NEPA Decarbonization Technology Analysis: Deliverable 4
Review Timelines for Categorical Exclusions, Environmental Assessments, and Environmental Impact Statements
Executive Summary
Key Findings
- Review duration climbs steeply from days to months to years. Median review duration is 20 days for CE, 116 days (~4 months) for EA, and 1,008 days (~2.8 years) for EIS — the expected complexity ladder, now quantified across the full database (complete timelines; Figure 5).
- Timeline completeness varies sharply by process, and the gaps are structural. Complete timelines (both an initiation and a decision date) reach 55% for CE, 59% for EA, and 36% for EIS (Figure 3). CE and EA are initiation-limited (no Notice-of-Intent requirement, often no recorded start); EIS is decision-limited — Records of Decision are usually separate documents outside the corpus, so EIS decision coverage is just 42.2%. LLM adjudication of the ambiguous cases lifted complete coverage roughly +7 points per process.
- Post-FRA, EIS documents got shorter while EAs held steady. Across all 3,678 EA/EIS projects (all energy types, regulatory pages from the full text), mean EIS length fell from 317 to 269 pages after the Fiscal Responsibility Act, while EA length stayed near 33–39 pages — though the post-FRA window is still small (Section 3).
- The longest reviews are prime case-study candidates. Genuinely protracted reviews — SunZia Southwest Transmission (14.6 yr), Energia Sierra Juarez, Grain Belt Express, Cushman Hydroelectric — are surfaced for CATF investigation and separated from extraction errors (an “initiation” anchored to a historical citation), which are excluded (Section 2.4).
- BLM field offices do not measurably speed up with experience. The apparent office “learning curve” disappears once calendar time is controlled: the speed-up is a corpus-wide secular decline in CE durations, not individual offices getting faster as their caseload grows (Section 5).
This report delivers:
Timelines for categorical exclusions, environmental assessments, and environmental impact statements — including segmentation by regulatory period (pre/post-FRA), energy category, and decision year — plus a timeline-outlier deliverable for case-study investigation and a document-length / FRA page-limit analysis refreshed on Phase 2 decision dates.
Methodology
This deliverable extends the Phase 1 timeline analysis — which examined review timelines for decarb projects only — to the entire NEPATEC 2.0 database, which includes decarb, fossil, and other project types. It does so with a rebuilt extraction pipeline that adds authoritative agency dates, section aware retrieval, and two machine-learning models.
How timeline dates are extracted
This section improves on the Phase 1 timeline method. For every project we identify two milestones — the initiation (when the review begins: an application received, a Notice of Intent, or scoping) and the decision (when it concludes: a signature, a Record of Decision, or a Finding of No Significant Impact). The pipeline finds these in five high-level steps:
- Extract candidate dates from every NEPA document in NEPATEC 2.0, capturing the sentence around each date as well as document location and metadata so its meaning is clear.
- Classify each date’s role — a machine-learning classifier labels every candidate as an
initiation,decision, orneither. - Rank the candidates — a second ranking model ranks initiation and decision candidates by their likely to be the final, decisive date for the project.
- Anchor on authoritative dates — where the BLM or DOE NEPA register records an official project-start or decision date, that date is used.
- Resolve the hardest cases with an LLM — for the projects where the operative date is still ambiguous, a large language model selects the best remaining candidate.
A project’s timeline is complete when both an initiation and a decision date are found.
How Phase 2 analysis improves upon Phase 1
- Full-database coverage — Phase 1 covered decarbonization technology projects only (20,725); Phase 2 covers the entire NEPATEC 2.0 database (61,881 projects).
- API registry dates — Phase 2 incorporates authoritative dates from BLM and DOE NEPA registers as an authoritative source of official project-start and decision dates. We pull in about 40,000 register dates, which is particularly helpful for CE and EA reviews.
- Cleaner retrieval and granularity handling — section-aware retrieval pulls fewer but better candidate dates based on document location and the pipeline uses year-month decision dates (e.g. a Final-EIS cover month) where no better dates are available.
- Improved classifier AND a ranking model — Phase 1 used a single date classifier (
DeBERTamodel); Phase 2 used an improve classifer (SetFit) and adds a dedicated ranking model (LightGBM) that decides which candidate is the operative initiation/decision, sharply improving which candidates get sent to the LLM adjudicating model.
Data Quality & Caveats
Fixable pipeline or data quality issues
- About 2,000 register date errors. About 2,000 document text dates are erroneously used in place of API register dates. These can be recovered with a little bit of coding elbow grease.
- Negative-duration stopgap. 233 decision-before-initiation rows are reclassified to
invalid_orderat analysis time; a source-level fix in script 05 is tracked. - Outlier extraction errors. — extra long “durations” where the initiation is a stray/historical date (a citation, a prior plan, a facility-construction date) rather than the NEPA start. The genuinely-long vs error split is curated in 10_outliers.R.
- LLM API call returned no date. About
2,000project dates were returned as null – hoping to dig into these to see if we can recover them with some coding elbow grease.
Known data or timeline quality issues
- Heavy reliance on proxy dates. ~55% of CE and ~68% of EIS complete timelines rest on a proxy date that stands in for a milestone the documents never recorded directly. The proxy differs by process (below).
- CE — imputed initiations. Most CEs record no distinct start, so the initiation is imputed: the pipeline takes the earliest qualifying signature/start date on the form (often the initiator’s signature, which precedes the approving official’s). ~13,100 CE initiations are proxies.
- EIS — Final-EIS-publication decisions. Only ~18% of EISs have a Record of Decision in the corpus, so for another ~18% the pipeline uses the Final-EIS publication date (often reported as month + year → imputed to the 15th) as a stand-in decision. ~58% of EISs still end up with no decision date at all.
- EAs are register-heavy. EAs have no Notice-of-Intent requirement, so a documented start date is frequently missing — 54% of EA initiations come from the BLM/DOE register rather than the EA document. Because those register “project start” dates are often late administrative entries made near the decision (some only 1–2 days before the FONSI), register-anchored EAs show artificially short durations (~60-day median vs ~12 months for document-based starts).
Next steps
- Lift EIS decision coverage — re-pull Final-EIS cover/publication dates to see if we can add more year-month dates for EIS.
- Fix the ordering check at the source — resolve the 233 decision-before-initiation rows in the selection step (currently handled by a downstream stopgap filter).
- Extend the analysis — the BLM field-office experience learning-curve test is now built (no experience effect once calendar time is controlled); geothermal timelines by field office and whether post-FRA adoption of categorical exclusions correlates with faster reviews (see Requested Extensions).
Date provenance
Every extracted date traces to one of two sources: the BLM/DOE NEPA registers (official agency dates pulled via the registers’ metadata API) or document text (parsed from the NEPA documents themselves, including LLM-adjudicated picks and Final-EIS publication dates). Figure 1 shows the split by process and endpoint — a transparency check on where the timeline rests. Registers anchor a meaningful share of CE dates; EA and EIS rely overwhelmingly on document text.
Timeline Analysis
Timeline completeness
Figure 2 shows the share of projects with complete timelines by review process; Figure 3 breaks down which component is missing. The patterns differ by process:
- CE (55% complete): decision coverage is high (91.4%) but initiation coverage is low (60.0%) — most CEs are single-signature determinations with no recorded start.
- EA (59% complete): the strongest of the three, but still initiation-limited — EAs have no Notice-of-Intent requirement and often skip scoping.
- EIS (36% complete): initiations are usually documented (an NOI is required), but decision coverage is only 42.2% because Records of Decision are frequently separate documents not in the corpus.
Comparison with Phase 1. Phase 1 measured completeness for decarb projects only and reported CE ~30% · EA ~62% · EIS ~48%. On the full database, Phase 2 reaches CE 55% · EA 59% · EIS 36%. CE is higher — the much larger universe is dominated by routine, register-dated determinations — but EA and EIS land below their Phase 1 decarb-only rates, for two reasons this analysis surfaced:
- The aggregate is diluted by non-decarb projects. Phase 1’s denominator was decarb energy; Phase 2’s is everything. Decarbonization EIS alone completes at 42.4% (much closer to Phase 1’s 48%) — it is the heterogeneous “Other” bucket that pulls the EIS aggregate down to 36%.
- The remaining gaps are structural, not fixable by better parsing. EIS completeness is capped by missing Records of Decision (a separate document, often outside the corpus — decision coverage only 42.2%); CE and EA completeness is capped by missing initiations, because neither requires a Notice of Intent and a start date is frequently never recorded. A minority of EIS “projects” are also comment letters or document fragments with no milestone date to extract. Lifting EIS coverage toward Phase 1 is the focus of the Next steps above.
Coverage also varies by energy category (Figure 4). Notably, decarbonization EIS projects complete at 42.4% — well above the EIS aggregate of 36% — because decarb EISs are better documented than the heterogeneous “Other” bucket that drags the aggregate down.
Duration analysis
Durations use all complete timelines (both dates present, including those resolved with a proxy date such as a Final-EIS publication date in place of a missing Record of Decision). Month-only dates are imputed to the mid-month 15th.
On that basis, durations track the expected complexity ladder from CE to EA to EIS — a median of about 20 days for CE, 116 days (~4 months) for EA, and 1,008 days (~2.8 years) for EIS (n = 27,052 / 1,730 / 1,321). Figure 5 summarises the spread; Figure 7 shows the full distributions; Figure 8 plots individual project spans.
Splitting duration by energy category (Figure 6) exposes a striking artifact: fossil EAs run far shorter than decarbonization or “other” EAs — a median of ~40 days, vs ~310 days for Decarb and ~209 days for Other. This is not because fossil EAs are genuinely faster: 84% of them are oil-and-gas well-abandonment EAs whose start date comes from the BLM register, where the recorded “project start” is a late administrative entry made within days of the decision. Because fossil is the largest EA group (754 of the ~1,730 complete EAs), those artificially short spans pull the overall EA median down — the truer EA review length is better reflected by the document-anchored (decarb/other) EAs.
The full duration distributions (Figure 7) show the same complexity ladder as a shape, not just a median: CE durations pile up near zero (most are days-to-weeks determinations), EA spreads out to a year or two, and EIS is broad and heavy-tailed — a meaningful mass of reviews runs 3–10+ years. That long EIS tail is real (large, contested reviews spanning multiple administrations) and is examined directly in the outlier case studies below.
Figure 8 plots each complete review as a horizontal bar from its initiation to its decision date, sorted by length within each process so the spread in durations is visible one project at a time rather than only in aggregate. CE spans collapse to short marks, EA bars range from a few months to a couple of years, and the EIS panel is dominated by multi-year bars — the same CE-to-EA-to-EIS complexity ladder seen in the histograms, now resolved to individual reviews.
Projects by decision year
Figure 9 shows project volume by decision year (all projects), faceted by review process. CE volumes ramp from the late-2000s (ARRA-era funding) with a second rise into the 2020s (BIL/IRA); EIS decisions lag CE and EA by several years, reflecting longer review cycles. Figure 10 restricts to the Department of Energy — the single largest agency in the dataset — where the ARRA/BIL/IRA grant-and-loan cycles are even more pronounced.
Timeline Outliers: Case-Study Candidates
Per the deliverable request, timeline outliers are surfaced for CATF staff to investigate via case studies (including whether NEPA itself was a cause of delay). Outliers are init→decision spans longer than 5,000 days (~13.7 years), produced reproducibly by code/deliverable04/10_outliers.R.
The critical distinction is real vs. extraction error. A multi-year span can be a genuinely protracted review or an artifact where the “initiation” was anchored to a historical citation (a facility’s construction date, a reserve’s founding, a prior plan, a regulatory-compliance milestone) rather than the NEPA start. The two are separable only by reading the evidence text, which the outlier table preserves.
How to read these (process matters)
- EIS outliers are mostly real. EIS median duration is already ~3 years; a 14–17-year EIS sits in the legitimate tail (transmission, restoration, resource-management plans genuinely run that long).
- CE outliers are mostly errors — CEs are fast determinations (median 20 days), so a multi-year CE almost always reflects a mis-extracted initiation. CE outliers are therefore excluded from the client list.
- Initiation dates before ~1990 are a strong error signal. Of the four EIS outliers with 1980s-or-earlier initiations, three are confirmed errors (the “initiation” is a park founding, a prior lock’s construction-completion date, or a RCRA compliance milestone).
The decarb EA/EIS candidates below are the highest-value case studies. SunZia (ROW application Sep 2008 → Record of Decision Apr 2023) is the textbook long-permitting saga; Energia Sierra Juarez, Grain Belt Express, and Cushman Hydroelectric are similar. Confirmed errors are excluded — e.g. the Palisades Nuclear Restart “19-year” span glues a 2005 license-renewal application to a 2024 restart Record of Decision (two different NEPA actions).
| Project | Process | Energy | Lead agency | Years | Initiation | Decision | project_id |
|---|---|---|---|---|---|---|---|
| Interstate Highway 35 Capital Express Central Project From United States Highway 290 East to United States Highway 290 West/State Highway 71 Travis County, Texas | EIS | Other | 29.8 | 1993-03-15 | 2023-01-05 | d4cb63aaadf006f332c6015042498d08 | |
| Closure of Nonradioactive Dangerous Waste Landfill (NRDWL) and Solid Waste Landfill (SWL) | EA | Other | Department of Energy | 24.5 | 1985-11-15 | 2010-05-01 | da4d59f465693f67613fb04a3be56a30 |
| Lower Cache Creek, Yolo County, Woodland and Vicinity, California Flood Risk Management | EIS | Other | 23.6 | 1996-05-06 | 2019-12-27 | 7da2bac1ab510ecd349877382a4569ae | |
| 2021 Land Management Plan Helena – Lewis and Clark National Forest | EIS | Other | Forest Service | 20.0 | 2001-10-26 | 2021-10-15 | ca3d703bbef5b90c42e4900a179cf696 |
| Rawlins Resource Management Plan | EIS | Other | 17.7 | 2002-02-25 | 2019-11-13 | 252300407e5f5b7ae87c30984faa46da | |
| Construction and Operation of The Molecular Foundry | EA | Other | Department of Energy | 16.3 | 1986-12-01 | 2003-03-06 | 8237e9b86021dba1b116398c74644f23 |
| West Mojave Route Network Project | EIS | Other | Bureau of Land Management | 16.3 | 2003-06-07 | 2019-10-02 | 48e2e33fc1555b1ef01a7c5bbc3e6c4f |
| West Mojave Route Network Project | EIS | Other | Bureau of Land Management | 16.3 | 2003-06-07 | 2019-10-02 | 34a8fa6c4bb7081831ba3e843af15db3 |
| Clearwater National Forest Travel Planning | EIS | Other | Forest Service | 16.1 | 2007-11-28 | 2024-01-15 | 9757c3901c0240237625ff424be4cdac |
| Clearwater National Forest Travel Planning | EIS | Other | Forest Service | 15.7 | 2007-11-28 | 2023-08-15 | 572a8342af3551b4c1dd71d910ea5645 |
| Pima County Multi-Species Conservation Plan | EIS | Other | 15.4 | 2000-09-07 | 2016-01-15 | 95c319c8d504a7e1436abf6de96f0a7c | |
| Placer County Conservation Program | EIS | Clean | United States Fish and Wildlife Service | 15.2 | 2005-03-15 | 2020-05-16 | 9e65f5e1614fae11f2d17be8d96fff63 |
| Malheur National Forest Site-Specific Invasive Plants Treatment Project | EIS | Other | Forest Service | 14.7 | 2000-06-26 | 2015-03-15 | 15b199816215e2ab7b2a01534eaaaeca |
| Sterling Highway Milepost 45-60 Project | EIS | Other | Federal Highway Administration | 14.6 | 2003-07-15 | 2018-03-04 | 4a660923bc39c4d901fa0e43743556c3 |
| Resource Management Plan for the Ring of Fire Planning Area | EIS | Other | Bureau of Land Management | 14.6 | 2009-06-26 | 2024-02-05 | f0519ad10cf558982119d91fc5070055 |
| Las Vegas Disposal Boundary | EIS | Fossil | Bureau of Land Management | 14.6 | 1990-05-08 | 2004-12-17 | 1ae98cccdd70178ad07bb9139acff4f1 |
| Energia Sierra Juarez U.S. Transmission Line Project | EIS | Clean | Department of Energy | 14.6 | 2009-02-25 | 2023-10-06 | 3b7e393de77613cc10641d9a352796ae |
| SunZia Southwest Transmission Project | EIS | Clean | Bureau of Land Management | 14.6 | 2008-09-15 | 2023-04-19 | 45d2e6f0c1355538141299ae292d7816 |
Document Length & FRA Page Limits
This section builds directly on Phase 1 Deliverable 5 (“Document Length Over Time and Fiscal Responsibility Act Impact”), which analyzed document length for decarb EA/EIS projects. Here we reproduce that analysis on the entire Phase 2 EA/EIS corpus, all energy types, refreshed on Phase 2 LLM-adjudicated decision dates (inclusion requires only a decision date — the time axis and FRA classification depend only on it). Each subsection below links to its Phase 1 counterpart for comparison.
Method & scope
- Regulatory pages = body word count ÷ 500, excluding embedded appendices and low-content pages, per 40 C.F.R. § 1508.1(bb) — the measure that matches the FRA limits (EA ≤ 75 pages; EIS ≤ 150, or 300 if extraordinarily complex).1
- These counts are computed from the actual page text by
fra/01_extract_pages.py(a DuckDB pass overphase2/data/processed/{ea,eis}/pages.parquet— 6.1M EIS pages — detecting the appendix boundary and low-content pages), producing regulatory pages for 5,032 EA/EIS projects (2,765 EA, 2,267 EIS) across all energy types — not just decarbonization. - The pre/post-FRA analysis below uses the 3,678 projects that also have a decision date (EA 2,239 / EIS 1,439).
- FRA date = enactment (June 3, 2023), matching Phase 1 D5 and the timeline-duration period split above.
Document length over time
Figure 11 tracks regulatory page length by decision date across all EA/EIS. Two patterns stand out. EA length sits low and essentially flat across the decade — the rolling mean hugs the low tens of pages, with only occasional long outliers — consistent with EAs being short by design. EIS length is far higher and far more dispersed: individual EISs run from tens of pages to well over a thousand, and the rolling mean hovers in the low-to-mid hundreds with no strong secular trend before the FRA. The post-FRA window (right of the red line) is denser — more recent decisions — and is where the EIS rolling mean steps down.
Compared with the decarb-only Phase 1 version, two differences follow from the wider Phase 2 scope. First, the all-energy EA cloud sits well below Phase 1’s: Phase 1’s decarb EAs averaged ~62 regulatory pages, whereas the full corpus averages ~33, because the large pool of short fossil and well-abandonment EAs (median ~7 pages) pulls the all-energy mean down. Second, the EIS picture is the same in shape — high, heavy-tailed, and dominated by project-to-project variation rather than a clear time trend — confirming Phase 1’s caution that average trends obscure substantial spread, now borne out on a roughly 5× larger sample.
Pre vs post-FRA
Across all EA/EIS, EIS regulatory length fell after FRA (mean 317 → 269 pages; median 260 → 212), while EA length was essentially flat (mean 33 → 39). The post-FRA window is short, so these are provisional (n = 218 EA, 97 EIS). The direction matches the decarb-only Phase 1 result closely (Phase 1 D5 — Pre vs Post FRA and its distribution view): Phase 1’s EISs fell from a mean of 368 → 270 regulatory pages, and the all-energy corpus here falls 317 → 269 — landing at essentially the same ~270-page post-FRA EIS mean despite a very different sample. EA stayed flat in both (Phase 1 62 → 57; here 33 → 39), reflecting that EAs were already well under the 75-page limit and had little to compress.
Breaking the same comparison out by energy category (Figure 13) shows the post-FRA EIS decline is broad — it appears in decarbonization and fossil EIS, not just one segment — while EA length is essentially flat across all categories. That consistency makes the EIS drop more credible as a genuine post-FRA effect rather than a composition artifact.
The full distributions (Figure 14) confirm the shift is not an outlier effect: the EIS distribution moves down post-FRA (lower median, thinner upper tail) while the EA distribution barely moves. Splitting that distribution by energy category — shown separately for decarbonization (Figure 15), fossil (Figure 16), and other (Figure 17) projects so each panel is legible — shows the EIS contraction holds in both decarbonization and fossil EIS, while EA length stays flat across all three.
Descriptive statistics
Table 2 reports the full descriptive statistics — mean, median, standard deviation, and interquartile range — by process and FRA period across all energy types (the all-energy counterpart to Phase 1 D5’s descriptive-statistics table).
| FRA Period | N | Regulatory Pages (body word count ÷ 500) | ||||
|---|---|---|---|---|---|---|
| Mean | Median | SD | P25 | P75 | ||
| EA | ||||||
| Pre-FRA | 2,021 | 33 | 19 | 44 | 7 | 41 |
| Post-FRA | 218 | 39 | 24 | 42 | 7 | 63 |
| EIS | ||||||
| Pre-FRA | 1,342 | 317 | 260 | 268 | 127 | 440 |
| Post-FRA | 97 | 269 | 212 | 250 | 130 | 394 |
| Post-FRA projects are those with a decision date on or after June 3, 2023. Regulatory pages exclude embedded appendix pages and low-content pages (maps, figures, blanks); documents whose filename already omits appendices use their page count directly, and documents without extractable text are excluded. | ||||||
FRA page-limit compliance
The Fiscal Responsibility Act sets presumptive page limits — 75 pages for an EA and 150 pages for an EIS (300 if the agency documents extraordinary complexity). Figure 18 scores the post-FRA EA/EIS projects with a decision date against those limits (regulatory pages).
- EAs comply at a high rate. 83% of post-FRA EAs (n = 218) fall within the 75-page limit; the 17% that exceed it generally do so modestly rather than dramatically.
- EISs are more mixed, but most are within reach of the complexity ceiling. Only 32% of post-FRA EISs (n = 97) land within the baseline 150-page limit, but a further 33% sit between 150 and 300 pages — i.e. within the extraordinary-complexity ceiling — so 65% are within the 300-page threshold overall. The remaining 35% exceed even 300 pages, consistent with the long right tail of complex transmission, restoration, and resource-management EISs seen in the duration analysis.
Two caveats temper this. First, the post-FRA window is short (218 EA / 97 EIS projects with both a regulatory page count and a decision date), so these rates are provisional and will firm up as more post-FRA decisions land. Second, the FRA limits are presumptive, not absolute — agencies may exceed them with senior-official approval or a documented extraordinary-complexity finding — so “exceeds limit” flags a document for review, not an automatic compliance failure.
Review Timelines by CEQ Regulatory Regime
Proposed section — awaiting your feedback
This section is a draft proposal, not yet built. It would test whether review durations shifted as the CEQ NEPA implementing regulations (40 C.F.R. §§ 1500–1508) were rewritten across administrations — a different breakpoint from the FRA statutory split used above. Please review the framing and planned figures and tell me whether to build it.
NEPA’s procedural rules were rewritten three times in four years. The natural regulatory regimes — keyed to each rule’s effective date — are summarized below. (The FRA, June 3, 2023, is a statutory amendment to NEPA, not a CEQ rule; it is already the pre/post-FRA breakpoint used in the duration and page-length analyses.)
| Regime | Effective | What changed |
|---|---|---|
| 1978 regulations | 1978–2020 | Original CEQ regulations; essentially unchanged for ~42 years. |
| 2020 Final Rule (Trump) | Sep 14, 2020 | Major rewrite: presumptive 2-yr EIS / 1-yr EA limits, page limits, narrowed effects/cumulative-impacts definitions. |
| 2022 Phase 1 Rule (Biden) | May 20, 2022 | Restored 1978 purpose-and-need and cumulative/indirect effects. |
| 2024 Phase 2 Rule | Jul 1, 2024 | Implemented the FRA's statutory NEPA amendments (codified page/time limits, CE sharing, lead-agency roles). |
| 2025 rescission | early 2025 | After EO 14154 and the Marin Audubon ruling, CEQ rescinded its regulations; agencies revert to their own NEPA procedures. |
Planned analysis. The pipeline already carries the scaffolding for this — 08_analyze.R computes a reg_period segmentation and writes d4_duration_by_period.csv, currently keyed to funding events (ARRA / BIL / IRA). We would re-point those cut dates to the CEQ rule effective dates (2020-09-14, 2022-05-20, 2024-07-01) and surface two views:
- Median duration by regulatory regime — a per-process interval/bar chart comparing median review length across the 1978 / 2020 / 2022 / 2024 regimes (from the already-computed
d4_duration_by_period.csv). - Duration trend by decision year with regime markers — the year-by-year median-duration line (
fig_d4_duration_trend.png, already rendered) with vertical markers at the three CEQ rule dates, replacing or augmenting the current ARRA/BIL/IRA markers.
Key caveat to flag up front. Duration is measured at the decision date, but a multi-year review mostly occurs under earlier rules — so this reads as “reviews decided under a regime,” not “reviews governed by” it. Post-2020 samples are also thin and confounded with the coverage ramp (the same provisional-sample limitation as the post-FRA split).
Feedback requested: (1) Is “reviews decided under a regime” the right framing, or should we anchor on the initiation date instead? (2) Keep the FRA statutory marker alongside the CEQ rule markers, or present them separately? (3) Which view do you want — the regime comparison, the trend-with-markers, or both?
BLM field-office experience
This extension answers a request from the deliverable scope: do BLM field offices process NEPA reviews faster as they accumulate experience — a “learning-curve” effect? We map each BLM-led project to the field office that handled it (from the structured DOI-BLM-<state>-<office>-<year>-<seq> case number), order each office’s reviews by the office’s own cumulative count, and test whether duration falls as that count rises — after separating the experience signal from two confounds that would otherwise manufacture a false result: the register-anchoring artifact and secular calendar-time trends. All conclusions here are associational, not causal.
Method & scope
- Field-office parse. The validated regex
BLM-([A-Z]{2}-?[A-Z]?[0-9]{2,4})on the case number (file name as fallback) maps 62.5% of BLM-led projects (16,249 of 26,016) to one of 207 field offices (15,131 from the structured case number, 1,118 from the file name). - Document-anchored durations are primary. BLM register “project start” dates are frequently late administrative entries (the ~40-day artifact documented in the duration analysis above), so the learning curve is computed on document-anchored initiation dates (
initiation_source_type ≠ metadata); the register-anchored version is a clearly-flagged secondary view. - EA and CE are analyzed separately (CE medians run in weeks, EA in months); EIS is excluded (too sparse per office). Offices qualify with ≥ 30 valid durations on the analysis measure.
- Model. A pooled regression with office fixed effects,
log(duration) ~ log(cum_count) + factor(decision_year) + project_energy_type, fit separately for EA and CE. Thelog(cum_count)coefficient is the experience effect net of secular calendar trends;factor(decision_year)is the calendar control,project_energy_typethe project-mix control.
Finding: the apparent learning curve is a calendar-time confound, not experience
Descriptively, CE reviews look like they speed up as an office gains experience: across the 32 offices with ≥ 30 document-anchored CE reviews, the median duration of an office’s earliest third of reviews is 53 days versus 29.25 days for its most-experienced third, and 62.5% of offices sped up (Figure 21). But that gradient is an artifact of calendar time. CE durations have fallen sharply over the 2010s corpus-wide, and because an office’s higher-experience reviews are also its later-calendar reviews (within-sample correlation ≈ 0.67), the raw pattern conflates the two (Figure 20). The office fixed-effects regression that adds a decision-year control dissolves the effect: the coefficient on log(cumulative reviews) is +12.7% per doubling (95% CI -1.3% to +28.6%) for the primary document-anchored CE measure (n = 1,788 reviews, 32 offices) — statistically indistinguishable from zero and, if anything, the opposite sign — even though the same model without the calendar control reproduces the apparent -7.2% speed-up. The register-anchored CE sensitivity points the same way (+10.0%, CI +4.5% to +15.9%). EA cannot support the office-level design: no office clears the ≥ 30 document-anchored-review bar (only 144 such EA durations exist corpus-wide), and the exploratory register-anchored EA estimate (-7.8%, CI -19.1% to +5.1%) is not distinguishable from zero and rests on artifact-prone start dates. Net of calendar time, there is no evidence that BLM field offices grow faster with experience.
Figure 19 shows the raw, uncontrolled gradient that motivates the question — and that the regression then explains away. Document-anchored CE durations decline across the cumulative-review deciles; the register-anchored lines sit far lower (especially for EA), a direct picture of the late-administrative-entry artifact.
The confound is direct in Figure 20: median CE duration falls with both calendar year and cumulative experience, because the two move together (later reviews are later in time). Only the regression, which holds decision-year fixed, can separate them — and once it does, the within-office experience coefficient is null/positive.
Per-office heterogeneity (Figure 21) reinforces the null: some high-volume offices show shorter durations for their most-experienced third, others longer — there is no consistent within-office speed-up once each office is its own baseline. Because a later tercile is also a later calendar period, most of the leftward movement that does appear is the same secular decline, not learning.
| Process | Measure | Min reviews/office | N reviews | N offices | %/doubling (no calendar control) | %/doubling (calendar-controlled) | 95% CI | Role |
|---|---|---|---|---|---|---|---|---|
| CE | Document-anchored | 30 | 1788 | 32 | -7.2% | +12.7% | -1.3 to +28.6% | Primary |
| CE | Register-anchored | 30 | 6395 | 69 | +10.7% | +10.0% | +4.5 to +15.9% | Sensitivity |
| EA | Register-anchored | 10 | 946 | 24 | -6.5% | -7.8% | -19.1 to +5.1% | Exploratory (artifact-prone) |
| EA has no office with ≥30 document-anchored durations (144 corpus-wide), so the EA row uses the register-anchored measure at a relaxed ≥10 threshold and is exploratory only. | ||||||||
Caveats
- Calendar-time confound (the headline). The raw “offices speed up with experience” pattern is secular calendar drift; net of decision-year there is no experience effect. This is the central caveat — the descriptive figures must not be read on their own.
- Register-anchoring artifact. Durations rest partly on register/proxy start dates that are late administrative entries; the primary analysis uses document-anchored dates to limit this, but the secondary register views inherit the artifact (visible as the much shorter register lines in Figure 19).
- ~38% of BLM projects are unparsed (37.5%). EISs account for most of the gap — only 7.4% parse (programmatic/large EISs are filed under titles, not DOI-BLM case numbers), versus 63.9% of CE and 66.8% of EA. Many unparsed EAs carry an office-like code in a non-standard form (bare
C069-2023-…,AZ-A010-…) lacking theBLM-prefix the regex anchors on, and are left out rather than guessed. - Project mix. Short well-abandonment EAs dominate some offices;
project_energy_typeis included as a covariate, but a “faster” office could reflect a shift toward simpler actions rather than learning. - Office identity may shift over time (renames, mergers, district reorganizations); a stable
office_codeis assumed.
Requested Extensions
The following items from the deliverable request require additional analysis beyond the current timeline build. They are scoped here for the next phase:
| Requested analysis | What it needs |
|---|---|
| Geothermal timelines by BLM field office | Geothermal project flags (Deliverable 1/3 tech tags) × field office × duration; limited by geothermal sample size per office. |
| CE adoption post-FRA vs. review speed | CE-adoption events (which agencies adopted other agencies' CEs post-FRA, BLM/DOE/USFS) × subsequent review duration; needs the CE-adoption register. |
Reproduction
Run from the repository root in the nepa conda environment:
# Timeline pipeline (candidates -> classify -> select -> LLM adjudicate)
conda run -n nepa python phase2/code/deliverable04/run_pipeline.py
ANTHROPIC_API_KEY=$(security find-generic-password -s nepa-anthropic -w) \
conda run -n nepa python phase2/code/deliverable04/06_adjudicate_llm.py \
--mode candidate_adjudication --process CE EA EIS --model claude-haiku-4-5-20251001 --workers 24
# Analysis + figures
Rscript phase2/code/deliverable04/08_analyze.R # coverage, durations, FRA, energy
Rscript phase2/code/deliverable04/10_outliers.R # timeline outliers (case-study candidates)
# FRA document-length analysis (regulatory pages from the full page text, all EA/EIS)
conda run -n nepa python phase2/code/deliverable04/fra/01_extract_pages.py --run
Rscript phase2/code/deliverable04/fra/02_pages_fra.R # document length / FRA page limits
# BLM field-office experience / learning-curve extension
conda run -n nepa python phase2/code/deliverable04/field_office/01_parse_offices.py --run
Rscript phase2/code/deliverable04/field_office/02_learning_curve.ROutputs are written to phase2/output/deliverable04/ (figures/, diagnostics/).
Draft generated 2026-06-26 | NEPA Decarbonization Technology Analysis — Phase 2, Deliverable 4
Footnotes
We use regulatory pages rather than raw PDF page counts because raw counts substantially overstate the length the statute limits — they include embedded appendices and sparse pages (covers, maps, dividers). Mean raw pages run roughly 2× regulatory pages for EAs and higher for EISs, so a raw count cannot be compared to the 75/150-page limits.↩︎