A deterministic-first reconciliation product: rules for clear records, evidence and human review for the exceptions.
Share of processed records resolved by rules before LLM routing.
Definition & measurement context ↓Researchers comparing monthly FDA Orange Book updates across fragmented sources.
Manual spreadsheet comparisons made recurring updates slow and hard to inspect. Ambiguous records needed judgment, while clear records did not need a model call.
In six discovery sessions, I mapped how researchers compared spreadsheet sources and handled records that deterministic matching could not resolve. I used those findings to define the PRD, acceptance criteria and exception-review workflow.
I separated the high-volume matching path from the cases that needed interpretation. We standardized source attributes, applied deterministic logic to clear records, and routed unresolved cases to an LLM with a human exception-review path.
Use rules for clear cases and reserve model reasoning for unresolved records. Stable cases did not benefit from probabilistic reasoning; ambiguous cases needed evidence-aware interpretation.
I used rules for clear records and routed ambiguity to model-assisted human review.
The proposal would treat these as the same ingredient, strength and dosage form across files. The visible strength disagreement prevents acceptance.
Tablet · Oral
10 mgTablet · Oral
20 mgConflict detectedA review note can explain a decision. It cannot correct a source field.
Use a note for investigation context. Any length of note leaves the conflict unresolved.
A 10 mg record and a 20 mg record cannot be accepted as the same product presentation. Reject the visible mismatch or inspect its provenance.
This newly designed portfolio demo illustrates the human-review boundary. Reject the mismatch immediately or inspect the local sample extracts and escalate it. No note can resolve the conflict or accept the link. It does not reproduce verified historical production controls.
Project-reported outcomes are separated from discovery counts and program scale. Definitions and measurement limits are included below.
Share of processed records resolved by rules before LLM routing.
Automated processing time, separate from human exception review.
Reported token-consumption reduction after deterministic-first routing.
End-to-end review time must include human exception work. It is not established by automated runtime, and deterministic coverage does not establish matching accuracy.
The next validation priority is a reproducible matching-quality evaluation and an end-to-end time study that includes human exception review. Pipeline speed and deterministic coverage answer different questions.