Codebook

This document defines every column in every data file. Read this before doing analysis.

data/moratorium_inventory.csv

One row per moratorium instrument. 533 rows total (v2026.07).

enacted_status values

This is the column most users want for filtering. The values:

• active — moratorium is currently in force (~44 rows)
• extended — moratorium was extended past its original sunset and is currently in force (~44 rows)
• replaced — moratorium has expired and a permanent ordinance has been adopted in its place (~37 rows)
• expired — moratorium has lapsed without a documented replacement (~20 rows)
• rescinded — moratorium was affirmatively repealed before its expiration date (~6 rows)
• pending — moratorium has been proposed but is not yet in force (~35 rows; e.g., a public hearing is scheduled but the vote hasn't happened)

To filter to "moratoria currently in force":

in_force = df[df["enacted_status"].isin(["active", "extended"])]

To filter to enacted-only (current or past):

enacted = df[~df["enacted_status"].eq("pending")]

Notes on the data

• What counts as a moratorium here. The inventory records instruments that pause the acceptance or approval of new development — something that stops a project from being permitted. Three things that look similar are deliberately excluded: a permanent prohibition or zoning ban (not a pause); a failed or tabled proposal that never produced an instrument (nothing was adopted); and a freeze on subsidy eligibility rather than on permitting. On the last: Arizona HB 4168 s.31 froze new applications to the state data-center tax relief program for three years, which stops no construction and is recorded in state_legislation.csv instead. See work/answers/rejected/ for the reasoning on excluded candidates.
• Multi-sector instruments (one ordinance covering data centers + solar + battery storage + wind) are recorded as one row. The sectors column gives a clean multi-label list; the trigger, legal_basis, and outcome text carry the prose.
• Several instruments adopted at once are recorded as separate rows, because the grain of this table is the instrument, not the jurisdiction or the meeting. Shelby County, Iowa adopted four separate resolutions on 2026-03-17 — 2026-14 (solar), 2026-15 (battery storage), 2026-16 (data centers and cryptocurrency mining), and 2026-17 (wind) — and those are four rows sharing a date and a jurisdiction but differing in legal_basis and sectors. Deduplicate on legal_basis, not on jurisdiction alone.
• Extensions of the same instrument (e.g., Resolution X adopts a 6-month moratorium, Resolution X-A extends it 6 more months) are recorded as updates to the original row's current_status and enacted_status, not as new rows.
• Successive new instruments are recorded as separate rows. If a county adopts a moratorium, lets it lift or expire, and then later adopts a new moratorium, each is a separate row. Oliver County, ND has three such phases — three separate moratoria, three rows, distinguished by (Phase 1) / (Phase 2) / (Phase 3) in the jurisdiction field and by the -p1 / -p2 / -p3 suffix on moratorium_id.
• Pending moratoria (proposed but not yet adopted) are intentionally included with enacted_status = pending because the public proposal is itself politically meaningful. Filter on enacted_status if you want enacted-only counts.
• [VERIFY] flags mark fields where we couldn't fully confirm the value from primary sources. Most often this is because the local government hasn't posted minutes or signed-ordinance text online yet.

data/state_legislation.csv

One row per state bill. 438 rows total (v2026.07).

This file tracks state-level legislation in 2025–2026 that relates to moratoria — bills authorizing local moratoria, prohibiting them, preempting them, or imposing state-level moratoria of their own. Most are still pending; a small number have been enacted.

A note on carried_over vs failed_died: whether a bill dies at the end of a calendar year depends on the state's session structure, not the calendar. States running two-year sessions (Ohio, Illinois, New Jersey, California, Pennsylvania, Virginia, Michigan) keep bills alive across the biennium; states running short fixed sessions (South Dakota, Maryland) do not. Getting this distinction right is the single largest source of error in a legislative tracker, so the refresh records the governing session context per state in the audit log under work/audit/.

data/summary_stats.json

A single JSON object with top-level aggregates. Useful for embedding live counts on a page or programmatic comparison across releases.

Top-level keys:

• total_local_moratoria: 533 — total rows in the inventory
• total_state_bills: 438 — total rows in state_legislation.csv
• states_with_moratoria: 42
• states_without_moratoria: 8
• moratoria_with_verify_tags: 180 — rows with at least one [VERIFY] flag remaining in any field (matches has_verify_tags = True in the inventory CSV)
• moratoria_without_verify_tags: 232
• enacted_status_breakdown: {active, extended, replaced, expired, rescinded, pending} — the breakdown of the 533 rows by enacted_status
• sweep_coverage: which states received a systematic month-by-month sweep for the release's window, plus swept_states_with_no_adoptions_in_window. Read this before interpreting a state's absence. A state with no rows in a window may have been searched and found empty, or may never have been searched; the inventory alone cannot tell you which. Note that swept_states_with_no_adoptions_in_window is about the window, not the state's whole history -- Idaho appears there while still carrying a 2025 instrument. Generated from data/sweep_coverage.json
• top_states_by_moratoria: list of [state_name, count] pairs, top 15
• top_states_by_bills: list of [state_name, bill_count] pairs, top 15
• states_by_activity_level: count of states in each closed-vocab activity bucket
• state_details: per-state object with abbreviation, activity_level, local_moratoria_count, state_bills_count, file, and (for special cases like NJ) an activity_notes field that preserves qualifying context

data/structured_extractions.jsonl

One JSON object per line. The file contains 864 lines:

• 526 lines are successful extractions (records with extraction != null and error == null)
• 338 lines are LLM-call errors retained for transparency (records with error != null)

⚠️ Scope mismatch as of v2026.07. This cohort was built from documents collected before 2026-04-28 and covers 211 jurisdictions. The inventory is now 416 rows. The clause-prevalence percentages below therefore describe roughly half the current inventory, not all of it — they were not recomputed in this release because the 50-state document sweep was still running, and extracting mid-sweep would have produced a cohort mixing two snapshots. Do not read these percentages as characterizing the 416-row inventory.

The n=348 cohort referenced in the working paper is the subset of successful extractions with confidence >= 0.4. Use this filter when computing clause-prevalence percentages:

import json
records = [
    json.loads(line) for line in open("data/structured_extractions.jsonl")
    if json.loads(line).get("error") is None
    and json.loads(line)["extraction"]["confidence"] >= 0.4
]
assert len(records) == 348

Each record has fields:

• source_file: relative path to the original document (in research/_originals/ of the private working repository)
• url: original URL the document was downloaded from (where known)
• extraction: the structured object — jurisdiction, state, instrument form, sectors, durations, findings, definitions, common clauses, sector-specific clauses, confidence, narrative summary
• error: null on success; an error message string on failure

The full schema (60+ fields organized into five tiers) is in data/moratorium-extraction.json (JSON Schema format). Records where error is not null should be excluded from analysis.

data/clause_extraction_analysis.json

The pre-computed aggregate analysis from structured_extractions.jsonl filtered at confidence >= 0.4. Used to generate the percentage tables in tables/.

Fields:

• total_records: 348
• min_confidence_filter: 0.4
• confidence: {min, max, mean, median, brackets}
• geographic: state counts, jurisdiction-type counts
• temporal: year breakdown of adoption dates
• legal: instrument forms, authority types, sector counts
• clause_prevalence: detailed breakdowns of every clause type in the 44-clause taxonomy (universal, common, sector-specific, definitional approaches, findings detail levels)

A note on dates

Where possible, dates are ISO-8601 (YYYY-MM-DD). When a day is unknown but month+year are known, we use YYYY-MM. When only the year is known, YYYY. When the date is uncertain (sourced from news rather than minutes), we may write something like 2025-12-17 reported or January 2026 [VERIFY]. Spreadsheet tools may need help parsing these; the examples/ folder shows how.