The engine at work

Defense contract records, read blind against an audit

A blind read: the engine was given the records only, never the audit. Read 2026-10-26, LEF SI Engine 0.62.0, with no AI model in the read. It goes to no one, and no customer has it.

The records. A per-day sample of the Department of Defense's contract actions for fiscal years 2010 to 2014, from USAspending.gov's public archive: 20 actions from each of 1,826 days (36,520), and the 12,661 sampled orders whose parent contract is in the same files, with those 6,316 parents.

The test. The Government Accountability Office reviewed DoD's contract data for those years (report GAO-15-484R). The test was locked before any file was opened: two things GAO says about how these records are kept, which the engine should find, and two fields it should leave alone.

What happened

• Found. GAO says the competition data for a task order is carried from its underlying contract. The engine named that field as carried from the parent contract in all 11,608 orders that record it alongside their parent, where chance alone would give about 45%.
• Left alone, correctly. It did not flag the fields GAO relied on, nor an order's own dollar amount, which GAO never says is carried from the parent.
• Missed. GAO ruled one-offer data before 2013 too unreliable to compare. The engine did not find it: the field is filled in nearly every record every year and its most common value never changes, so the problem lies in what the values say, which a check of how a field is kept cannot see.

GAO did not review the engine. A match shows the engine reads the record the way an auditor did, nothing more. These are counts in a sample of the archive as of October 2026, never DoD's totals.

The national dam inventory, read against its own dictionary

Read 2026-10-26, LEF SI Engine 0.62.0 and 0.62.1, with no AI model in the read. It goes to no one, and no customer has it.

See how this build works, as a flow chart

The records. The National Inventory of Dams, the public file the U.S. Army Corps of Engineers keeps from state and federal submissions: all 92,693 records, data dated 23 October 2026. No dam name, owner, place or coordinate was read.

The test. The engine read the records alone first, and that read was frozen before the inventory's data dictionary was opened. Then it read them again, with the dictionary.

What happened

• No false rule. The dictionary states five rules joining a dam's operational status to its condition rating, and the records break all five (494, 163, 17, 24 and 137 records). Read alone, the engine held none of them as a rule: none of its 363 gaps joins those two fields.
• Every disagreement named. Given the dictionary, the engine named all five places the records and the dictionary disagree, each with its count, and used none of them to explain anything away. It judges neither side.
• A proposed closing for each. A closing tool then proposed how each of the 835 records would be closed: for 624, the record's own dated rating points at its undated status; for 211, the record cannot tell, so both are offered. Each is a question for the record's owner, and none changes the record. A date shows a field was set recently, not that it was set right: where the dated field is the one in error, the proposal points at the other.
• Missed at first. On the first engine version we ran, the report printed four of the five disagreements without saying a fifth existed. Fixed the same day.

Statements about how a file is kept as of 23 October 2026, never about the condition or safety of any dam.

Defense vehicle contract awards

A test build: it goes to no one, and no customer has it. Read 2026-10-26, LEF SI Engine 0.62.1, with no AI model in the read.

See how this build works, as a flow chart

The records. 5,000 Defense Department contract award records for combat, tactical and wheeled vehicles (fiscal 2016 to 2026) from USAspending.gov, a public federal source, drawn as a fixed sample of the awards held locally. The build was made for an Army ask; the Army awarded part of these contracts, not all.

What the build does. The engine's own read, with no tool of its own: which combinations of agency, supplier, industry code, place of performance and award type never appear together, each checked against every record.

What the engine found

• The engine read all 5,000 records.
• It looked for pairs of common values that never appear in the same record and found 87.
• The records themselves explain 23 of them. 39 are too thin or too uncertain to stand on, and are leads only. 25 hold in these records.

10 of the engine's checks could not run on these records, each for a reason the run states.

A hospital file, read for a quality director

A test build: it goes to no one, and no customer has it. Read 2026-10-26, LEF SI Engine 0.62.1, with no AI model in the read.

See how this build works, as a flow chart

The records. CMS's public Hospital General Information file: 5,419 hospitals, their star ratings, measure groups and footnotes.

What the build does. A peer comparison for each hospital: how its measure groups stand against hospitals of the same ownership, type, emergency services and state, using only the counts and footnotes the file itself publishes, and saying where the file is silent.

From the build’s own read. Footnotes are common in this file: the overall rating carries one for 42.7% of hospitals, and the patient-experience and timeliness groups publish no better-or-worse counts at all, so the brief says so instead of comparing them.

What the engine found

• The engine read all 5,419 records.
• It looked for pairs of common values that never appear in the same record and found 1,282.
• The records themselves explain 823 of them. 8 are too thin or too uncertain to stand on, and are leads only. 451 hold in these records (9 of those are the same gap told again through a field carrying the same information, so 442 distinct).

17 of the engine's checks could not run on these records, each for a reason the run states.

Hazardous materials incidents

A test build: it goes to no one, and no customer has it. Read 2026-10-26, LEF SI Engine 0.62.1, with no AI model in the read.

See how this build works, as a flow chart

The records. 1,500 rows of PHMSA's hazardous materials incident reports from January 2022 (PHMSA is the federal pipeline and hazardous materials safety agency); they cover 1,464 reports, because one report can take more than one row.

What the build does. An atlas of which ways of shipping (mode, hazard class, packaging) go with injuries or fatalities in these reports, and an inventory of which of the report form's fields are ever filled.

From the build’s own read. Of the 1,464 reports from January 2022, 31 distinct dates, the configuration with the most reports of injury or fatality is Highway / class 8 / (not recorded): 2 with an injury, 0 with a fatality, of 25 reports. These are counts in these records, not rates.

What the engine found

• The engine read all 1,500 records.
• It looked for pairs of common values that never appear in the same record and found 5,107.
• The records themselves explain 1,981 of them. 2,661 are too thin or too uncertain to stand on, and are leads only. 465 hold in these records (34 of those are the same gap told again through a field carrying the same information, so 431 distinct).

14 of the engine's checks could not run on these records, each for a reason the run states.