# Gemral Edge — Data Due Diligence Questionnaire (DDQ)

**Version:** 1.7 · **Last reviewed:** 20 October 2026 · **Contact:** support@gemral.com

This questionnaire answers the questions a fund's data, compliance and legal teams usually ask before licensing an alternative dataset. It covers the Gemral Edge datasets delivered through the REST API v1 and the full-history bulk delivery under the Internal Use License (https://www.gemral.com/edge-api-license.md, section 8). The physical infrastructure dataset for the AI-infrastructure universe (data-center sites, power plants, grid load, equipment imports and satellite measurements) is covered in section 13. Where an answer depends on a signed agreement, the order form governs.

**Which license applies to a fund.** A fund or investment business that uses the data for its own research, models or trading decisions needs the Internal Use License. The standard API plan is a license to display the data inside the subscriber's own products, and through it an API key reads only the most recent 24 months of history.

Figures that change daily (row counts, coverage by year, freshness) are not copied here. They are published live at https://www.gemral.com/edge/data-coverage and in the machine-readable data dictionary shipped with every bulk delivery.

---

## 1. Company and contacts

| Question | Answer |
|---|---|
| Legal entity | Stated in the order form and on our vendor tax form (available on request). |
| Product | Gemral Edge: an entity-resolved, point-in-time record of US public-sector disclosures linked to companies, people and events. |
| Commercial contact | support@gemral.com |
| Compliance and data questions | support@gemral.com (subject line "DDQ") — usually answered within one business day. |
| Security contact | support@gemral.com (subject line "Security") |

## 2. Data provenance

**2.1 What are the underlying sources?**
Public records, described by category:

- U.S. congressional financial disclosure filings.
- U.S. federal procurement and contract-award public records.
- U.S. federal arms-transfer public notices.
- U.S. securities-regulator public filings (company reports and insider-ownership reports).
- U.S. federal regulatory and legislative publications.
- U.S. federal lobbying disclosure filings and U.S. federal campaign-finance disclosure filings.
- U.S. market-regulator public trading statistics.
- For the physical infrastructure dataset (section 13): U.S. federal environmental compliance records, U.S. state environmental permit records, U.S. county environmental permit records, U.S. federal energy and electricity statistics, U.S. federal infrastructure reference data, U.S. federal international trade statistics, open map data and public satellite imagery.

The source category behind each field is listed in the data dictionary delivered with the data. Rows in the bulk files carry a stable record reference (columns ending in `_ref`, for example `source_ref`) instead of a link to the source document.

**2.1a Do you name the specific agencies, portals or documents behind each record?**
Not in the product or in our general documentation. As is common practice for alternative-data vendors, sources are described by category. A fund's compliance team that has signed an NDA can request the detailed source list, and any bulk-file row can be traced to its source document through its record reference (for API rows, through the record `id`).

**2.2 What does Gemral Edge add?**
Collection, parsing, deduplication, entity resolution (filing → company → ticker → person, as of the filing date), event classification, scores, convergence signals, forecasts and a public forecast-resolution ledger. The licensed product is this added layer together with the normalised records.

**2.3 Are there any composite or proprietary signals?**
Yes. Some signals are composites built from several of the inputs above. Their methodology is documented at a level that lets you reproduce the event timing; the exact weighting of proprietary composites is disclosed only under NDA. No composite uses non-public information (see section 4).

**2.4 Do you use web scraping?**
We retrieve public records from official publication endpoints and bulk files. Part of the congressional trading history was loaded from publicly available compilations of those same public filings; those rows are labelled with the same source category and can be identified on request under NDA. We do not log in to, or circumvent access controls on, any website, and we do not collect data from social media accounts or private communications.

## 3. Legal right to license

**3.1 Who owns the underlying data?**
Nobody owns US federal government public records; they are published for public use. Our rights cover the added layer described in 2.2.

**3.2 Are there third-party licenses inside the product?**
The disclosure datasets are derived only from public records (government bodies and, for short-sale volume statistics, a U.S. market regulator). The physical infrastructure dataset also uses open community map data and public satellite imagery; their licenses and how we use them are described in section 13.3. Any field that depends on third-party market data (for example, price-based outcome measures) is identified in the data dictionary, and the relevant vendor terms are disclosed on request.

**3.3 Can we use the data for trading?**
Yes, under the Internal Use License: research, model building, backtests, portfolio construction, risk and live trading decisions inside your organisation. Redistribution of raw data outside your organisation is not permitted. Publishing aggregated derived results is permitted.

## 4. MNPI and PII

**4.1 Does the data contain material non-public information?**
No. Every record originates from a document or dataset that a government body or a U.S. market regulator has already made public, or from open map data and public satellite imagery. We do not buy, solicit or accept non-public information, and we do not obtain information from company insiders, government employees or anyone under a duty of confidentiality.

**4.2 How do you make sure nothing becomes "early"?**
Each record carries the time the source made it public (`source_published_at`). Records are never shown before that time in point-in-time views. See section 5.

**4.3 Does the data contain personal information?**
It contains names of public officials, filers, registrants and corporate insiders exactly as they appear in public filings, together with the transactions and roles reported in those filings. We do not add contact details, home addresses, government identification numbers or other personal data that is not in the public filing. We hold no data about private individuals beyond what a public filing names.

**4.4 Data subject requests**
Requests to correct a record that does not match its public source are handled at support@gemral.com. Because records mirror public filings, we correct mismatches against the source rather than delete accurate public records.

## 5. Point-in-time integrity

**5.1 Which timestamps does each record carry?**

| Field | Meaning |
|---|---|
| `source_published_at` | When the source made the record public (filing, posting or announcement time). This is the field to use for "what could the market know at time t". |
| `first_seen_at` | When Gemral Edge first stored the record. |
| `is_backfill` | `true` when the record was loaded historically, after the fact, rather than captured live. |
| `record_hash` | Hash of the record content, used to detect later changes. |

These fields exist on the source tables for congressional trades, federal contracts, pre-award notices, insider trades, lobbying filings and regulatory events. Coverage of other tables is listed in the data dictionary.

**5.2 What happens when a source amends or withdraws a record?**
Amendments do not overwrite history silently. When the content of a record changes, the previous version is written to a revision log (`edge_source_revisions`) with its hash and the time we observed the change, including records that the source later withdrew. Point-in-time queries can therefore reconstruct what was known at any past date.

**5.3 How do you avoid look-ahead in entity matching?**
Filings are matched to companies and tickers using the identifiers and corporate relationships valid on the filing date (renamed companies, ticker changes, parent/subsidiary links), not today's mapping.

**5.4 How do you avoid look-ahead in scores and backtests?**
Scores and historical backtests are computed only from records whose `source_published_at` is on or before the evaluation time. Backfilled history is flagged (`is_backfill`) so you can exclude it or test it separately.

**5.5 Limitations you should know about**
- Historical records loaded by backfill have `first_seen_at` later than `source_published_at`. For those rows, `source_published_at` is taken from the source document, not from our capture time.
- The revision log started in October 2026. Amendments that happened before then are visible only as the current version of the record.

## 6. Survivorship bias

**6.1 Do you keep companies that were delisted, merged or renamed?**
Yes. Disclosure records are kept for every company regardless of its current listing status, and entity resolution maps old tickers to the correct historical company.

**6.2 Is any part of the history affected by survivorship bias?**
Yes, and we state it plainly: **price-based outcome measures are incomplete for securities delisted after 2018.** As of 16 October 2026, 893 tickers referenced in our disclosure records lack end-of-day prices, most of them delisted after 2018. The disclosure records themselves are complete; what is missing is the price series needed to measure post-event returns for those securities. Rows without prices are flagged in the data dictionary rather than silently dropped, so your own backtests can treat them explicitly. Closing this gap is on our roadmap and will be announced in the changelog.

## 7. Coverage and update schedule

**7.1 Historical depth**
Depth varies by source and is extended over time by backfills. The live start date and row count per dataset are published at https://www.gemral.com/edge/data-coverage and in each bulk delivery's manifest.

**7.2 How often is the data updated?**
New public records are collected throughout each US business day and are available through the API as soon as they are processed. Daily scores are recomputed once per US trading day. Webhook subscribers receive new events within minutes of processing (dispatch every 5 minutes).

**7.3 Change management**
Schema changes are versioned. Breaking changes are announced by email to account owners in advance, and the data dictionary carries the version of every field.

Change on 20 October 2026:
- API v1 and webhooks: the fields `source_url`, `doc_url`, `basis_url`, `filing_url`, `pdf_url`, `source_family`, `source_record_id`, `reason_detail` and `sources` keep their names and positions (including CSV/JSONL column order) but are now always `null`. Parsers that read them keep working.
- Bulk files: link columns are replaced by stable record references (`source_url` → `source_ref`, `url` → `record_ref`, and so on; the data dictionary lists each one), source labels carry the source category names from the data dictionary, and a few tables and columns use category-neutral names (listed in the data dictionary). A record reference stays the same across deliveries for the same source document.

## 8. Delivery

| Channel | Details |
|---|---|
| REST API v1 | HTTPS JSON (CSV/JSONL export on supported endpoints), API key authentication. Under the API plan a key reads the most recent 24 months of history; older requests return `history_limit_exceeded` with the earliest readable date. No history limit under the Internal Use License. Quickstart: https://www.gemral.com/edge-api-quickstart.md |
| Webhooks | HMAC-SHA256 signed event payloads with timestamp, automatic retries. Link and source-label fields are present but always `null` (see 7.3). |
| Full-history bulk files | Yearly Parquet files with a manifest (row counts, sha256 checksums) and the data dictionary, delivered through short-lived signed download links (S3-compatible object storage). Available under the Internal Use License. |
| Data dictionary | Machine-readable description of every table and field, including point-in-time semantics and known gaps. |

## 9. Security

- Data is hosted with a managed cloud database and object storage provider; data is encrypted in transit (TLS) and at rest by the provider.
- API keys are stored hashed and can be revoked by the account owner at any time. Each key has rate limits, export quotas and a history-depth limit enforced on our servers.
- Bulk download links are signed and expire after a short time window.
- Access to production systems is limited to named operators, and secrets are kept in environment configuration, never in source code.
- We do not receive or store client portfolio data. Your queries are logged for rate limiting and billing only.

## 10. Service levels and support

- Target availability: 99.5% per calendar month, without service credits. Live status: https://www.gemral.com/edge/status
- Support by email at support@gemral.com, usually within one business day. There is no 24/7 on-call.
- Credit-backed service levels are available only through a separately signed agreement.

## 11. Trials and evaluation

An approved institutional trial is for evaluation only: internal testing and backtesting, no production use. Trial length is listed at https://www.gemral.com/edge/pricing.

## 12. Licensing and commercial terms

- License text: https://www.gemral.com/edge-api-license.md (section 8 covers the Internal Use License for funds).
- Term: 12 months, renewing for further 12-month terms unless either side gives 30 days' written notice.
- Fees and any renewal price adjustment are set in the order form.
- On termination you keep your derived results and delete raw data within 30 days unless the order form says otherwise.

## 13. Physical infrastructure dataset (AI-infrastructure universe)

This dataset places the physical footprint of AI infrastructure next to the disclosure records on the same point-in-time axis. It covers a universe of 49 US-listed companies: hyperscalers and AI cloud providers, data-center operators, chip, server and networking suppliers, optical suppliers, and power, generation and cooling suppliers. The ticker list ships in each bulk delivery's manifest.

**13.1 What the dataset contains**

| Table | Content | Refresh |
|---|---|---|
| `edge_physical_sites` | Data-center sites and power plants: location, grid area, nearest high-voltage substation (name, voltage, distance), building codes parsed from the site name, and the listed company linked to the site. Duplicate registry records of the same facility are folded into one primary site. | Site registries weekly; power-plant inventory monthly |
| `edge_physical_site_events` | Dated site events: new and modified data-center air permits from U.S. state environmental permit records (Virginia, Georgia and Texas) and U.S. county environmental permit records (Arizona): registration/permit and version number, issue date, and—for Virginia and Georgia permits—permitted backup generator capacity in `value_num` with `value_unit = 'mw_backup_permitted'`; a permit is `permit_new` only at version 1 or with no version number, any higher version is `permit_modification`), federal air-program registrations from U.S. federal environmental compliance records (`air_program_registered`, `is_backfill = true`, covering 79 of 81 primary data-center sites in TX, GA, OH and AZ), and power-plant status changes from U.S. federal energy and electricity statistics (`plant_planned`, `plant_under_construction`, `plant_operating`, `plant_retired`) from February 2021 onward. | Virginia permit page daily; other state and federal registries weekly; power plants monthly |
| `edge_grid_zone_load_daily` | Daily electricity demand per grid sub-region, or per balancing area where it has no sub-regions, since 2019 (U.S. federal energy and electricity statistics). | Daily |
| `edge_trade_hs_district_monthly` | Monthly US import value from U.S. federal international trade statistics by 4-digit HS code and customs district for a watchlist of AI-infrastructure equipment codes (transformers, generators, switchgear, cooling, servers and parts). National totals: importer names are not available. | Monthly |
| `edge_trade_hs_country_monthly` | Monthly US import value from U.S. federal international trade statistics by 6-digit HS code (`hs6`, 17 codes) and origin country (`cty_code`, 201 partner countries plus `'-'` for the all-country total; multi-country group aggregates such as APEC, OECD, USMCA and EURO AREA are excluded) over a rolling 24-month window (17,255 rows from September 2024 through August 2026). Delivered as sector context (`attribution = 'sector_context_not_company'`), not company revenue or purchases. | Monthly |
| `edge_site_satellite_monthly` | Per-site built-up index; bright-roof share (`roof_frac`, `roof_frac_wide`: the share of clear pixels within 500 m and 1,000 m of the site point that are bright and non-vegetated, which is how roofs and concrete look from orbit); and land-surface temperature difference against a vegetated ring around the site. | Monthly, finished periods only |
| `edge_capex_quarterly` | Single-quarter capital expenditure (`capex_usd`) extracted from companies' quarterly and annual financial statements in XBRL form (U.S. securities-regulator public filings, original and amended) for 44 of the 49 universe tickers (2,332 rows), keyed by `(ticker, period_end, filed_at)` so every filing and restatement is kept by its filing date. Included in the bulk sample pack. | As filed |
| `edge_land_dc_buildings` · `edge_land_dc_parcels` · `edge_land_dc_pipeline_areas` | U.S. county land-use and building records about data centers, delivered field-for-field as published: permitted data-center buildings (239 records: building permit number, permit issue date, floor area, build status, parcel number, address), data-center land parcels (139 records: recorded owner entity, project name, permit numbers and floor area per permit, zoning, acreage, parcel polygon) and planned data-center project areas (85 records: application number, status, zoning case, proposed floor area, acreage, area polygon). One row per version of each record, so later corrections by the issuer are kept next to the earlier version. | Re-read weekly; a new row only when a record changes |
| `edge_land_dc_links` | Links from those records to `edge_physical_sites` by distance (`distance_100m`, or `distance_300m_approx` for 100–300 m) and to tickers through the parcel's recorded owner entity (`owner_rule`; buildings inherit the owner of the parcel with the same parcel number, `parcel_pin_owner_rule`). | With each re-read |

Reported annual emissions are collected but are not part of the bulk delivery, because they have no accuracy sample yet (see 13.4). Permit events ship under the permit and site-owner samples. Power-plant status events ship under the power-plant owner sample, which checks the owner and ticker link; the status itself is copied from the source inventory and has no sample of its own.

- **Permitted backup generator capacity (`capacity`, unit `mw_backup_permitted`).**
  - *Definition:* Total nameplate rating of emergency/standby generator sets stated in a facility's air permit (`basis = 'permitted_emergency_generator_rating'`). It is an upper bound on on-site backup power capacity, **not** actual IT load or utility grid draw. Values are stored on permit events with `value_unit = 'mw_backup_permitted'` (kept distinct from `'mw'`, which denotes power-plant generation capacity).
  - *Extraction and aggregation:* Equipment tables are parsed by deterministic regular expressions from each Virginia air-permit document (`raw.parse.generators`, verified by matching a verbatim substring `raw.parse.quote` against the document text; 171 of 171 populated events match verbatim). When a line gives both engine horsepower (`bhp`) and electrical output (`ekW`/`kWe`/`kW`), electrical output is used; `kVA`-only ratings are multiplied by the standard 0.8 power factor (`pf_assumed = 0.8`), and `bhp`-only ratings by $0.7457 \times 0.95$ (`from_engine_hp = true`). Fire pumps, boilers and cooling towers are excluded. Because each permit document lists the cumulative equipment of the entire campus and the state page publishes only the latest permit version per facility, a site's current `permitted_backup_mw` is taken from its latest permit issued on or before `as_of` (never summed across versions), and `permitted_backup_mw_total` sums primary sites where the company's role is `owner` (other roles are reported separately in `by_role`). Incremental capacity `added_mw` (summed over 365 days as `added_mw_12m`) is extracted from the modification's project description, set equal to `total_site_mw` when the permit is version 1, or computed by subtracting the immediately preceding version in the database; where none of these is available, `added_mw` is `null` and excluded from `added_mw_12m` (123 of 171 Virginia permits with MW carry `added_mw`).
  - *Lower bounds and non-emergency units:* Where a permit table has OCR corruption, grouped rating rows or missing unit counts (`count_missing`, `ambiguous_rating`, `grouped_rating_rows`), the parsed sum is flagged `lower_bound = true` (17 Virginia permit documents, 16 of which carry a numeric MW value): true permitted backup capacity is at least the reported figure. Non-emergency (prime or continuous) generator units are recorded separately in `raw.parse.non_emergency_mw` (28 Virginia documents totalling 7,341 MW) and are **not** included in `permitted_backup_mw`.
  - *Current scope:* Virginia air permits: 171 of 194 permit events have `mw_backup_permitted`, covering 163 of 183 primary Virginia data-center sites with a permit and 107 of 116 ticker-linked primary sites (92.2%). Georgia air permits: 10 of 46 permit events (8 data-center sites) have `mw_backup_permitted`, read from the equipment table of the permit narrative document. Of the other 33 narratives with a document, 26 use table layouts the parser does not read yet and 7 state only heat input (MMBtu/hr), so they carry no MW rather than an estimate. Generators that a Georgia narrative schedules for installation after the permit year (19 MW at two sites) are reported separately in `raw.parse.future_mw` and are not counted in the total.
  - *Conflicting figures in a permit:* When a permit's summary sentence gives a different per-unit rating from its equipment table for the same number of units, we keep the table figure and set `mw_status = 'conflict'`, with both figures in `raw.parse.conflict`. Two Georgia permits are marked this way (Vantage Data Centers GA1 and GA2: table 66 × 2.75 MW, summary sentence 66 × 2,500 kW); no Virginia permit is.
  - *Why we do not convert to IT load (MW IT):* Data centers install backup generators with N+1, N+2 or 2N redundancy and size them to cover cooling and auxiliary loads alongside IT racks, so effective IT load is typically 50–80% of permitted backup generator capacity depending on campus design. Rather than apply an arbitrary conversion factor in the API or UI, we report permitted backup generator MW as measured and let funds apply their own engineering assumptions. For independent reference ranges on backup-to-IT ratios and data-center power sizing, see the December 2024 report on data centers by Virginia's state legislative audit commission and the 2024 U.S. data-center energy usage report by a U.S. Department of Energy national laboratory (full citations on request).
- **Sector equipment imports (`trade_context`).**
  - *Definition and attribution:* Derived from `edge_trade_hs_country_monthly` (U.S. federal international trade statistics, 6-digit HS × origin country, 24 months) and tagged `attribution = 'sector_context_not_company'`. These numbers measure aggregate US imports of an equipment category, **not** an individual company's revenue, shipments or purchases; when multiple tickers map to the same HS6 code, each sees the full sector figure without market-share allocation.
  - *Company-to-HS6 mapping and two-layer verification:* In `hs_watchlist`, each link between a ticker and an HS6 code carries `hs_relation = 'supplier'` (companies that manufacture or sell that equipment category) or `hs_relation = 'input_demand'` (hyperscalers, AI clouds and data-center operators that purchase or deploy it) together with a verbatim quote (`basis.quote`) from the company's annual report filed with the U.S. securities regulator. Every link is machine-checked in two layers: L1 verifies the exact quote exists in the filing, and L2 verifies the quote contains product keywords specific to that HS6 code. Across the 49-ticker universe, 38 tickers have at least one verified link (85 links across 17 HS6 codes) and return `available = true`; 11 tickers (`BX`, `DBRG`, `NEE`, `D`, `SO`, `VST`, `CEG`, `NRG`, `SMR`, `OKLO`, `TLN`) have no annual-report sentence naming those specific tariffed equipment categories and return `available = false` with `reason = 'no_company_filing_evidence'`.
  - *Growth metric (`yoy_3m`):* For a ticker's linked HS6 codes, `imports_12m_usd` sums the latest 12 published months, `yoy` compares the latest month with the same month a year earlier, and the headline growth rate `yoy_3m` compares the rolling 3-month sum with the same 3-month window one year earlier ($\sum_{t-2}^{t} \text{USD} / \sum_{t-14}^{t-12} \text{USD} - 1$) to smooth single-month vessel-arrival noise, alongside `top_countries` (top 3 origin countries over 12 months) and `top_districts` (top 3 customs districts over 12 months from the 4-digit HS district table).
- **County land-use and building records (`edge_land_dc_*`, bulk sample pack and `land_records` in the footprint API).**
  - *What is delivered:* every field of every record as the issuing county publishes it, without recalculation. Floor area (`square_feet`, `overall_sq_ft`) is floor area, **not** IT load or MW, and is not converted. Placeholder values used by the source are kept as published (for example `bp_final_dt = 1901-01-01` and `year_built = 1000` on buildings not yet finished). Polygons are delivered in WGS84 GeoJSON; `centroid_lat`/`centroid_lon` are our approximate label points.
  - *Versions:* each row is one version of a source record (`record_key`, with `record_hash` over every field and the geometry). When the issuer edits a record we add a new row and set `superseded_at` on the old one (`superseded_reason = 'changed'`); a record that disappears is closed with `'removed'`. A version holds on `[first_seen_at, superseded_at)`.
  - *Links to tickers:* a parcel is linked to a ticker only when its recorded owner entity matches one of our documented owner rules (an annual-report subsidiary list, an acquisition filing, or the company's own brand in the entity name), from that rule's `valid_from`; 46 of 139 parcels match a rule (8 tickers: `AMZN`, `BX`, `CSQR`, `DLR`, `EQIX`, `MSFT`, `V`, `VRSN`, plus private and non-U.S. owner groups with no ticker). Other owner entities are delivered verbatim without a ticker rather than guessed. In `get_physical_footprint`, `land_records` lists the parcels and buildings held by the company's legal entities and the planned project areas within 300 m of its sites.
- **Quarterly capital expenditure (`edge_capex_quarterly`, bulk sample pack).**
  - *Definition and calculation:* Extracted from XBRL company financial facts in U.S. securities-regulator public filings (`PaymentsToAcquirePropertyPlantAndEquipment`, `PaymentsToAcquireProductiveAssets`, `PaymentsToAcquirePropertyPlantAndEquipmentAndIntangibleAssets`, evaluated per quarter so issuers that changed XBRL tags over time remain continuous). Within each fiscal year and tag, single-quarter capex (`capex_usd`) is taken from the ~90-day period fact (80–105 days) when reported, or derived by subtracting cumulative 3-, 6- or 9-month cash-flow facts from the 6-, 9- or 12-month cumulative total. Every quarterly and annual filing and every amendment is retained as a separate row keyed by `(ticker, period_end, filed_at)` so point-in-time queries use only filings with `filed_at` on or before the evaluation date.

**13.2 Point-in-time fields**

Every physical table carries `source_published_at`, `first_seen_at`, `is_backfill` and `record_hash` with the meanings in section 5.1 (in `edge_capex_quarterly`, `filed_at` is the filing's publication date). Where the source gives no publication time, we set an estimate that can only be late, never early: a grid day is available at 00:00 UTC on the next day, an import month (both district HS4 and country HS6 tables) at month end plus 40 days, a satellite period at the first day of the next period plus 3 days. Each link from a site or power plant to a ticker has a `valid_from` date: the date of the public document (subsidiary list attached to an annual report, permit or press release) that establishes it. The ticker on `edge_physical_sites` is the current owner link, not a point-in-time one. Point-in-time ticker links for site events are built in `edge_resolved_events` under the event family `physical`, and bulk deliveries carry that family for permit events and power-plant status events (not emissions). Each row has the ticker valid on its publication date (`ticker_asof`): a ticker is attached only once its link's `valid_from` has passed, so an event published before the link was established carries no ticker. For power-plant subsidiary rules (`pw-*`, 21 rules across 14 subsidiary families), `valid_from` is anchored to the earliest annual-report subsidiary list naming that subsidiary (or the acquisition closing date where a plant portfolio changed hands, such as Calpine to `CEG` in 2026); for direct plant-inventory links without a subsidiary-list rule, `valid_from` is the date we recorded the owner, which can be later than the date the owner took control. A join from a site event to a ticker through `edge_physical_sites` uses today's owner link and can look ahead when a site changed hands; use `edge_resolved_events` for point-in-time work.

**13.3 Licenses of the inputs**

- Federal, state and county public records (permit records, energy and grid statistics, trade statistics, securities-regulator filings): federal records are not subject to copyright; state and county records are redistributed with attribution.
- U.S. county land-use and building records (`edge_land_dc_*`): local-government public records; the issuer publishes no reuse license for them. They are delivered as published under the Internal Use License. Recorded owners are legal entities; the tables contain no names of private individuals (each load stops if an owner name looks like a person).
- Open map data, licensed under the Open Database License (ODbL) 1.0. We use it only to derive results (substation name, voltage and distance to the site). The map layer itself is not part of any delivery.
- Public satellite imagery: only the derived values per site are delivered, never the imagery. The temperature measure is derived from public-domain U.S. government imagery.

Required attributions (keep them when you publish results that show the substation or satellite fields):

- © OpenStreetMap contributors
- Contains modified Copernicus Sentinel data 2023-2026

**13.4 Accuracy gate and its limits**

The accuracy gate used for the disclosure datasets applies: a record type ships only with a sampled precision of at least 95% on a sample no older than 30 days. Twelve types are sampled for this dataset: site owner, power-plant owner, permit, substation, grid area, grid load, import totals, satellite values, imports by country, quarterly capex, permitted backup generator capacity and county land-use and building records. In the 18 October 2026 sample the first eight scored 100%; in the 19 October 2026 sample the three added types also scored 100% (backup generator capacity: 58 of 58 checkable rows, 2 of 60 could not be checked); in the 20 October 2026 sample the county land-use and building records scored 60 of 60 (20 rows per table: every field and the geometry hash matched a fresh read of the record, and the 17 sampled owner links recomputed to the same ticker). Read that figure with these limits:

- Owner links (site and power plant) are checked against primary documents with a verbatim quote. The other nine types are re-measured by machine against a fresh download of the source using our own rules, so they catch transport errors, drift and changed source data, but not an error in the rules themselves. For quarterly capex and backup generator capacity the re-measurement uses the same extraction code that produced the figures (the capacity check also confirms that every quoted sentence appears verbatim in the permit text), so those two checks catch data errors and changed sources, not a flaw in the extraction logic.
- The county land-use and building check confirms that we copied the issuer's records faithfully and apply the owner rules consistently; it does not confirm that the issuer's records are themselves correct or complete (a parcel layer can lag new permits), and it was taken on the day the records were first loaded.
- Power-plant owners and satellite values are sampled (200 rows each); the three types added on 19 October were sampled at 60 rows each; the other types were checked in full.
- If the sample is not renewed within 30 days, the affected types leave the bulk delivery automatically.

**13.5 Known gaps**

- **Owner coverage.** About 88.3% of primary data-center sites (489 of 554, measured 19 October 2026; 283 linked to a listed ticker and 206 carrying a stated non-listed owner type, up from 450 of 554 after round 3 and 420 of 554 on 18 October) are resolved. The remaining 65 sites are registered to holding entities with no public document naming the parent; they are marked `undisclosed` rather than attributed from press reports. The owner accuracy above covers linked sites only.
- **State permit coverage, federal registration backfill and missing issue dates.**
  - *Virginia (state permit records):* 194 permit events (`51` `permit_new`, `143` `permit_modification`) across 183 primary data-center sites, with full permit document text and emergency-generator MW parsed on 171 events (107 of 116 ticker-linked primary sites; 9 ticker-linked sites lack MW because the document has no equipment table, lacks generator ratings, or has severe scan corruption).
  - *Georgia (state permit records, SIC 7374):* 46 permit events (`25` `permit_new`, `21` `permit_modification`) with permit numbers and issue dates (`2006-10-25` to `2026-08-25`) matched to 25 primary data-center sites from the federal compliance records via federal facility IDs; generator MW from the permit narrative document for 10 of these events across 8 sites (see the capacity scope above).
  - *Arizona (county environmental permit records, facility code 93):* 15 permit events (`6` `permit_new`, `9` `permit_modification`) matched to 15 primary data-center sites from the federal compliance records. Issue dates are extracted from signed final permits in the county document index on 10 of the 15 rows; on the remaining 5 rows where the county document index has no record, `issued_on` is left `null` rather than guessed or back-dated.
  - *Texas (state permit records) and Texas, Georgia, Ohio and Arizona federal air-program layer (federal environmental compliance records):* In Texas, 34 state air-permit events (`11` `permit_new`, `23` `permit_modification`) with issue dates are matched across 8 primary data-center sites after filtering out prior non-data-center occupants at the same address. Across the four states, 79 of 81 primary data-center sites (**TX 12/12, GA 29/29, OH 23/23, AZ 15/17**; 2 Arizona sites have no federal air-program row) carry an `air_program_registered` event (`is_backfill = true`) dated to the facility's earliest federal air-program begin date (most commonly NSPS Subpart IIII or MACT Subpart ZZZZ for stationary internal-combustion engines), without MW or permit version numbers. Because Ohio's state permit document service was unavailable across five checks and Texas's state permit search lacks an industry code filter, Ohio and the remaining Texas sites rely on this federal registration layer.
- **Only the latest permit version is listed.** The state permit page shows one version per facility, the most recent. When that version is higher than 1, the facility already held an earlier permit that the page no longer lists, so we record it as a modification, not as a new permit, and the original issue date is not in the data. Before 19 October 2026 our loader marked the first version it saw as new; 132 of 183 events marked new were modifications, and the 12-month count of new permits fell from 48 to 14 when we corrected it. The correction relabelled the events without changing their dates (`first_seen_at`, `source_published_at`) and is noted on each event in `raw.parse.event_kind_fix`.
- **Current site names on permit pages.** The state permit page shows a facility's current name, which may differ from its name when an older permit was issued (for example, a 2014 permit now listed under a later company name). A permit event can therefore carry a name the market did not see on the issue date. Ticker links still follow `valid_from` (13.2), but treat the site name on older permits as the current name. Closing this needs archived copies of the page and is on our roadmap.
- **Power-plant status history starts in February 2021.** Power-plant status events from U.S. federal energy and electricity statistics (`plant_planned`, `plant_under_construction`, `plant_operating`, `plant_retired`; 25,110 events across 7,765 plants) begin on `2021-02-01` because the January 2021 monthly inventory serves as the baseline comparison month for detecting status transitions.
- **Quarterly capex gaps (`edge_capex_quarterly`).** Five of the 49 universe tickers (`TSM`, `NBIS`, `DLR`, `NEE`, `DBRG`) have no quarterly capex series under the standard US-GAAP cash-flow tags in XBRL filings (for example, foreign private issuers filing under IFRS or REITs/utilities reporting capital additions under custom tags), and two tickers (`D` after `2020-03-31`, `GLW` after `2020-09-30`) stop in 2020 when their filings switched to non-standard extensions.
- **Postcode-level coordinates.** Some registry records give only a postcode centroid. Those sites have `geo_precision = 'zip5'`, and their grid area and substation distance are approximate.
- **Satellite seasonality.** The built-up index swings with the seasons by more than the construction signal and also rises when land is cleared. Compare the same period year over year; a new permit does not always show up as new construction. The value is measured over a 500 m disc around the site point, so buildings placed farther out on a large campus are missed. In arid regions (for example the US Mountain West and Southwest) bare dry ground already scores like built land and bright white roofs can score lower than the desert around them, so the index can stay flat or fall while a campus grows; our hand check of five campuses on 18 October 2026 found the index matched dated imagery at three eastern sites and missed construction at both western sites. The bright-roof share was added for this case: it counts new roofs, not cleared land. Re-measured on the same campuses, it rose with the new buildings at both western sites (Arizona from 3.5% to 17% of the 500 m disc between summer 2023 and summer 2025; Colorado from 0.3% to 1.4% when the roof inside the disc was finished in 2026), stayed flat at an operating campus in Virginia, and did not move when eastern forest was cleared. It has its own limits: grey or dark roofs and steel frames below the brightness cut-off are not counted, so it confirms a finished roof rather than the start of works; the processing that masks clouds sometimes flags large white roofs as cloud, which can only make the share too low, never too high; and winter values are noisier, so compare summer with summer. `roof_frac_wide` uses a 1,000 m disc to reach buildings set back from the site point, at the cost of more neighbouring land.
- **County land-use and building records cover one county.** Other counties do not publish comparable layers yet. 88 primary Feed 2 data-center sites with point coordinates in and around that county are linked by distance (85 of them within 100 m); sites with only a postcode centre are never linked by distance. Distance links are approximate: registry points often sit at a gate or road, and on dense campuses one point can fall within 300 m of several buildings. History starts on 20 October 2026; earlier versions of the records are not available.
- **Grid load is regional.** Public demand data exists per grid area, not per substation or site.