02 · Evidence
Every rule can say where it came from.
The Standard's specs are read by people, so they don't mention paragraphs or job steps. A separate evidence ledger holds that link, and AI can do most of the reading, as long as a person confirms what it found.
The ledger
A claim is what it does, not what it should do.
Keep the ledger's language flat. "The fee calculation drops fractions of a cent" is evidence. "The fee should be truncated" is a decision, and it belongs to the process owner.
That line is what makes AI extraction safe. A model can read a million lines and propose claims all day, because a proposed claim commits nobody to anything.
Specs stay clean, too. The ledger lives in migration/evidence/, beside domains/ and apps/, and it can be removed when the source is gone.
## EV-014 — The late fee drops fractions of a cent - Claim: The fee is 5% of the amount due, and anything past two decimals is discarded rather than rounded. - Source: ACLP210 2200-ASSESS-FEE (no ROUNDED on COMPUTE) - Kind: code - Extracted by: AI — 2026-11-05 - Status: confirmed - Confirmed by: Marcus Bell — 2026-11-05 - Disagrees with: The fee schedule says "5%", no rounding stated - Supports: LPF-003
Where to look
The main program is only one witness.
Tidewater's late fee and its follow-up rules were scattered across six kinds of source. Reading only the COBOL would have missed most of them.
| Source kind | Found at Tidewater | Would be missed by |
|---|---|---|
| Program | 5% fee and $10 minimum, in ACLP210 | Nothing; everyone looks here |
| Sort card | Guaranteed loans skip follow-up, in INCLUDE COND | Reading only the COBOL |
| Code table | Each bank's grace days, in TWCODES | Reading only source libraries |
| Server config | 10-day quote validity, in quote.properties | Reading only the repository |
| Stored procedure | A loan needs a zero balance to close, in SP_CLOSE_LOAN | Reading only the table definitions |
| Data profile | An undocumented status H on 41 loans | Reading only code and comments |
Rules
MIG-005 to MIG-008
- MIG-005
Every proposed rule cites at least one evidence entry.
- MIG-006
A claim extracted by an AI stays
proposeduntil a named person confirms it. - MIG-007
Evidence records what the system does; only a person decides whether it should keep doing it.
- MIG-008
Where sources disagree, production behavior is the evidence of what happens, and the disagreement becomes an open question.
## EV-nnn — <short claim> - **Claim:** <what the system does, in plain words; no judgment about whether it should> - **Source:** <artifact and lines, for example ACLP210 2200-ASSESS-FEE, or a job log and date> - **Kind:** code | config | sort card | stored procedure | data profile | observed behavior | interview - **Extracted by:** <person or AI model> — YYYY-MM-DD - **Status:** proposed | confirmed | disputed (MIG-006) - **Confirmed by:** <named person> — YYYY-MM-DD - **Disagrees with:** <other evidence, or none>; if so, open a question (MIG-008) - **Supports:** <rule ID, or "out of scope" with the reason (MIG-039)>