Zoning Data Factory: Addi Architecture Disclosure

Overview · Modules · Zoning Data Factory

What it is

Ordinances in, canonical records out

The zoning data factory is our in-house pipeline for turning municipal zoning ordinances into canonical, investor-readable jurisdiction records. Ordinances are written by and for city attorneys; investors need to know what a parcel actually permits, in plain words, with an honest statement of how sure we are. The factory closes that gap jurisdiction by jurisdiction and the whole platform reads the result. This is one of the two proprietary data assets named in architecture section 12, and it is an owned, compounding asset, not a licensed feed: every finished jurisdiction stays in the corpus at a quality bar we control.

The obvious alternative is licensing zoning data from a third party. We build the corpus in-house instead because a licensed feed cannot carry our provenance discipline: we would be re-serving someone else's confidence claims without being able to defend them. Building the factory cost us time up front. What it bought is a data asset where every claim traces to an ordinance a human can go read.

The questions it answers

For the investor and the operator

Through the surfaces that consume the corpus:

What does this jurisdiction actually allow on this parcel, stated in the language an investor uses, not the language the ordinance uses?
Which zoning district does this parcel sit in, and how confident are we in that match?
When the record says a rule does not exist here, was that checked, or just not found?

For the reader of this package: how we produce a trustworthy legal-text data asset, and why we believe the quality holds as the corpus grows.

Scope and boundaries

What it owns, and what it does not

The unit owns the path from a municipal ordinance source to a canonical jurisdiction record accepted into the corpus, and it holds that path to the guarantees below: nothing invented, nothing silently missing, independent review, and honest confidence. It ends at the corpus: the analysis engine (architecture section 11) reads canonical records through the warehouse boundary and renders the zoning report from them; the coverage unit reports which jurisdictions are covered, partial, or not yet. The factory never writes into deals, reports, or any tenant's data. It produces reference truth; other modules decide what to do with it.

What it is NOT

Not a licensed or aggregated feed. The corpus is produced in-house from primary sources.
Not a generative summarizer. Canonical records are structured data with quality gates, not prose written by a model about an ordinance.
Not a legal opinion. Records state what the ordinance says and how confident we are in the reading; they do not advise.
Not a per-tenant system. The corpus is reference data shared by the platform; tenancy applies to the deals that consume it, not to the records themselves.

Withheld by design Tier 3

The extraction specification's contents, the gate list, source acquisition methods, how extraction work is organized and scaled, and tooling choices inside the pipeline are withheld by design and held at Tier 3, as is a walkthrough of the pipeline itself against a real jurisdiction.

The contract

The unit's public shape

No fields, no schemas, no payloads; responsibilities, the write door, events, guarantees, and refusals only.

Responsibilities

Produce canonical, investor-readable jurisdiction records from municipal zoning ordinances, against a binding extraction specification.
Hold every candidate record to the ladder of quality gates before it can enter the corpus.
Route every candidate through independent quality review. The extractor that produced a record never approves it.
Match parcels to zoning districts against GIS conservatively, carrying an honest confidence statement with every match.
Maintain the corpus as the platform's single source of zoning truth, read by the analysis engine through the warehouse boundary and reflected in the coverage unit's honest coverage states.

The write door

Like every module, the factory has one write door, and the corpus is written through it or not at all. The door admits a small set of business-named commands, and what it enforces is the point: nothing becomes readable by the platform except a record that passed the quality gates and independent review, and a record may claim an absence only when the absence was actually checked, so downstream readers can distinguish "checked and absent" from "not yet examined."

Withheld by design Tier 2

The command roster is held at Tier 2.

Events announced

Events carry identity plus the business fact, never a data dump.

A jurisdiction record was accepted into the corpus.
A jurisdiction's coverage state changed. The coverage unit consumes this to keep its covered, partial, and not-yet states truthful without reaching into the factory's tables.

Guarantees

Nothing is invented. Every value in an accepted record traces to the source ordinance.
Nothing is silently missing. Every absence in an accepted record was concluded from an actual check, never from silence.
Independent review. Every accepted record passed independent review by a reviewer who did not produce it.
Honest confidence. Every district match states the confidence the evidence supports, in words, and downstream surfaces render it no stronger than stated.
Reference data behind the boundary. The corpus is reference data behind the warehouse boundary: consumers read it through the access module and never through the factory's internals, so the factory can change without breaking a single consumer.

Refusals

Refuses any write that has not passed the full gate ladder. There is no expedited path into the corpus.
Refuses self-approval. An extraction reviewed only by its own producer does not land, regardless of how clean it looks.
Refuses unproven absence. "We did not find it" is not "it does not exist," and the door will not record the second when only the first is true.
Refuses to overstate a district match. Where the GIS evidence is ambiguous, the record says so rather than picking a winner.
Refuses direct table access from any other module, in both directions. Consumers get the corpus through the boundary; the factory writes nothing outside its own ownership.

What done means

The gates a change passes

A jurisdiction is done when an investor can read its canonical record and act on it: what the districts permit, stated in investor language; where the parcel boundaries of confidence sit; and, for every rule the record does not carry, the assurance that we looked. Done is not "we extracted the ordinance." Done is "the record survived every gate, a second set of eyes with no stake in the extraction confirmed it, and what the reader sees says what the record says."

We hold this bar because the failure mode of a zoning data product is quiet: a wrong permitted-use claim does not crash anything, it just misleads an investor months later. The gates exist to make that failure loud at build time instead.

Every candidate record climbs a ladder of quality gates defined by a binding extraction specification. The specification is the contract the extractor works against and the standard the reviewer audits against; neither party improvises. The ladder checks the record at more than one altitude, because we have designed for the case where a record is right in one place and wrong in another; a record that is correct in only one of the places a reader could meet it has not shipped.

Above the mechanical ladder sits independent review: a reviewer who did not produce the extraction audits it. This is our standing rule that the writer is never the auditor, applied to data production. We adopted it because self-review converges on the extractor's own blind spots, and legal text is exactly where blind spots hide.

District matching against GIS is gated on conservatism: the match is recorded at the confidence the spatial evidence supports. Where the evidence is ambiguous, the gate requires the record to say so.

Withheld by design Tier 3

The specification's contents and the concrete gate list are held at Tier 3, as is a walkthrough of the pipeline against a real jurisdiction.

The failing cases

An invented value fails the build. If a value in the candidate record cannot be traced to the source ordinance, the record does not land, whether the value is wrong or happens to be right.
A silent omission fails the build. A record asserting or implying absence that was never actually checked does not land.
A self-reviewed extraction fails the build. If the producer also performed the review, the write is refused regardless of the record's content.
A record that reads differently in different places fails the build. If what a reader sees does not faithfully present the accepted record, the record does not land even though the data may be correct.
An overconfident district match fails the build. A parcel-to-district match asserted above what the GIS evidence supports does not land.

Honest-state behaviors

Not yet covered says so. A jurisdiction we have not processed is not yet covered, and the platform says so. The coverage unit renders covered, partial, and not-yet states from the factory's announced facts; there is no fabricated completeness.
Partial stays partial. A partially processed jurisdiction is presented as partial, never rounded up to covered.
Confidence travels with the record in words, and downstream surfaces render it no stronger than the record states, in line with the same platform-wide honesty doctrine that keeps coarse-geography data from being presented at finer precision than it has.
Absence is concluded, not assumed. When the record says a rule is absent, the reader can see that the absence was concluded from search, not from silence.

Where it connects

The spine sections this unit realizes

Section 12, proprietary data assets: the corpus as a concept and the quality-gate story.
Section 11, the compute engine and the warehouse seam: how the analysis engine reaches canonical records, and why producers never touch storage directly.
Section 3, the failure model: unmapped jurisdictions and wrong or stale sources are named failure modes; this unit and the coverage unit are the answer to them.
Section 9, the evidence chain: the zoning report rendered from the corpus enters the chain as a versioned artifact like any other.