Learning lab
From mismatch to next step
A working demonstration of how a reviewer moves from a discrepancy between two systems to a defensible next step. All records and all procedures here are fictional and built in. Nothing on this page is a federal rule, and nothing can be uploaded.
Pick a scenario
Six built-in fictional cases.
What is fictional here
Fixture set lab-fixtures-1.0.0.
- The two systems, the record ids, the amounts and the statuses are invented.
- The procedure passages are invented and are labelled FP-xxx to make that plain.
- No record is ever modified by this lab, by code or by the model.
- Feedback you give below stays in this browser tab and is not saved anywhere.
Amount mismatch: 1,000 versus 800
Ledger A and Ledger B both hold record R-1001, but the amounts differ. What is the next step?
Both systems agree the record exists and is posted. Only the amount differs.
Step 1. Deterministic comparison (no AI)
Records are matched by id, amounts compared as integer cents, duplicate ids rejected, one-sided records reported in both directions.
| Record | Fictional Ledger A | Fictional Ledger B | Difference | Finding |
|---|---|---|---|---|
| R-1001Charge line | $1,000.00posted | $800.00posted | $200.00 | Amount differs |
Compared 1 record(s). Matched 0. Paired-record absolute difference $200.00. Unmatched amount (unconfirmed, not a paired difference): $0.00. Every figure on this table is computed in code from integer cents, with no model involved.
Step 2. Retrieved fictional procedures (no AI)
Method actually used: applicability predicates, then metadata tag match plus literal keyword overlap over 6 fictional passages (no embeddings, no vector index, no reranking).
FP-101 Fictional Procedure 101: amount variance under one thousand units
When two systems report a different amount for the same record id and the absolute variance is under 1,000 units, the reviewer opens a variance note, records both amounts in cents, and asks the source system owner which figure was entered last. No value is edited by the reconciliation tool itself.
Matched tags: amount_mismatch. Score 15.
Step 3. Explanation or rule-based escalation
Conflicting or missing evidence produces a rule-based result without calling AI. Normal-case AI explanations require a shared usage reservation before calling the model.
Shared usage limits: at most one call per click, one call per visitor every 20 seconds, and a conservative global cap of 200 model calls per day for the whole lab.
This lab is a teaching demonstration on synthetic data. It is not a reconciliation system, not official guidance and not connected to any institutional record. It shares this site with the Schedule of Reductions calculators but has no access to their engine or their data.