WaitingForPower — an Energy Project Tracker

Open source, MIT licensed — see LICENSE. Contributions, forks, and issues welcome.

Tracks proposed U.S. energy projects — generation, transmission, storage, LNG, and pipelines, every fuel type — and how long each has been waiting for approval, and why.

The argument is structural, not partisan: solar, wind, storage, gas, nuclear, hydro, LNG, pipelines, and transmission all get stuck in the same handful of bottlenecks. Every tracked delay is mapped to one of seven named cause categories (src/lib/data/causeCategories.ts). The site's policy argument — six specific, bipartisan reform proposals, one per structural bottleneck, each with a stated problem, proposal, strengths, weaknesses, and bill links — lives on a single page, /policies (src/lib/data/policies.ts), deliberately kept separate from the neutral cause-category data so "why a project is stuck" and "what we're arguing should change" aren't the same object.

Quick start

npm install
cp .env.example .env      # set DATABASE_URL to your own Postgres instance
npx prisma migrate deploy # applies the committed migrations to your database
npm run ingest:permitting-dashboard  # populate from a live, no-key-needed source
npm run dev

Needs a real Postgres connection string in DATABASE_URL — there's no bundled local database file. The deployed app and local dev both point at the same hosted Postgres instance for this project (see "Architecture" below for why); for your own fork, any Postgres works (Neon, Supabase, Vercel/Prisma Postgres, RDS, local postgres via Docker, etc.).

Data & sourcing

All project data comes from live, re-runnable sources — no hand-curated one-off research. An earlier version of this project shipped a small hand-researched seed set; it was removed deliberately in favor of sources that stay current on their own, rather than a snapshot that goes stale. See src/lib/ingest/README.md for the full per-source table, open questions, and how each is scheduled:

Every source above — 4 federal/national workbook sources plus the Permitting Dashboard's live API plus all forty state docket modules — runs on Vercel Cron (vercel.json) with no manual step, staggered by 30 minutes so no two sources' runs overlap. Cadence is matched to how often each source actually publishes, not one uniform schedule: the 40 state modules and the Permitting Dashboard run daily, since same-day filings are exactly what the homepage changes feed is built to surface; the 4 sources whose underlying data only republishes monthly, quarterly, or annually (EIA-860M, LBNL Queued Up, ORNL HydroSource, EIA's pipeline tracker) stay weekly, since checking them daily wouldn't catch anything new. This moved off Hobby-plan-driven weekly checks (see src/lib/ingest/README.md for the full cadence history) once the site upgraded to Vercel Pro. Every ingestion run upserts by a stable per-source ID (see the matchKey/manualOverrides.csv identity mechanism in open question #1), so a re-run updates existing projects in place instead of duplicating them.

Every ingested project links back to its public source (see each project's detail page). Where a date or figure wasn't confidently available, the project is marked dateConfidence: "approximate" or carries a dataQualityNote saying so, rather than presenting invented precision as fact. Notably absent so far: sub-250MW generation/storage projects (see each source's capacity floor / scope above — this includes the large majority of hydropower relicensing dockets, which skew small), and any cause-category assignment for automatically-ingested projects (none of the five sources publishes why a project is delayed — see open question #4). This site only tracks projects still waiting on approval — see "RESOLVED_STAGES" in src/lib/ingest/README.md for what's deliberately excluded and why.

Interconnection queue detail. lbnlQueuedUp.ts carries two extra fields for LBNL-sourced projects, added 2026-08-21: interconnectionQueueStage (the workbook's own study-phase label, e.g. "Feasibility Study") and networkUpgradeCostUsd (reserved for a join against LBNL's separate, irregularly-updated interconnection cost-analysis datasets — not yet populated by any module). Per explicit product decision, a suspended interconnection request is treated the same as withdrawn: not shown as "waiting." Every ingestion run now also actively removes a previously-shown project the moment a later edition reports it withdrawn, suspended, or operational, rather than leaving it frozen in its last-known state — see src/lib/ingest/README.md for the full detail and open question #9.

Open questions

Flagged deliberately rather than guessed at — see also src/lib/ingest/README.md for the per-data-source version of this list.

1. Cross-source project identity matching is a real, ongoing problem, not fully solved. EIA, the Permitting Dashboard, and now the growing state-docket series each use their own name/ID for what might be the same physical project. src/lib/ingest/manualOverrides.ts + .csv lets a human declare two source records the same project via a shared matchKey — there's no automated fuzzy-matching, deliberately (name similarity + geographic proximity + capacity similarity is flagged as the highest-value follow-up engineering task, not attempted here). A real merge bug in this mechanism was found and fixed 2026-08-23: upsertNormalizedProject (common.ts) used to look up existing projects by slug, but slug is derived from a project's name, which differs between sources — so two records sharing a manually-declared matchKey still produced two separate rows instead of merging into one, silently defeating the override file's entire purpose (caught by hand when Illinois's new CPCN docket for Grain Belt Express turned out to duplicate an existing Permitting Dashboard entry for the same line). Fixed by adding a real matchKey column to Project (prisma/schema. prisma) and keying the upsert on that instead, with a slug-based fallback lookup so every project ingested before this column existed self-heals (gets its real matchKey backfilled) the next time its own source naturally re-ingests it — no separate backfill migration needed. ProjectSource rows are now upserted by (projectId, label) rather than deleted-and-recreated on every run too, so a merged project keeps both sources' links instead of the more-recently-run source's write wiping out the other's. manualOverrides.csv currently has one real entry (Grain Belt Express: Permitting Dashboard project_id 109441 + Illinois ICC docket 22-0499) — the two other originally-flagged duplicates, SouthCoast Wind and Ocean Wind 1, have no current overlap to merge (see the "NOTE on cross-source identity matching" comment in permittingDashboard.ts), so nothing is declared for them yet.
2. Permitting Dashboard's Socrata dataset is a denormalized join, not one row per project — a single query can return dozens of byte-for-byte duplicate rows per project. The ingestion module dedupes before normalizing; watch for this if the row count from a fresh run ever looks far higher than expected.
3. Permitting Dashboard has no public milestone/timeline or application-filed-date field on the open Socrata dataset this project used — that data likely exists behind the token-gated /api/v1/project/{id} endpoint mentioned in the dashboard's own docs, which wasn't registered for in this pass.
4. EIA-860M, the Permitting Dashboard, ORNL's hydropower relicensing dataset, and EIA's pipeline projects tracker don't publish a cause category. Every project ingested from any of the four ships with causeSlugs: [] and an explicit note that it needs manual/derived assignment, rather than a guessed default. (LBNL Queued Up is the exception — every row it produces is tagged interconnection_queue_backlog, since that's definitionally what an interconnection queue entry is waiting on.)
5. LBNL Queued Up and ORNL hydropower relicensing column names are unverified against a future downloaded workbook — both parsers were written from familiarity with past/current editions of their respective codebooks and fail loudly (naming the missing column) rather than silently misreading a shifted one. Check the current workbook's own codebook/field-descriptions tab before relying on either after a new annual edition ships.
6. Redistribution terms aren't fully confirmed for any source. Federal (.gov) data is generally public domain under 17 U.S.C. §105, consistent with default federal open-data licensing norms, but no dataset-specific terms page was found for data.permits.performance.gov or the EIA API; LBNL's Queued Up asks for citation in a way that reads like an academic norm, not a formal license; and ORNL HydroSource links a Data Use Policy whose exact redistribution terms weren't independently confirmed either. Get an explicit answer per source before redistributing bulk data via this site's own API at scale.
7. Investment-waiting only covers generation/storage projects with MW capacity and a published construction-cost figure. Transmission, pipeline, and LNG projects show "not estimated" rather than a number built on assumptions this project couldn't defend as well — see src/lib/calc/investmentWaiting.ts and /methodology.
8. LBNL's interconnection cost-analysis datasets aren't a single reliably-updated source. Confirmed 2026-08-21: networkUpgradeCostUsd is reserved on the schema but not yet populated by any ingestion module. The underlying data (emp.lbl.gov/interconnection_costs) is six independent per-region publications (MISO, PJM, SPP, ISO-NE, NYISO, non-ISO BAs), not one combined/annually-refreshed file like Queued Up — several editions are years stale (MISO's is from 2021). The join to existing LBNL Queued Up projects (by entity+q_id) does check out — spot-checked against PJM — but coverage against currently-active queue entries will be sparse, since most rows in these cost studies are long-since-operational or withdrawn projects. See src/lib/ingest/README.md.
9. SQLite + serverless deployment (resolved). v1 originally shipped with a committed SQLite file for zero-config local dev. On the first Vercel deploy this broke completely: Next.js's serverless file tracer doesn't know to bundle a file that's only referenced via a connection string (not imported), so prisma/dev.db was silently missing from the deployed function and every DB read 500'd — a strictly worse failure mode than the "writes won't persist" issue originally flagged here. Fixed by moving to a hosted Postgres instance (Prisma Postgres via Vercel's Storage integration) used by both local dev and production.
10. State PUC/PSC dockets: thirty-eight states down, 12 to go, each with its own hard problem. Confirmed 2026-08-24: no national aggregator exists for state utility-commission dockets — each state runs its own system, and FERC eLibrary covers the federal side alone. vaSccDockets.ts, txPuctDockets.ts, coPucDockets.ts, ohOpsbCases.ts, scPscDockets.ts, azAccLineSiting.ts, waEfsecFacilities.ts, nmPrcDockets.ts, ilIccDockets.ts, flPscDockets.ts, nyDpsDockets.ts, nvPucnDockets.ts, orEfscFacilities.ts, maEfsbDockets.ts, okOccDockets.ts, utPscDockets.ts, wiPscDockets.ts, kyPscDockets.ts, moPscDockets.ts, inIurcDockets.ts, njBpuDockets.ts, mdPscDockets.ts, ctCscDockets.ts, wvPscDockets.ts, tnTpucDockets.ts, caCecDockets.ts, nhSecDockets.ts, idPucDockets.ts, nePrbDockets.ts, laPscDockets.ts, alPscDockets.ts, arPscDockets.ts, dePscDockets.ts, meDepSiteLawPermits.ts, riEfsbDockets.ts, vtPucDockets.ts, sdPucDockets.ts, and ndPscDockets.ts are all plain-HTTP-fetch sources, not scraping projects — no headless browser needed for any of them, same shape as this site's other sources — but none was "just add a module":
    • Virginia has a real, structured Status field, but its search scope (caption contains the exact phrase "Certificate of Public Convenience and Necessity") is precise and narrow: only 46 cases in Virginia's entire history match it, and only 1 is currently active.
    • Texas has no status field at all — "still waiting" is inferred from scanning each docket's full filing history for a closing signal (a final order, order on rehearing, or similar), calibrated by hand against real dockets rather than guessed at (see txPuctDockets.ts's header for the specific false-negative this caught and fixed before shipping). In exchange, its yield is far higher — over 100 recent candidates vs. Virginia's single-digit count.
    • Colorado has status (Active/Closed/Effective/Withdrawn/Suspended/Appealed) already as a column in the search results themselves — no per-candidate detail fetch needed just to know whether a docket is still open, unlike either Virginia or Texas.
    • Ohio turned out the simplest yet: its Power Siting Board publishes the entire case history (227 cases) as one unauthenticated JSON request, no search/pagination/session at all, with both status and fuel/project type as real structured fields — the first source in this series where fuel type isn't a keyword guess. The catch was finding it: Ohio's regular PUCO docketing system sits behind a bot-defense WAF that blocks every search regardless of headers, a real dead end confirmed by hand before pivoting to OPSB.
    • South Carolina has no reliable status field either (its "Status" column reads "Open" even on a docket granted years earlier), so "still waiting" is inferred from an embedded Orders sub-table already present on the same detail-page fetch — cheaper than Texas's full filing-history scan since it's one small table, not every filing. Its bigger gotcha was scoping: a server-side exact-phrase search for the CPCN phrase Virginia and Texas both use almost entirely missed SC's real captions, which use a longer statutory name ("Certificate of Environmental Compatibility and Public Convenience and Necessity"); and a broader keyword search let through two petitions that merely argued about a certificate rather than applying for one, caught in a real post-run data-quality check and fixed by requiring captions start with "Application of."
    • Arizona has a real JSON API (unlike VA/TX/SC's HTML), but its docket-level docketStatus field turned out to have the exact same unreliability as South Carolina's Status field — independently rediscovered rather than assumed to carry over: a docket filed in December 2022 still reads docketStatus: "Open" today despite a certificate having been granted five months after filing. The real signal is a separate decisions array (empty = still pending, any entry = a Commission ruling occurred) — confirmed against a real 64-docket batch, where the deceptive "Compliance Due" status (81% of that batch) reliably meant "already granted, now in post-approval compliance monitoring," not "still waiting." Its search endpoint also silently returns zero rows if rowsPerPage is omitted or zero, despite a correct nonzero total count and no error — caught before it could look like "no candidates found."
    • Washington has no CPCN/siting authority in its own utility commission at all — WUTC's ~36,500 dockets are tariffs, rate cases, and affiliated-interest filings, confirmed by hand to contain essentially nothing siting-related. That authority instead sits with a separate body, the Energy Facility Site Evaluation Council (EFSEC), whose entire all-time facility history is only 19 records (RCW 80.50 only reaches major energy facilities), cheap enough to ingest without any date-based lookback at all. Washington is also the one state so far where this series' now-familiar "don't trust the status field" lesson ran backwards: EFSEC's structured status field turned out to be the reliable one, and it was a free-text description paragraph that was caught lying — a facility whose narrative still described it as active and awaiting construction had actually had its site certification terminated four months earlier, correctly reflected only in the structured field.
    • New Mexico has a real JSON API behind an Angular SPA front end (found by capturing the app's own network requests, not guessed), and — unusually for this series — its status field held up under independent testing rather than lying. Its real gotchas were scoping ones instead: its CCN category also covers water/sewer utility certificates (excluded by caption keyword) and, separately, an e-filing intake rejection that never got a real docket number assigned but still appeared in search results as if it were a case.
    • Illinois turned out to be this series' first real cross-source duplicate: its CPCN docket for Grain Belt Express is the same physical interstate transmission line already tracked via the federal Permitting Dashboard. Confirming and fixing that surfaced a genuine bug in manualOverrides.csv's merge mechanism itself (see open question #1) — worth more than the state module in its own right. Illinois's own scoping problem: its CPCN case-type bucket also catches a declaratory-ruling petition and a pure eminent-domain petition, both naturally excluded by requiring the actual CPCN phrase; capacity is published as voltage (kV) rather than MW, a first for this series (Illinois's 1997 generation deregulation means its CPCN docket is now almost entirely a transmission-siting instrument — 58 of 59 real candidates since 2000 are transmission lines, not generation).
    • Florida has no CPCN process at the PSC at all — siting runs through a separate DEP process, and the PSC's "determination of need" is only a small opening sub-docket, confirmed by hand to cover exactly two dockets in the PSC's entire 30,555-docket history (solar is statutorily exempt from mandatory siting certification in Florida unless the developer opts in, so most of Florida's actual solar/storage build-out never touches this process at all). Also the one state where the agency-of-record's own status field is the one caught lying, not a second source: the PSC's own docket for an FPL transmission line shows a 2026-06-01 close date, but DEP's live Applications-in-Process page — the real multi-agency process the PSC docket only opens — still shows filings from 8/18/2026, ten weeks later.
    • New York covers two live siting-certificate tracks in one module since both live in the same underlying system: Article VII (transmission) and Article VIII of the Public Service Law (renewable generation, formerly Executive Law § 94-c until a 2024 state law repealed and replaced it — DPS's own records straddle both names inconsistently, confirmed by hand, so this module queries both). NY publishes no status field at all, and — a real gotcha found only by checking a specific granted case, Alfred Oaks Solar — the renewable track's actual grant order isn't even filed as an Order or Decision document type; it's plain correspondence titled "...Final_Siting_Permit_-_Signed," found only by scanning every filed document's title regardless of type. Real request volume also forced a scheduling tradeoff: a full run took 236s against a 300s cron budget, so its candidate cap was tightened for safety margin, a documented, accepted limitation (see the module header).
    • Nevada has no CPCN either — its equivalent, a Utility Environmental Protection Act (UEPA) permit, spans two entirely separate PUCN systems (a legacy WebForms docket list good only through ~October 2023, plus a modern OnBase JSON API for real status). Its hardest problem was multi-phase transmission reviews: a docket can have a bare "GRANTED" order and still be genuinely active months later because a new phase was filed after it — confirmed against five real GridLiance West dockets that would have been wrongly deleted as resolved under a naive "does any order say GRANTED" check. Fixed by requiring no later substantive filing (excluding routine same-day companion documents like service lists) after the most recent disposition.
    • Oregon has no siting authority in its own utility commission either, same pattern as Washington — the real body is the Energy Facility Siting Council (EFSC). Unusually, both of its own structured status fields turned out unreliable rather than just one: confirmed by cross-checking all 97 tracked facilities against each other, 37 disagreed, and every spot-checked disagreement was resolved correctly by a free-text narrative field and incorrectly by at least one structured field. Also this series' second real cross-state duplicate (after Grain Belt Express): Cascade Renewable Transmission, a 400kV HVDC line crossing the Columbia River, is the same physical project as an existing Washington EFSEC entry — merged via manualOverrides.csv into one row carrying both states' source links.
    • Massachusetts also has a real siting board (EFSB) separate from its DPU, same shape as WA/OR — confirmed by checking rather than assuming, since DPU's own "Siting" docket track mostly turned out to be companion filings to an EFSB docket already covered here. Its own "Closed Date" field can stay null for years after a real certificate grant (a docket kept receiving post-approval compliance filings indefinitely) — resolution is instead inferred from scanning every filed document's own type for a "Final Decision," cross-checked against a case that closed the opposite way (a formal withdrawal notice, no final decision ever filed). A near-miss caught by hand: a docket's most recent filing was "Notice of Withdrawal of Counsel," an attorney leaving the case, not the project being withdrawn — a loose keyword match would have wrongly closed it.
    • Oklahoma has no generic CPCN process at all — its own "CCN" relief type is used almost exclusively by telecom carriers (zero of ~60 real filings sampled were electric). The real electric-siting equivalent is a narrower certificate under the High-Voltage Transmission Line Siting Act, and only 4 cases have ever been filed under it since the state's imaged-document system began in 2022 — all 4 already resolved as of shipping, so this module correctly upserts zero projects today, a real result confirmed by hand, not a bug. Real gotchas found anyway: case numbers aren't unique across docket types (a bare case-number search silently merged two unrelated dockets), and a real case's opening filing was clerked as "Other Document" rather than "Application," missed entirely by the obvious document-type filter.
    • Utah turned out similarly sparse: across the state's entire electric-docket history back to 1987, only 12 dockets are genuine new CPCN applications, and all 12 are already granted — another real, confirmed zero-candidate result (PacifiCorp/Rocky Mountain Power, which owns virtually all Utah retail generation and transmission, gets its resource decisions blessed through periodic Integrated Resource Plan acknowledgment rather than case-by-case CPCN siting). This project has no PDF-parsing dependency, so final orders — unstructured PDFs with no machine-readable grant/deny field — are read by decompressing their own FlateDecode content streams directly with Node's built-in zlib and pulling text out of the raw PDF operators, no new dependency added.
    • Wisconsin shares one docket case-type code across two statutes — the large-facility CPCN (Wis. Stat. § 196.491) and the smaller-facility Certificate of Authority (§ 196.49) — so no separate module or search was needed for each. Its own "Status" field turned out to be a records-retention lifecycle flag, not a case-decision one: two independently-known-decided dockets (one energized since 2023, one operating since shortly after its 2019 grant) both still show "Active" 7+ years later. The real signal is a filed order titled "Final Decision" — confirmed via a real docket whose title has an actual typo, "Signed ad Served," which is why the detection regex matches on "final decision" alone rather than the fuller phrase.
    • Kentucky is the first state in this series where the obvious status signal actually held up: its case-search "Include Closed" filter was checked both directions against real dockets (a years-old still-open case with no PSC action since 2022; a closed case with a real "Final Order Entered" granting a certificate) and used as the primary signal, with a text-based grant/deny scan kept only as a defensive secondary check. Its own scoping problem instead: a broad "Construct" case-type code also covers new headquarters buildings, AMI rollouts, fiber/broadband construction, and cooling-tower retrofits, none of them a generation/transmission project — filtered out by requiring the real construction phrase.
    • Missouri required a real post-shipping fix, caught in this project's own standard verification step, not left to production: an anchored regex for "Order Approving Stipulation and Agreement" missed real title variants with a modifier word inserted ("...Approving Third Stipulation..." / "...Approving Unanimous Stipulation..."), and a second signal — "Closing File" — was filed under both "Order" and "Notice" filing types, but the original resolution check only ever scanned type "Order". Together these left several already-resolved 2018/2019 dockets showing as still waiting until a post-run data-quality check against the live DB caught it, fixed both, and confirmed the fix removed exactly the stale rows and no others.
    • Indiana turned out to have a soft security gap rather than a data gotcha: its public docket-search page shows a Google reCAPTCHA widget, but that widget is checked only in the page's own client-side JS — the real backing search API (a separate companion Azure App Service the portal's JS calls cross-origin) never receives or validates a token, confirmed by posting to it directly. Its "Case Status" field held up under independent checking against filed Final Orders; an "Appealed" case gets its own dedicated stage ("litigation") instead of being deleted like every other resolved status, since the Commission's Final Order already exists but is still being challenged in court.
    • New Jersey has no CPCN process at all; the closest equivalents are two distinct docket types (a 40:55D-19 "reasonably necessary for the public" determination, and a Competitive Solar Incentive Program siting-prohibition waiver) covered by one module. Its own "Case Status" field was found stale by nine years on a real granted docket, so resolution is instead read from the most recent Board Order PDF's own text — decompressed with Node's built-in zlib, since this project has no PDF-parsing dependency and Utah's module (utPscDockets.ts) had already proven the same technique works.
    • Maryland has no case "Status" field at all; "still waiting" is inferred from scanning every filed document for a dispositive Commission/Public Utility Law Judge order, calibrated against a full scan of all 175 real cases (not a sample) after four different heuristics were tried and rejected — real dispositive orders use surprisingly varied phrasing, including some filed with no descriptive subject at all beyond the order number itself. A post-run data-quality check against the live DB (this project's own standard verification step, not left to production) caught a real county-extraction bug — a free-form "capitalized words before COUNTY" regex swept in preceding caption text since these captions are themselves ALL CAPS — fixed with a whitelist of Maryland's 23 real county names, which also caught a genuine source typo ("DORCESTER" for "Dorchester" in one real caption).
    • Connecticut follows the same "real siting authority isn't the obvious one" pattern as Washington, Oregon, and Massachusetts — PURA is only a commenter into the Connecticut Siting Council's own process. CSC has no queryable docket search at all, only hand-typed CMS pages, and its own disclaimer that it may not stay up to date was confirmed true by hand: a petition granted in 2013 was still listed as an open matter in 2026, caught only by cross-checking every candidate against CSC's own historical Decision and Order List. A structural bug — not a parsing bug — was found and fixed during this project's own verification step: the module was silently excluding resolved candidates from its output entirely rather than passing them through with a resolved stage, which meant a project already tracked from a prior run that later resolved would never be revisited or deleted (upsertNormalizedProject in common.ts only deletes a project when it's passed in with a resolved stage — it never diffs "everything previously tracked, minus what showed up this run"). Fixed by pushing every resolved candidate through with currentStage: "cancelled".
    • West Virginia splits its construction-certificate authority across a general CPCN and a separate Siting Certificate for merchant generators, and produced this series' richest real STATUS dataset yet: a confirmed real denial, a case resolved via an ALJ Recommended Decision that auto-finalizes with no separate Commission order, and a confirmed false-positive (an unrelated Pro Hac Vice attorney-admission motion using the word "granted" in the same docket the resolution regex is written to ignore). Also caught two real bugs before shipping: a hybrid gas+solar filing was tagged "solar" because fuel keywords were checked in a fixed declaration order rather than whichever technology is actually named first in the caption; and the same structural "vanished candidate" bug found in Connecticut, here triggered a different way — WV's own case search is scoped Active-only, so a case whose Active flag flips to Closed disappears from every future search before this module's own resolution check ever gets a chance to run on it, rather than the check itself missing it. Fixed by diffing this source's previously-tracked matchKeys (queried directly from the DB) against each run's active-candidate list and pushing a resolved stub for anything that vanished.
    • Tennessee is a genuine, confirmed zero-yield source, not a scraping gap: TVA, a federal instrumentality exempt from TPUC's certificate jurisdiction, supplies the overwhelming majority of the state's generation, and its ~150 local power companies hold exclusive pre-assigned territories under TVA contracts — so a new-entrant electric CCN essentially never triggers. Scanning the entire 160-docket active population by hand (not a sample) found zero currently-open electric generation/transmission/storage CCN candidates; every real CCN-type caption is a water utility expanding service territory or a telecom carrier's franchise application. Kept live anyway as a "standing watch for a rare event" source, the same convention this project already uses for ORNL hydro's own thin population, and given the same preventive "vanished candidate" fix as Connecticut/West Virginia even though nothing exists yet to have gone stale.
    • California required ruling out the obvious agency first: CPUC's own "Proceeding Information Search" is an Oracle APEX app whose real search submit path is client-side Dynamic Action logic invisible to a plain fetch() — confirmed two ways (a raw POST replay that gets redirected but never actually persists the search terms, and a real Chromium session where the search click never reached the actual postback endpoint at all, losing the typed value both times). No CAPTCHA, no login wall — just a fragile SPA-only interface with no plain-HTTP path, correctly deferred per this project's standing guidance rather than forced. CEC turned out to be the real gate instead: it has exclusive jurisdiction over ≥50MW thermal/geothermal plants and, since AB 205, ≥50MW solar/wind and ≥200MWh storage, and nearly every large project files there rather than at CPUC. A false positive was caught before shipping — a local air district's own "Notice of Decision," unrelated to CEC's actual siting decision, an earlier version of the resolution regex would have wrongly matched — along with the same "vanished candidate" structural bug already found in Connecticut/West Virginia (CEC's own listing query is scoped to Under Review/Suspended Proceedings status only).
    • New Hampshire is the fourth real confirmed instance of "the real siting authority isn't the obvious utility commission" (after Washington, Oregon, Massachusetts, and Connecticut): RSA 162-H assigns siting authority exclusively to the Site Evaluation Committee, and the PUC's 3 commissioners are only 3 of SEC's 5 statutory members — they cannot alone constitute a quorum. A December 2025 restructuring moved SEC's own docket records onto the PUC's website under an "SEC" prefix, which is exactly what makes it look like "the PUC does siting" at first glance. Caught a real false-positive before shipping: a docket "rejected" as procedurally incomplete reads exactly like a final denial by keyword match ("Application" + "Certificate" + "Rejecting" all present) but isn't one — confirmed against a real docket that continued for 190+ more filings after its own "rejection." Also confirmed and documented a real access-tooling gotcha: a bare curl GET against any nh.gov subdomain returns a hard TLS-fingerprint bot block, but the exact same request via Node's own fetch() (the real runtime this module and Vercel's serverless functions use) returns a clean 200 with no special handling — not a real blocker, just a curl-specific false alarm, documented so a future maintainer doesn't mistake it for one.
    • Idaho is one