
Why this caught my attention
My previous Gemini article left a question open: when a number changes, is it a new reporting period or a correction to an old one? I would give Hydrogen one bounded job: compare two public reports from one shipping company and prepare a proposed update to a revenue-and-operating-profit table.
How I would use it
I would supply the reports myself and request one row per figure: metric definition, reported value, currency, scale, period, publication date, source URL and page, plus the exact supporting passage. Missing evidence should produce a blank value and an explanation. Originally reported and restated figures would occupy separate rows.
In Excel, I would preserve the raw extraction, normalize units in a separate sheet and recalculate operating margins from verified revenue and operating profit. A changed definition or a mismatch with the source would stop that row from entering the model. I would compare the two releases using company, metric definition and period together, rather than matching labels alone.
There is a reason to test this before assuming efficiency. Vals reports Hydrogen at 65.12% on Finance Agent v2.1, versus Argon's 67.66%; their recorded costs per test are $4.36 and $4.32 respectively. These are benchmark observations, not my costs or the probability of getting my spreadsheet right.
Vals uses weighted partial credit across 449 held-out questions, a common six-tool setup and three LLM judges. This differs from All-Pass, which requires every check to pass. Missing a critical condition gives a question zero credit. The v2.1 scores cannot be compared directly with v2. The page displays November 11, before the public launch; I cannot establish why, so these are results observed on November 14, not a verified release-day test.
Where I want to take this
I would track incorrect periods, missing evidence, silent revisions, total API cost and review minutes. A model that produces more rows but needs more correction may not improve my workflow. Only after a manual trial would I consider n8n for collection and notifications, with approval before changing the model. That integration is a proposed design, not a tested capability.
The next question is concrete: how should the data dictionary retain both the original number and a later correction without rewriting the history of my analysis?
Sources
- Introducing Gemini 4 Hydrogen — Google · Published 2026-11-12; updated 2026-11-13; accessed 2026-11-14
- Finance Agent v2.1 — results and methodology — Vals AI · Published Page displays 2026-11-11; observed 2026-11-14; accessed 2026-11-14