The abstract shows the original rent, not the rent being paid

Accounts is billing one figure and the abstract holds another. Before anyone looks at the model, subtract them: if the difference is exactly one escalation step, one negotiated reduction or one deferral, this is a document-set problem and the extraction did its job on the wrong pile of paper. A ragged difference with odd decimals is a different investigation entirely.

That distinction matters because the two fixes touch different systems. A stale value means an instrument is missing or was combined wrongly, which is intake and merge logic. A garbled value means the number was misread, which is extraction. Teams routinely spend a fortnight on the second while the first is sitting in an unopened folder.

Count what the chain says exists against what you actually hold

One count settles it, and it takes about 15 minutes per lease. The point is to separate 2 states that look identical from the outside: the amendment was never ingested, or it was ingested and the merge lost it.

1. Open the most recent amendment you hold and read its recitals. Amendments almost always name the ones before them, so Amendment 4 tells you that 1, 2 and 3 exist even if you have never seen them.
2. Count the files in the lease record against that number. Fewer files than the chain recites is a missing document, and you are done diagnosing — the rest of this page is about the case where the count matches.
3. Check for instruments the chain would not recite. Rent deferral letters, waivers and side letters are frequently unnumbered, filed as correspondence, and never mentioned by any amendment.
4. Ask which file each abstracted rent value came from. A per-field citation makes this instant; without one, this step is a manual read, which is one reason a citation that will not resolve is treated as its own defect in the cited page that does not contain the clause.
5. Confirm the schedule you compared against is the whole schedule. A rent table that stops 3 years early looks like a stale value and is not one — it is the structural failure in the rent schedule that lives in an exhibit.
6. Write down which of the 2 states you are in. Missing document routes to re-ingest. Complete set routes to re-merge. Everything below hangs off that answer.

5 reasons the merge produced last decade's number

Rows 1 and 2 are intake failures and end in a re-ingest. Rows 3 to 5 are merge failures with every file already present, and they are the ones that recur, because nothing about the folder looks wrong afterwards.

Everything is in the folder and the merge still lost

Ordering is where complete document sets go wrong, because the obvious keys are all wrong. Upload order reflects who scanned what. File name order sorts Amendment 10 before Amendment 2. Execution date is close but breaks whenever a deed is signed after the change it makes has already taken effect, which is common on rent concessions agreed mid-quarter and papered later.

Effective date is the correct key, and it has to be extracted as its own field rather than inferred from the signature block. A lease already carries several dates that are routinely collapsed into one — the separation argued in the three start dates a lease actually has — and an amendment adds its own. Where 2 instruments share an effective date, fall back to execution date, then to the recital chain, and record which rule decided the outcome.

Contradictions are the case nobody designs for and every long chain contains. Two amendments that both purport to set the rent for the same year, or an amendment referring to a document that was never signed, cannot be resolved by ordering alone; they are resolved by a person, and the system's job is to surface them rather than to pick. The full reconstruction of a chain that behaves like this is worked through in a retail lease with eleven amendments.

The folder was complete, the extraction was accurate, and the number was 4 years old. Nothing in the pipeline was wrong except the order it read things in.

Re-ingest or re-merge, decided in 4 questions

1. Does the file count match the recited chain? No — stop here and re-ingest. Nothing downstream can be assessed until the missing instrument is in the record.
2. Does any abstracted rent value cite a file older than the newest instrument? Yes — re-merge. The newer document was present and was not applied.
3. Does the same lease produce different rents on 2 runs? Yes — re-merge, and look for 2 files that both claim to be the whole agreement, because a non-deterministic result almost always means an unresolved tie.
4. Is the value right and the date it starts from wrong? Yes — re-merge after fixing date extraction. This one is worth isolating because it produces correct-looking rent with incorrect arrears.

Re-ingest is cheap and one-off. Re-merge is the one worth engineering, because it will run again after every future amendment, and a merge that cannot be re-run safely over an existing record turns each new instrument into a manual reconciliation. Building that as a repeatable operation rather than a rescue is ordinary internal tools and operations software.

What the corrected rent still does not prove

One repaired lease is one repaired lease. The same fault is almost certainly present across the portfolio, silently, on every lease with a chain — and the detector is not the abstract, it is the comparison against what is being billed, which is why a rent roll that does not tie to the leases is the standing check rather than a one-off exercise.

The durable fix is in the schema. A field set that stores rent as a value, with no room for the instrument that set it, the date it took effect and the instrument it superseded, will drift again after the next amendment however carefully this one was repaired. Designing for lineage before extraction begins is the argument in designing the lease field schema before you extract anything, and it sits with the rest of lease abstraction and property document AI in our work with real estate teams.

• amendments
• lease abstraction
• rent
• diagnosis