Method

How a grade is reached.

Every rule that can fire, written down before it fires. A rule nobody can read is a rule nobody can argue with.

The one idea underneath all of it

A public dollar does not exist in one state. It passes through stages, and each stage has its own record and its own figure. Almost every misleading money claim is a real figure quoted at the wrong stage — which is why a flat true-or-false stamp gets it wrong in both directions.

1 Authorized A ceiling a legislature or board voted. Nobody has promised it to anyone. Where it is written down: appropriation act, budget ordinance, board resolution authorized, appropriated, budgeted

2 Obligated Committed under a signed contract or grant award. Money is promised, not gone. Where it is written down: contract register, grant award notice, obligation report obligated, awarded, committed

3 Disbursed Actually paid out. The only rung that means the money left the account. Where it is written down: disbursement ledger, warrant register, check register, payment file spent, disbursed, paid

4 Received Landed with a named recipient. Not the same as disbursed when there is a pass-through. Where it is written down: recipient award record, subrecipient report, single audit received, collected, got

The gap between the top rung and the bottom rung is not waste or a scandal. It is normal: money is authorized long before it is paid. The gap only becomes a story when someone describes one rung using the verb belonging to another.

The rules, in the order they are tested

The worst grade any rule reaches is the grade of the claim. Press "Show me why" on any decomposition and you get exactly this list, filtered to the rules that fired.

Rule	Reaches	Fires when
R1	Unsupported	No record in the ledger covers the claim’s scope at any stage.
R2	Misleading	The claim names one stage and the record at that magnitude is a different stage.
R3	Misleading	The figure is more than 10% away from the nearest record at any stage.
R4	Misleading	A ratio word ("doubled") is more than 15% away from the ratio the records show.
R5	Compressed	The figure matches, but it is the sum of two or more record lines.
R6	Compressed / Misleading	The claim gives a figure but never says which stage it reached.
R7	Accurate	The figure matches a single record line at the stage the claim names, within 3%.
R8	Accurate	The ratio word holds against the records.
R9	Unsupported	A comparison is asserted and the ledger holds fewer than two dated figures to compare.
R10	Misleading	The claim asserts a cause and no attribution record is attached.
R11	Unsupported	Nothing in the ledger speaks to the claim at all.

A figure this close to a single record line is stating that record.

10%

Rounding, or a merge that a fair reader would accept.

15%

How far a ratio word may sit from the ratio the records show.

What this engine will not do

It will not decide who is lying.

A grade describes the distance between a sentence and a record. Intent is not in the record, so it is not in the grade.

It will not invent a record.

If the ledger is silent the verdict is Unsupported, and "unsupported" is a statement about the ledger, not about the claim being false.

It will not carry a cause on a figure.

A number can show that something changed. It cannot show what caused it. An attribution needs its own record or it is named as missing.

It will not call a model.

There is no model and no network call in the grading path. The same file runs here and in the build gates, so a grade cannot differ between what we test and what you see. You can check that yourself.