The archived dataset states every finding, its measured figures, and the rule a reader needs to use the data safely — but not the story of how each finding was reached: the hypotheses tested and killed, the corrections that were themselves corrected, the leads not yet run down. That history lives here, where it can keep being amended, and each entry below is pointed at from the archive by a stable anchor.
Entries are extracted verbatim from the dataset's working record at the date shown and lightly amended only below a dated line. The frozen archive is the authority for every figure; this page is the authority for the narrative.
POLICY.ar.caveat — extracted 2026-08-18
how the withdrawn EIA 169-barrel over-claim was found to be a rounding artifact
It read: reproduces the AOGC GRAND TOTAL to 169 barrels in 6.78 million (0.002%). EIA PUBLISHES ARKANSAS OIL IN WHOLE THOUSANDS - every one of the twelve 2014 months is an exact multiple of 1,000 - so no 6.78 million figure can sit further than 500 barrels from it, and 169 is merely where 6,775,169 falls against the 6,775,000 boundary. It measured rounding, not agreement, and it would have read just as well against a series that was 400 barrels wrong. It was also quoted against the wrong vintage - the xlsx rather than the printed report, from which EIA differs by 21,620 barrels in 2014.
POLICY.ar.caveat — extracted 2026-08-18
how the per-year median statistic was found biased (58th percentile) and corrected
(This read SIX until the per-year median was corrected: byYear() had been taking the upper-middle value of an even count, the 7th of 12, which is the 58th percentile and not a median. 1995 and 2011 were artefacts of that bias.)
POLICY.ar.caveat — extracted 2026-08-18
the enumeration-blindness code-fix mechanics (mergeInvisible vs enumerateCached)
The CODE defect is closed: --probe-gaps now folds its discovery into the index through mergeInvisible and persists it, where enumerateCached used to overwrite without merging.
POLICY.nd.caveat — extracted 2026-08-18
the discovery headline: the archive was short 284 MMbbl / 371 Bcf until 2026-08-14 while every gate said ok
NORTH DAKOTA WAS SHORT 284,385,537 BBL AND 370,688,682 MCF UNTIL 2026-08-14, AND EVERY GATE IN THIS REPO SAID "ok".
POLICY.nd.caveat — extracted 2026-08-18
how the share defect was found: fetcher summed only Oil and Gas; per-county ratios exposed it where statewide could not
fetch-nd-dmr.mjs summed Oil and Gas and nothing else. IT WAS A SHARE DEFECT, WHICH IS WHY ONLY A PER-COUNTY INSTRUMENT COULD SEE IT. Confidential wells are new completions, so the loss concentrated where the drilling was: against DMR's own county publication over 2015-05..2025-08, williams ran at 0.9108 of DMR and mckenzie 0.9263, while ward ran 0.9993 and slope 0.9996. The per-county ratio IQR was 0.0447 on oil. No statewide ratio can contain that information, and every reconciliation instrument in this repo is statewide.
POLICY.al.caveat — extracted 2026-08-18
the 2026-08-08 offshore-fields discovery story (dated historical statement about an archive the file no longer holds)
ALABAMA GAINED 36.4 MMBBL ON 2026-08-08 — the archive was short, not EIA. The fetcher discovered fields from WELLS, which cannot see a field that has no wells: 18 offshore fields (CountySearch matches county names exactly, and "Mobile" never returns a Mobile-OS well) plus four of OGB's own catch-all field codes that hold production not assigned to any named field. Fields are now enumerated from /Production/getFields/. That is a dated historical statement about an archive this file no longer holds, and it is the last one here: every figure after this sentence is measured at generation time.
POLICY.al.caveat — extracted 2026-08-18
the withdrawal history of the old no-year-exceeds-1.04 sentence (the permutation-blindness rule it motivates stays)
THE OLD "NO YEAR EXCEEDS 1.04" SENTENCE IS WITHDRAWN, AND NOT BECAUSE IT WAS FALSE. It read: nothing overshoots, no year exceeds 1.04, which is the check that nothing was double-counted. It still holds numerically — the highest Alabama gas year is 2024 at 1.0141 on the per-year median, and across BOTH fields and BOTH statistics the largest single-year figure anywhere in Alabama is oil 2025 at 1.0309 on the annual total — and it checks nothing.
POLICY.al.caveat — extracted 2026-08-18
hypotheses tested and killed for the 2024 gas excess
Three explanations were tested and rejected: the fixture is not stale (EIA series re-pulled 2026-07-31, all twelve months identical digit for digit); it is not a generic unrevised-tail artifact (MI and VA fixtures end on the same month and reconcile); and it is not the MS/CO CO2 pattern (that produces 30-65%, not 2-3%).
POLICY.al.caveat — extracted 2026-08-18
the withdrawn Clarke County explanation
THE CLARKE COUNTY EXPLANATION IS WITHDRAWN. This entry used to attribute about 39% of the excess to Clarke County moving 922 to 2,061 MMcf between 2023 and 2024. Clarke reports ZERO gas in every month from 2021-01 onward, so that paragraph was describing volume the archive does not contain — a county-level story told about a state whose county axis has since been rebuilt from permits.
POLICY.oh.caveat — extracted 2026-08-18
how the not-circular inference was found wrong (the 4-of-195 statistic misread)
This entry used to assert that the comparison was not circular, citing 4 of 195 monthly sums rounding to EIA's published thousand with a longest run of 1, and concluded from that the agreement was evidence rather than an echo. The statistic is real; the inference from it is not.
POLICY.oh.caveat — extracted 2026-08-18
the 18-vs-58-vs-62 comparator correction story (the 55-county convention it lands on stays)
THE COMPARATOR IS LOAD-BEARING AND THIS SENTENCE USED TO GET IT WRONG: it read "18 counties rather than 58". 58 was not invented — it is the PEAK single year (2013, 58 counties) — and 62, every Ohio county that has ever reported, is the archive roster and the answer a re-derivation would reach for. Neither is the right comparand.
POLICY.pa.caveat — extracted 2026-08-18
how the annual-sums-rescue implication was found wrong (whole-series medians quoted as an era verdict)
ANNUAL SUMS DO NOT RESCUE THOSE YEARS, AND THIS ENTRY IMPLIED THEY DID UNTIL 2026-08-15. It read: judge those years on annual sums instead, from the 1996-01 source boundary they give 0.982 oil and 1.004 gas. Those two figures are not pre-2012 annual sums at all — they are the WHOLE-SERIES medians of the monthly ratios, i.e. section 6's baseline column, quoted as though they were a verdict on the era the sentence was excusing.
POLICY.pa.caveat — extracted 2026-08-18
the withdrawn 2005-2008 four-year-era wording
THIS USED TO READ "2005-2008 ... 1.08-1.13", A FOUR-YEAR ERA THE TABLE EXCLUDES A YEAR FROM.
POLICY.pa.caveat — extracted 2026-08-18
the editorial principle about spans contradicting the table (no operational content for a citer)
An unexplained era with a clean year in the middle of it is a different finding from a contiguous one, and the prose must not assert a span the table beside it contradicts.
POLICY.mi.caveat — extracted 2026-08-18
the 2026-08-13 under-collection defect story (county-filter omission, Kalamazoo at 25%) and its repo pointer
Until 2026-08-13 the archive under-collected Michigan oil by up to 13%, and the caveat here read that EIA "stops echoing" from 2020-01 and the two disagree by 7-15%, growing. That divergence was OURS: fetch-mi-egle.mjs swept EGLE Data Miner once per county and its server-side county filter silently omitted newer high-numbered PRUs, so the state growth story - Wolverine and Savoy horizontals in Kalamazoo/Calhoun, Core Energy CO2-EOR in Otsego - never entered the file. Kalamazoo, the largest oil county in the state, published at 25% of its true volume for six years. See docs/PLAN-MI-OIL-GAP.md.
POLICY.mi.caveat — extracted 2026-08-18
probe-next notes: candidate third sources
The candidates are Michigan Treasury severance-tax filings (Form 381, aggregates unpublished, so FOIA) or a commercial dataset. MPSC covers gas wells only and its series stopped in June 2022.
POLICY.ms.caveat — extracted 2026-08-18
the replaced blind-by-construction county-exclusion instrument and its measured cost
THIS REPLACES A COUNTY-LEVEL EXCLUSION THAT WAS BLIND BY CONSTRUCTION. Until 2026-08 the comparison dropped all of Rankin and Madison from 2013-01, which papered over the CO2 but also deleted 22,280,483 Mcf of genuine hydrocarbon those counties produced — and, worse, made the reconciliation structurally incapable of seeing the defect it was compensating for. Its cost was measurable: 2013-2016 read 0.876 / 0.872 / 0.880 / 0.877 under the county rule and read 0.982 / 0.981 / 0.981 / 0.980 under the classifier. From 2017 Rankin's hydrocarbon gas is near zero and the two instruments coincide.
ACCEPTED_YEARS.wy.gas — extracted 2026-08-18
the Lincoln/county-exclusion hypothesis measured and killed, including the correction of its earlier understatement
Measured anyway, and re-measured 2026-08-15: excluding Lincoln leaves 2017 at 1.030, 2018 1.034, 2022 1.023 and 2025 1.071 — still outside the band — while pushing 1993 to 0.951 and EVERY ONE of the 22 currently-reconciling years below 0.99, as far down as 0.877 (1992) and 0.902 (1991). That last clause read "1993 to 0.931 and four reconciling years to 0.979", which named the three mildest of the twenty-two and made the over-correction look survivable; it is not. Lincoln+Sublette gives 0.43-0.65 across the series; the Powder River CBM counties give 0.877-0.965. No county subset is ~0 in 2016, 92,552 MMcf in 2017, ~0 in 2020 and 115,321 MMcf in 2025.
ACCEPTED_YEARS.wy.gas — extracted 2026-08-18
the withdrawn seventeen-year count and the provenance search that failed to reproduce it
THE COUNT USED TO READ SEVENTEEN AND NOTHING REPRODUCES IT. Taken separately the two halves give 13 years (median inside 1.000-1.004) and 10 years (spread at or under 0.010 — the census CIRCULAR_SUSPECTED.wy.gas quotes); their union is 15 and their intersection is the 8 named above. No provenance for seventeen survives, so it is withdrawn rather than reinterpreted.
ACCEPTED_YEARS.co.gas — extracted 2026-08-18
the dead Montezuma-dropout hypothesis (three independent refutations), the generalised-form refutation, the two corrected circulated figures, and the probe-next notes
THE MONTEZUMA-DROPOUT HYPOTHESIS WAS TESTED AND IS DEAD. DO NOT RE-RUN IT. The idea is the obvious one: §2a of this document reports 18 months where Montezuma gas collapses to 0.0028-0.0068 of its own neighbours, so perhaps this era is that hole. It is not, for two INDEPENDENTLY SUFFICIENT reasons and a third that disposes of the generalised form. (1) THE EXCLUSION IS BLANKET, SO MONTEZUMA CONTRIBUTES EXACTLY ZERO. POLICY.co.exclude.gas names montezuma and dolores with NO start period, so neither county is inside this comparison in any month and neither can move it. Repairing all 18 dropouts to the lesser of their trailing and leading neighbour medians — 520,330,649 Mcf written in, against the 2,332,106 Mcf actually filed in those months — leaves all twelve per-year medians and the whole-series baseline of 1.0166 BIT-IDENTICAL. Repairing dolores as well (22 months, 528,423,252 Mcf) changes nothing either. (2) THE SIGN IS BACKWARDS. Every one of these twelve years sits ABOVE the 1.02 ceiling, 1.0238 (2004) to 1.1186 (2008), and a dropout repair only ADDS gas. Run the comparison with the exclusion lifted, where Montezuma can move it, and the repair makes the era WORSE: 2010 goes 1.2222 to 1.3118, 2008 1.4101 to 1.4217, 2009 1.3549 to 1.3556, and no year in the era improves by any amount. A hypothesis that has to move the ratio DOWN cannot be tested with an instrument that only moves it UP. (3) COVERAGE FAILS TOO, AND WOULD FAIL EVEN IF THE SIGN WERE RIGHT. The dropouts span 2008-06..2014-12, which is not this era: NINE of these twelve years contain zero Montezuma dropouts, and eight of the 18 fall in 2013 and 2014, outside the era entirely. Only 2008 (2), 2009 (2) and 2010 (6) overlap it at all. GENERALISING THE HYPOTHESIS DOES NOT RESCUE IT. Repairing EVERY severe month-low dropout the self-consistency detector finds across Colorado, Mississippi, Alabama and Texas — 340 county-months, 245 of them gas totalling 556,459,136 Mcf and 95 oil totalling 1,456,324 bbl — closes ZERO out-of-band years in any of those four states and breaks none. Widening the repair to `notable` severity as well (1,512 county-months) also closes zero. County-month dropouts are not what any of these eras are made of. TWO FIGURES THAT CIRCULATED WITH THIS HYPOTHESIS WERE WRONG, and are corrected here so they do not get re-derived: the Montezuma dropouts are 18 months over 2008-06..2014-12, NOT ~12 months in 1999-2011; and Texas oldham and hale carry ZERO dropouts in 2004 (oldham: 2018-04, 2018-06, 2020-05; hale: 1998-12, 2000-03, 2013-07, 2013-12, 2018-04, 2022-03), so they cannot explain the tx.gas 2004 finding either. WHERE THE NEXT PROBE SHOULD GO: AT THE REFERENCE, NOT AT OUR COUNTIES. The divergence ENDS at 2014 and stays ended — 2014 through 2026 run 0.9905-1.0129 in every single year, inside the band, with nothing changing on our side at that boundary and no county entering or leaving. That is the exact signature ACCEPTED_YEARS.wy.gas documents for Wyoming, where EIA's own supply-and-disposition components (Nonhydrocarbon Gases Removed, Repressuring) are visibly restated across the same windows in which its gross withdrawals disagree with the regulator. Colorado's counterpart components — Nonhydrocarbon Gases Removed, Repressuring, Vented and Flared — are NOT in scripts/fixtures/, which carry gross withdrawals only, so testing this needs a fetch first. That is the open door. The county-dropout door is closed.