Claude Reads ARIES
CRUDECUTS № 002 · CLAUDE READS ARIES EFF DATE · AUG 2026

Crude Cut № 002 · tutorial · five questions, one Access file

Claude
Reads ARIES

Every evaluation shop has one: a .accdb — a Microsoft Access database that only ARIES itself will open. Before anyone opens it, there is a first question worth asking the file directly: what is actually in here? We pointed Claude at a demo evaluation — 20 producing oil wells, five operator groups, three states — plus the two files that ship beside it in every deal, the oneliner Excel and the Reserves & Economics PDF. Then we asked five questions: four about what is actually in the file, then the one everyone wants to ask first. Each tab below is one question: the prompt as typed, the answer as returned. Setup is one line: brew install mdb-tools.

Question one · the manifest

Sixty tables. Twenty-eight have anything in them.

Here's an ARIES database. Before I open it: what's in it, and what's actually populated? — plus the file path

Claude lists the tables with mdb-tables, counts every one, and sorts them by what ARIES itself says they are. The file is a 5.7 MB ARIES 2017.1 database: one DBS, one project (YE26_NON-OP, a transported project), one scenario (2026Q3), effective 08/2026. Of 60 tables, 28 hold rows and 91% of all rows are daily production. The per-property tier is where a deal lives, and it is thin: five of nineteen tables populated — the master, the economics lines, saved fits, monthly and daily history. Every run-output table is empty. The PV10 on the PDF cover exists only in the exports; the database holds inputs and history, not results.

60
Tables · 28 populated · 32 empty
32,367
Rows · 91% of them daily production
0
Rows in the four run-output tables

Table manifest · every table in the file · rows in each · hover for what it holds

populated empty run output · empty

RESULT: per-property tables populated — AC_PROPERTY 20 · AC_ECONOMIC 590 · AC_FCST 128 · AC_PRODUCT 961 · AC_DAILY 29,580. Empty: ONELINE, ECOSUM, MONTHLY, DETAIL (the run results), OWNER, WELL, TEST, NOTE, RESERVES, PZFCST, RATIO. Instructions: 9 scenario lines, 112 side-file lines holding five price decks (two of them unreferenced), one lookup table (Oklahoma taxes).

Run this on your own database

This page ran on a demo file, but the moves are the same on any ARIES database: one brew install mdb-tools, then the five prompts above, verbatim, in a Claude session pointed at your .accdb. The full recipe — prompts, the commands underneath, and what to check — ships with this cut.

crudecode.dev
Methodology

The files: a demo ARIES evaluation — alchemist energy-20260721.accdb (5.7 MB, exported 07/21/2026), its oneliner Excel (20 rows × 49 columns), and its Reserves & Economics PDF (12 pages, DBS "DEMO", scenario 2026Q3, effective 08/2026). The wells and operators are real and public record; the ownership interests and economics are demonstration constructs. Tools: mdb-tools (Homebrew) to read the .accdb, openpyxl for the Excel; every number on this page comes out of one extraction script shipped with the cut.

The manifest (Q1): mdb-tables for the list, mdb-count per table. Tiers are ours, derived from ARIES's own registry (ARSYSTBL): tables it parents on the master (PARENT = M) are per-property data; other registered tables are economics instructions; unregistered tables are housekeeping (projects, graphs, filters, units, locks). "Run output" = AC_ONELINE, AC_ECOSUM, AC_MONTHLY, AC_DETAIL — the tables ARIES writes results into.

The case (Q2): a master field counts as filled when it is non-blank after mdb-export; partial counts are wells with the field blank. Economics lines are AC_ECONOMIC rows grouped by SECTION; the section names are ARIES's. "Same skeleton" = keyword set per section is identical across cases except the exceptions named. The scenario table routes every section to qualifier 2026Q3 (section 4 also lists a fallback qualifier, ALCHEMIST, which has no lines).

The history (Q3): monthly = AC_PRODUCT, keyed by P_DATE month; daily = AC_DAILY summed to the same well-month by D_DATE. A well-month "ties" when oil differs by ≤1% (or ≤0.5 bbl); ">10%" is on the larger of the two. The chart marks a group-month when any well in the group is >10% off. Coverage lines are first→last row per group; the notch is the month with no monthly row. The leap-day duplicate is two AC_PRODUCT rows for one well dated 02/28/24 and 02/29/24 with identical volumes. Forecast: AC_FCST rows by QUALIFIER; the live forecast is the §4 8-word expressions (initial rate, units, limit, method, b, secant-effective annual decline) with a START line per phase.

The audit (Q4): headline tie-out = sums of the oneliner's 20 rows vs the PDF grand-total page (p.12); net revenue compares against the PDF's TOTAL NET REVENUE row (its metrics block prints the oil-only figure, 5,767.620, on the same page). Interests ran = AC_ECONOMIC section 7 NET lines (words: WI%, net oil %, net gas %; ditto lines are reversion steps). Master NRI = AC_PROPERTY.NRI × 100. Disagreement threshold 0.005 pp; 13 of 20 wells tie.

The valuation (Q5): per-well PV10, net volumes and net revenue from the oneliner export, grouped by its ASSET_GROUP column; group PV10 totals cross-checked against the PDF's per-group pages (Horton p.1, Kale p.3, Dreadnaught p.5, Burton p.7, Shooter p.9) — all five tie. PV10 = discounted cash flow at 10%, M$ = thousands of dollars; shares are of the 2,254.8 M$ total. The "had the master NRI been used" figure is the Shooter group's net revenue (66.025 M$) scaled by 180.7, the ratio of master NRI to the interest that ran — a linear scaling of revenue, not a rerun; costs sit on the working interest and would not scale with it.

What we'd push back on ourselves: this is a demo database, and both catches — the 180× metadata mismatch and the two histories disagreeing — may well be planted. But they are exactly the classes of error this workflow exists to catch in real books, where the master table drifts from the economics across quarters of revisions and a daily feed and an allocated monthly feed stop agreeing. We do not know which history is right; the daily table has the more plausible October 2024, but "more plausible" is not evidence — the state filings for these public wells would settle it, and a connected session can run that check. Our tier labels are a reading of ARSYSTBL, not ARIES documentation. And we keyed the PDF values in by hand from the printed report, page references given.