These stages are distinct and must not be blurred:

inventory       documents source candidates, access paths, provenance expectations,
                format, update posture, and limitations.  No data is pulled.

retrieval       pulls data from a documented source.  Produces raw artifacts.

extraction      transforms retrieved raw artifacts into chain entries
                (ImplementationSignal records, ImplementationFinding interpretations).
                Source-anchored, assumption-anchored.

interpretation  packages extraction into the Housing Element Implementation Risk Brief
                and related interpretive surfaces.

This document is the inventory stage only. An inventory entry says: "this source is reachable, structured like X, with limits Y." It does not say: "this source proves assumption Z."

Source inventory ≠ source sufficiency. Sufficiency is a finding-level question, decided per-assumption during chain extraction.

Controlled vocabulary for v0_status:

pinned                  canonical URL or resource ID verified; structure inspected;
                        usable in v0 at the access level recorded
partial                 access path partially verified; some pattern works, others not yet
candidate               pinned at catalog level; not deep-probed for fields
needs verification      category real; specific source URL or resource ID not pinned
category-level only     recorded as a category without a specific source identified
deferred (Tier-2 case)  not pinned in v0 by default; revisited only for Tier-2 deeper
                        interpretation

Controlled vocabulary for source_classification — what a source is in the inventory model:

raw evidence            primary source carrying the data itself
operational aggregate   curated / aggregated view over raw upstream feeds
companion source        pairs with another canonical source for a chain pair
identity anchor         provides cross-source join keys; not a signal source
enrichment layer        context layer (zoning, parcels, administrative overlays)
meta-source / discovery viewer / dashboard / catalog over raw sources
reference / lookup      controlled-vocabulary / code-lookup datasets
deferred (Tier-2 case)  category-level only; not pinned in v0
candidate               pinned at catalog level; intended classification pending deep probe

source_classification answers what a source is. v0_status answers what is verified about access to it. The two are independent fields.

Each entry below belongs to the housing / planning subset of the Tier-A signal-visibility posture described in source-strategy.md. These seven pinned source categories are the sources exercised in v0; they do not exhaust the cross-system Tier-A posture. Specific dataset choices are working candidates, not v0-canonical doctrine. (A full re-tag of every source's tier field onto the revised visibility / interpretation ladder is a separate follow-on; the pinned content here is unchanged.)

These are navigation / interpretive entry points, not raw Tier A evidence. They surface or aggregate raw sources cataloged above. On their own they do not bear on Assumption → ImplementationSignal → ImplementationFinding interpretation; they help locate the raw sources that do.

Tier B is the dependency-interpretation layer (causal / delivery-risk / cross-system) — gated and not exercised in v0 by default (see source-strategy.md). The candidate sources below illuminate the conditions under which implementation occurs; reading them as dependency interpretation is Tier-B work, triggered by a specific finding, claim, reviewer need, or demo concept.

Tier C sources are selective for Tier-2 deeper interpretation of flagged sites. They are listed at the category level in source-strategy.md. Specific Tier C datasets are not pinned in v0 by default. All Tier C entries carry source_classification: deferred (Tier-2 case) and v0_status: deferred (Tier-2 case).

ParticipatorySignal candidate sources (public comments, meeting transcripts, comment letters, survey summaries) are reserved by default in v0. If activated, representativeness and bias documentation are required per source.

The field set used throughout this inventory:

source_id                       short, lowercase, underscore-delimited; stable;
                                used as canonical short reference
source_name                     human-readable name
source_owner                    publishing organization (and underlying agency if different)
source_category                 category from docs/source-strategy.md
source_classification           raw evidence / operational aggregate / companion source /
                                identity anchor / enrichment layer / meta-source / discovery /
                                reference / lookup / deferred (Tier-2 case) / candidate
tier                            A / B / C
url_or_identifier               canonical URL or DataSF / ArcGIS resource ID
platform                        DataSF/Socrata, ArcGIS, HCD portal, static download, etc.
access_type                     direct download, API, web viewer, etc.
format                          CSV, JSON, GeoJSON, XLSX, PDF, GIS layer, etc.
update_frequency                nightly, quarterly, per-cycle, etc. (when known)
coverage_period                 time range covered by the data (when known)
geographic_scope                SF whole-city, neighborhood, statewide, etc.
primary_keys                    join keys; BLKLOT format variant noted where relevant
relevant_objects                which object-model primitives this source supports
relevant_chain_stage            Assumption / ImplementationSignal / ImplementationFinding /
                                InterventionCandidate
reported_category_relevance     how this source maps to APR / Pipeline / Production reporting
known_limitations               structural limits
data_quality_caveats            data-quality notes (verbatim where dataset publishes them)
provenance_requirements         what to cite when using this source
v0_status                       pinned / partial / candidate / needs verification /
                                category-level only / deferred (Tier-2 case)

Notes on field use:

• source_classification is the role discipline — what the source is in the inventory model. v0_status is the access-verification state. Separate fields, separate questions.
• relevant_chain_stage is restricted to the four chain stages from object-model.md: Assumption / ImplementationSignal / ImplementationFinding / InterventionCandidate. Supporting primitives (Outcome, Constraint, OpportunitySite, etc.) belong in relevant_objects, not relevant_chain_stage.
• Meta-source entries omit fields not applicable to viewers and substitute chain_relevance (descriptive prose) for relevant_chain_stage since meta-sources do not bear on the chain directly.

This field set is documentation, not an executable schema. The inventory is Markdown-first and consistent with data-dictionary.md. Promotion to a runnable manifest (YAML, JSON, CSV) is deferred until a v0 consumer requires it.

The cautions below belong with the inventory itself; they are not commentary about it.

1. Source inventory ≠ source sufficiency. Naming a source in the inventory does not certify that the source carries enough signal to weaken or support a given assumption. Sufficiency is a finding-level question, decided per-assumption during chain extraction.
2. Metadata accessibility ≠ row-level use. A confirmed-accessible metadata endpoint says nothing about whether the rows behind it are clean, current, or interpretable.
3. DataSF / Socrata resource IDs may change. IDs are generally stable but not guaranteed permanent. Cite IDs with retrieval / verification dates.
4. Source dashboards are often JS-rendered. Cite the linked dataset, not the dashboard view, when source-anchoring a signal.
5. PIM is site-level via the viewer; bulk is via the canonical Tier A datasets. Do not treat PIM as a bulk dataset.
6. Appendix B.4 is large and statewide-format. ~99.6% of rows are < 5-unit ADU-strategy parcels. Only ~171 sites have capacity ≥ 20; 80% of those are already Pipeline-flagged. The aggregate ADU strategy means testing the chain at any individual RH-1 site does not really test SF's strategy.
7. SF Pipeline has data-quality caveats. Verbatim metadata: "Significant data and systems improvements are underway; data may change at any time." Any signal sourced from Pipeline must carry this caveat.
8. Public-input sources are deferred / cautious. ParticipatorySignal candidate sources are reserved by default in v0.
9. BLKLOT formats vary. Any cross-source join requires explicit normalization.

Lots with letter suffix (e.g., 0044-002A / 0044002A) require consistent handling.

10. Category-level entry ≠ pinned source. Tier B abstract entries (B.1, B.2, B.4) and Tier C entries are category-level placeholders, not pinned sources. Do not let category-level entries imply more verification than has happened.
11. Pipeline (6jgi-cpb4) is an operational aggregate, not a primary source. Pipeline curates upstream raw feeds: PPTS (Planning Accela: qvu5-m3a2 + y673-d69b) and PTS (DBI: i98e-djp9). For Tier-2 deeper interpretation, dropping back to the raw feeds may surface signal that Pipeline collapses.

This inventory is a working snapshot. Specific dataset choices remain candidate per the posture set in source-strategy.md. Entries will be revised as prototype work tests them. No data has been retrieved as part of this document; retrieval is a separate, later stage.