hold capital that favors large, stable, income-producing assets, increasingly with an ESG mandate.
Insurance-company general account #
An insurer’s core investment pool backing its policies — similar appetite to a pension fund: large, safe, long-duration income, and a major source of long-term real-estate debt.
Stabilized #
The point where a building has leased up to its normal occupancy and income has settled — the moment most long-term buyers and lenders underwrite to.
Lease-up #
The period after opening when a building fills from empty to stabilized occupancy.
Refi #
Refinancing — replacing one loan with another, usually a cheaper long-term loan once the building is stabilized.
Exit #
The end of the hold — selling the asset (or its loan) to realize the return.
Sponsor exposure #
How much the developer (“sponsor”) has at risk in the deal. Construction lenders watch top-of-stack sponsor equity closely.
Ticket size #
The dollar amount a given investor wants to deploy per deal. Some capital won’t look below (or above) a certain ticket.
Asset-size threshold #
The minimum (or maximum) deal value a capital source will consider — many institutional buyers won’t touch anything below, say, ~$50M.
Dealbreaker #
A condition that disqualifies a deal for a given capital source no matter how good the rest is — e.g. no affordability for a LIHTC investor, or an uninsurable site for a pension fund.
Deal-mode toggle #
A switch (single asset / portfolio / district) that tells the model how many assets to underwrite and whether to show a roll-up — so the same engine serves a single building and a whole district.
District economics #
The cross-cutting economics of a multi-asset, neighborhood-scale project — shared infrastructure, phasing, cross-collateralization, a district-scale capital source — that only appear when more than one asset is in play.
Single-asset deal #
One building underwritten on its own; the default case (N=1).
Portfolio deal #
Several assets underwritten together and rolled up — a district is the portfolio case with shared economics.
Backwards-propagation of expectations #
Planning a project backward from the capital that will eventually own it: if you want a pension-fund take-out, you bake in the asset size, ESG data and cash-flow shape it needs at the very start, before the design locks.
Stage alignment #
Making sure the project’s metrics and decisions at each stage (pre-dev, construction, lease-up, stabilized, refi, exit) match what the capital that owns that stage expects.

Climate & insurance

Climate insurability #
Whether a property can actually be insured against climate risk (flood, wildfire, wind) at a price the deal can carry, for the whole hold. Increasingly a yes/no deal-gate, not just a cost line.
Carrier appetite #
How willing insurance carriers are to write coverage in a given submarket. In some markets it is shrinking toward zero, which can strand an otherwise-good deal.
Parametric coverage #
Insurance that pays a fixed amount when a measured trigger is hit (e.g. wind speed over a threshold) rather than reimbursing assessed losses — sometimes the only coverage available in high-risk markets.

Engine & underwriting concepts

Cascade #
The knock-on chain when one input moves — raise efficiency → more rentable area → more rent → higher value. The model traces it automatically.
Orphaned lever #
An input the engine supports but no agent proposes a value for, so it just sits at its default until a human moves it.
Density bonus #
Extra units or height a city grants in exchange for something (often affordable units) — a key way an entitlements move can unlock the affordable → tax-credit financing path.
Exit cap rate #
The yield a future buyer demands; value ≈ annual income ÷ cap rate, so a lower cap means a higher sale price. Often the single biggest driver of the return.
DSCR floor #
Debt-Service-Coverage-Ratio floor — the minimum ratio of operating income to loan payments a lender requires (1.20 conventional, 1.176 HUD here). Below it, the loan sizes down.
Facility cap #
The maximum loan size the model allows (~$95M here) — a hard constraint the agents cannot exceed.
Validator #
A hard rule the engine enforces (the DSCR floor, the facility cap). A move that breaks one is penalized and can’t count as a “go.”
Lead Agent composer #
The orchestrator that sequences the agents — reads the binding constraint, asks agents for moves, and composes the cheapest path that pencils.
Two-tier orchestrator #
The planned shape with the ten-agent fleet: cross-cutting agents set constraints / objectives, lever-owning agents propose moves within them, and the Lead Agent composes.
Cross-cutting agent #
An agent that doesn’t own sliders; it sets targets / constraints the other agents’ moves must satisfy (e.g. Capital Sourcing, ESG, Insurance).
Lever-owning agent #
An agent that proposes actual input values — it moves sliders (e.g. Program, Cost, Revenue). After the steel-man pass these are called “composers.”
Composer agent #
An agent that proposes actual engine inputs — it moves sliders: Program, Cost, Revenue, Operations, Capital Stack, Risk & Returns. The “real agents” in the propose-a-number sense.
Policy agent #
An agent that sets constraints or objectives the composers must work within, rather than moving sliders itself (Capital Sourcing, ESG). A policy, wearing a persona for the user.
Gate agent #
An agent that is a binary checkpoint — a yes/no deal-killer (Insurance: insurable or not). A validator with judgment, not a move-proposer.

Engineering — how it’s built

Transpiler #
A program that translates code from one language to another — here, the small tool that reads the Python engine and writes out the equivalent JavaScript automatically.
AST walker #
A transpiler that reads code by its structure (its Abstract Syntax Tree) rather than as plain text, so the translation is reliable rather than find-and-replace.
Fuzz test #
A test that throws thousands of random inputs at the system to catch edge cases that hand-picked examples would miss.
Banker’s rounding #
Rounding a value that is exactly halfway to the nearest even number (2.5 → 2, 3.5 → 4) instead of always up. The standard accounting convention; what the engine now uses.
Frozen oracle #
The old hand-written engine, kept untouched as a reference the tests check against — like keeping the original signed contract in a drawer. It doesn’t ship.
Generator-fidelity test #
The test proving the generated JavaScript matches the Python engine number-for-number, over the scenarios plus thousands of random inputs.
Parity tolerance #
How close two computed numbers must be to count as equal (here one part in a million) — needed because different platforms’ math can differ in the last digit.
Snapshot test #
A test that freezes today’s known-good outputs and flags if a change moves any number, so nothing drifts silently.
Idempotent #
An operation you can run repeatedly with the same result — e.g. a build script that regenerates a section cleanly instead of duplicating it.
Steel-man #
Arguing the strongest possible case for a position before judging it — here, making the best case both for keeping the agent fleet and for tearing it down, then acting on what survived.
No terms match that search.