D4: Timeline Coverage & Limitations
As of 2026-07-23 (Tier-C section-retrieval restore; full-universe reconciliation applied 2026-07-15; 48 new adjudications applied 2026-07-23). All numbers on this page are recomputed from the current timeline_project_dates.parquet / timeline_candidates.parquet or read from the report’s diagnostics (output/deliverable04/diagnostics/).
This page explains why timelines are missing where they are missing: what coverage each review process reaches, where the gaps come from, and which gaps are structural (the date is not in the corpus) versus recoverable. Companion pages: Date Sourcing & Provenance (what the dates mean and where each one comes from) and Known Issues & Deferred Items (what is known-broken and why it was deferred).
Headline coverage
A project’s timeline is complete when both an initiation and a decision date are present — the definition used in the report’s coverage figures.
| Review type | Projects | Has initiation | Has decision | Complete (both) |
|---|---|---|---|---|
| CE | 54,668 | 32,613 (59.7%) | 49,392 (90.3%) | 29,916 (54.7%) |
| EA | 3,083 | 2,060 (66.8%) | 2,458 (79.7%) | 1,798 (58.3%) |
| EIS | 4,130 | 2,585 (62.6%) | 1,749 (42.3%) | 1,487 (36.0%) |
| All | 61,881 | 37,258 (60.2%) | 53,599 (86.6%) | 33,201 (53.7%) |
A small share of the “both dates” rows have the decision before the initiation (invalid_order: CE 1,414 · EA 56 · EIS 155, mostly month-imputation artifacts). They count as covered above but are excluded from every duration figure; restricting completeness to valid-order rows gives CE 52.0% · EA 56.5% · EIS 32.3%.
Reading the table: for CE and EA the decision is well covered (81–91%) and the initiation is the scarce endpoint; EIS is the reverse — initiations are well covered (a Notice of Intent is required) but decision coverage is only 42.2%. For duration context, the headline medians on complete timelines are CE 20 days (n = 27,275), EA 117 days (n = 1,736), EIS 1,021 days (n = 1,329); duration n’s are slightly below the complete counts because year-granularity endpoints are excluded from durations (see the duration-frame definition in Date Sourcing & Provenance).
The denominator: the full inventory
The table’s 61,881 projects are the complete NEPATEC 2.0 project inventory — coverage denominators and the report’s headline project count are the same universe. This holds by construction since 2026-07-15: the pipeline’s universe reconciliation (reconcile_universe in 05_select_dates.py, EIS-only in earlier runs) now covers all three process types, so the 628 CE and 65 EA projects whose documents yield zero date candidates appear in the output as missing_both stubs rather than silently vanishing. (Earlier published figures used the 61,187-project pipeline universe; the harmonization added only dateless stubs, so every date, duration, and median is unchanged.)
Main reason for low coverage, per review process
| Process | The cap is on… | One-line reason |
|---|---|---|
| CE | initiation | Most CEs are fast determinations with no initiation date recorded — structural. |
| EA | initiation | EAs have no Notice-of-Intent requirement and often skip scoping, so a start date is rarely documented; it exists mainly where the BLM/DOE register captured it — structural. |
| EIS | decision | EIS initiations are usually documented (an NOI is required), but the Record of Decision is frequently a separate document not in the corpus, so the Final-EIS publication date is used as a proxy decision; some EIS “projects” are also comment letters/fragments with no milestone. |
So CE and EA are initiation-limited (structural), and EIS is decision-document-limited — three different, defensible root causes, none of which is a pipeline defect.
Where the no-dates come from
Projects with no date at all (missing_both): CE 4.7% · EA 11.8% · EIS 31.1% of each process’s population. EIS is the structural outlier.
Zero-candidate projects — projects whose documents produced no date candidates at all, materialized as missing_both: EIS 401 (9.7%; 41 of the 753 clean-energy EIS), CE 628 (1.1%), EA 65 (2.1%). Reading the EIS set, they are non-milestone documents — EPA/agency comment letters, draft-review correspondence, short fragments, and a few OCR-garbled scans. A comment letter’s date is not a NEPA milestone, so extracting nothing is the correct behavior.
Structural share of the remaining gaps — a missing endpoint is structural when the project has no candidate of that role anywhere in its documents (nothing for selection or LLM adjudication to pick):
| Process | Missing initiation | …with no initiation-role candidate | Missing decision | …with no decision-role candidate |
|---|---|---|---|---|
| CE | 22,055 | 20,584 (93%) | 5,276 | 1,030 (20%) |
| EA | 1,023 | 481 (47%) | 625 | 231 (37%) |
| EIS | 1,545 | 937 (61%) | 2,381 | 1,277 (54%) |
(An “initiation-role candidate” is an extracted date whose role cue is a clear or proxy initiation; a “decision-role candidate” is a clear/proxy decision or the body-text decision holding category.)
The LLM adjudication layer has already been applied: every completable ambiguous case was sent to Claude Haiku (11,264 cached adjudications cumulative — 11,216 from the June 2026 full run + 48 from the 2026-07-23 Tier-C restore; 10,334 projects in the current output carry an LLM-sourced date; cumulative cost $18.28). The remaining gaps are therefore not “waiting on the LLM” — they are dominated by the structural buckets above.
For EA specifically, a June 2026 full-text audit of 100 EA decision-only projects found 82 had no initiation signal anywhere (no application-received, scoping, NOI, or pre-filing date); several of the remaining 18 signals were not true NEPA initiations (“external scoping was deemed unnecessary”, water-permit applications). EA initiation is genuinely sparse at the source, not under-extracted.
The EA register-anchor artifact
54% of EA initiation dates (1,112 of 2,053) come from the BLM/DOE register rather than the EA document. Those register “project start” entries are often late administrative entries made near the decision, which compresses the measured span. On complete EA timelines (headline duration frame), split by initiation_source_type:
- Register-anchored (
metadata): 1,104 projects, median 60 days initiation→decision. - Document-anchored (
document_text): 420 projects, median ~380 days (~12.5 months). - The remaining ~212 complete EAs have LLM-adjudicated or hand-verified initiations (medians ~180 days, in between).
Register-based EA durations therefore understate the true process length by roughly a factor of six; the document-anchored figure is the more defensible review-length view (the report’s Figure 6 shows both for fossil EAs, where the artifact is strongest).
EIS: the FEIS-publication fallback and month granularity
- Only 732 of 4,130 EIS (17.7%) have a Record of Decision in the corpus. The pipeline searches ROD-first (register ROD → ROD-typed document → ROD language in narrative), and only when no ROD exists anywhere falls back to the Final-EIS publication date as the decision, flagged
decision_is_feis_fallback. - In the current output, 757 of the 1,749 EIS decisions (43%) are FEIS-publication fallbacks, and 503 EIS decisions (29%) are month-granularity dates imputed to the 15th (a further 4 are year-only and excluded from durations). Durations built on them are month-precision.
- Why the gap cannot be cheaply closed (a selection-only fix was tested and abandoned) is documented in Known Issues & Deferred Items.
Coverage in the clean-energy subset
Complete-timeline coverage for clean-energy (decarbonization) projects: CE 49.2% (9,554/19,399) · EA 47.8% (274/573) · EIS 43.0% (324/753). Note that decarb EIS (43.0%) sits well above the EIS aggregate (36.0%) — clean-energy EISs are better documented than the heterogeneous “Other” bucket that drags the aggregate down.
Phase 1 vs Phase 2 — the two published figures are not the same universe
- Phase 1’s complete-timeline figure covered clean-energy projects only (denominators 19,399 CE + 573 EA + 753 EIS = 20,725).
- Phase 2’s figure covers the entire database (54,668 CE + 3,083 EA + 4,130 EIS = 61,881 projects: clean + fossil + other).
Comparing the two figures directly is misleading. Two comparisons:
As published (different universes):
| Review type | Phase 1 (clean only) | Phase 2 (all projects) |
|---|---|---|
| CE | 5,899/19,399 (30%) | 29,916/54,668 (55%) |
| EA | 355/573 (62%) | 1,798/3,083 (58%) |
| EIS | 362/753 (48%) | 1,487/4,130 (36%) |
Apples-to-apples (clean-energy only, both phases):
| Review type | Phase 1 clean | Phase 2 clean | Δ pts |
|---|---|---|---|
| CE | 5,899/19,399 (30.4%) | 9,554/19,399 (49.2%) | +18.8 |
| EA | 355/573 (62.0%) | 274/573 (47.8%) | −14.2 |
| EIS | 362/753 (48.1%) | 324/753 (43.0%) | −5.1 |
On a like-for-like clean-energy basis, Phase 2 CE completion is far above Phase 1 (authoritative register dates plus the recovered CE-form extraction), while EA and EIS land below their Phase 1 rates for the structural reasons above: EA initiation is source-sparse (Phase 1’s higher EA rate leaned on noisier candidate dates that Phase 2 correctly declines), and EIS is capped by ROD availability in the corpus (the EIS shortfall is a corpus limitation, not an extraction regression — see the methodological note in Known Issues & Deferred Items).
There is no universe drift. Every one of Phase 1’s 20,725 clean projects is present in the Phase 2 output, and the two phases’ clean-energy denominators are identical. (Before the June 2026 candidate-extraction fixes, 1,376 were missing; the last 148 were zero-candidate projects, materialized as missing_both by the 2026-07-15 universe reconciliation.)
Caveats to keep stating
- Coverage is not accuracy. “Complete” means both dates were found; end-to-end date accuracy has not yet been formally validated against a held-out gold set (see Known Issues & Deferred Items on why the current gold set cannot serve that role). Proxy and fallback dates are flagged and reported in tiers.
- Register-anchored EA durations understate the true process length; prefer document-anchored views where sample size allows.
- A large share of EIS decisions are Final-EIS-publication proxies at month granularity; EIS durations are best-available estimates given what NEPATEC 2.0 contains.
- Clean-energy EA completion (47.6%) lands under Phase 1’s 62% because EA initiation is structurally sparse — this is expected, not a regression.
- EA coverage is not governmentwide. In NEPATEC 2.0, EA (and CE) intake is concentrated in DOE, BLM, and a small USDA/Forest Service contribution, whereas EPA supplies a broad, governmentwide EIS crawl. Every EA figure therefore describes EAs in this corpus — heavily BLM/DOE — not the full federal EA population, and the corpus contains fewer EAs (3,083) than EISs (4,130), the inverse of the true national ratio.