Lab / Synthetic data
Finance variance review
A small workbench for reviewing plan-to-actual exceptions, checking their source references, and recording a human decision.
Prototype boundary: every record is synthetic. Variances are calculated by deterministic code; the explanation is fixture text, not model output. There are no uploads, external requests, persistent storage, or accounting-system writebacks.
Sample close
September 2026 · Sample operating unit
- Exceptions
- 3
- Needs review
- 3
- Reviewed
- 0
- Net variance
- +$14,850
Review queue
Materiality is not inferred here. Reviewers decide what requires action.
SYN-1042
Vendor services
Operations · Sep 2026
- Plan
- $42,000
- Actual
- $54,600
- Variance
- +$12,600
- Against plan
- +30%
Context in fixture
Additional service days were included in the synthetic activity detail.
Rule-based context, not an AI-generated explanation.
Source references
- SYN-GL-1042
- SYN-ACT-308
IDs are synthetic fixtures; no source system is connected.
No records match this filter.
Select a record to review. Changes remain in this browser session.
Design boundary
The prototype demonstrates a review flow: calculate plan-to-actual variance, expose the fixture context and source identifiers, then let a person mark a record reviewed or escalate it. The values, teams, periods, and source IDs are invented for this demo.
It does not call an LLM, connect to finance systems, persist decisions, establish accounting policy, or measure a user's time saved. A real deployment would need permission-aware source integration, materiality rules agreed with finance, audit persistence, security review, failure handling, and an evaluation set reviewed by domain experts.