# Verified Construction Core v0.3 preview

This tracked preview contains **14 physical sites** and
**15 linked projects** selected from `2026-07-22-open-seed-v97`.
Every selected project has a physical-status observation no more than 90 days
old at the 2026-08-20 review date. 3 project rows carry official
parcel or surveyed boundary geometry; this describes the geometry attached to the row, not a claim
that construction occupies the entire parcel. The other 12 use explicitly labelled
site-, parcel-, address-, or first-party campus-location points with their precision limits preserved.
Locality centroids and model-only lifecycle states fail the selector.

This is **not** the final Verified Construction Core v1. It does not change the historical Atlas
`construction_verified=false` field, infer a continuously current state, claim independent imagery
verification, or claim completeness. The final release remains blocked at
14/100 sites,
10/40 countries, and
11/50 non-US sites.

Files:

- `sites.csv` and `projects.csv`: the reviewed physical-site/project cohort.
- `sites.geojson` and `map.html`: matching clean-clone-readable map products.
- `evidence.csv`: status, geometry, operating-model, workload, and typed-metric evidence.
- `schema.json`: machine-readable field, relationship, GeoJSON, and map contract.
- `selection-report.json`: accounting for every source pipeline row and every final gate.
- `manifest.json` and `manifest.sha256`: byte and SHA-256 bindings.

Missing power, annual energy, operating model, workload, and development type remain explicit
unknowns; the builder never converts missing values to zero. Existing satellite reviews remain
non-claiming analyst evidence. The selection report binds each non-default imagery outcome to its
exact tracked review source and preserves locally unsealed identity lineage and unadjudicated
conflicts explicitly. Two exact Canadian geometry overlays were reviewed but excluded: one recent
scene was unusable and one usable comparison showed no filtered recent-change component.

Every workload observation now carries a validator-bound `deployment_scope`; all three current
observations are `intended`, never operational. `role_claims_json` binds each published role to an
evidence ID and relationship scope. Five intended-operator claims pass that gate. Four prior
CDC/AST owner or operator strings lacked role-specific evidence, so v0.3 clears them and preserves
the rejected claims and reasons in the provenance contract rather than laundering status evidence
into role evidence.

The preview validates from a public clean clone. Rebuilding it still requires the locally hydrated
v97 payload, so clean-clone rebuildability is an explicit failed final-release gate rather than an
implied capability.