Audit Day
An exchange auditor and a defending agent go through three years of one firm's market data use. Watch what it owes, why, and what would have prevented it.
A venue auditor agent and a defence agent work through a synthetic firm's entitlements, usage logs and declarations. Findings are raised with row-level evidence, priced by deterministic code, defended or conceded, and settled. Then the same three years replay with monthly reconciliation switched on.
How this was built
When an exchange audits a firm's market data use, it looks back three years. Every leaver who kept a login, every shared password, every application declared in the cheaper category becomes a back-fee, with interest and a penalty on top. Firms rarely know how exposed they are until the letter arrives. AI agents make this harder: they consume real-time data at volumes no person does, and venues are still writing the rules for them.
This lab stages the audit. Harbor Ridge Capital, a fictional 300-person asset manager, is audited by Meridian Exchange, a fictional venue. An Auditor agent works through a seven-test audit programme and raises findings, citing the rows that support each one. A Defender agent, acting for the firm's market data provider, searches the same records for evidence that clears each finding, or concedes and names the control that would have caught it. When the Defender clears a finding, the Auditor checks the defence and either withdraws the finding or says why it stands.
The agents argue; plain code counts. Every dollar comes from an exposure engine that applies the fee schedule to the data the same way every time, and every round is checked against an answer key written with the data: five problems are real, and two have evidence that clears them. The Protected switch replays the same three years with a monthly reconciliation, which catches most problems in the month they start. Everything is synthetic, built on a personal AWS account, and runs on Kimi K2.5, an open-weight model in Amazon Bedrock, chosen after trying four candidates on the same audit.
Exploring how agents can protect a firm in a market data audit, November 2026. Runs on Kimi K2.5 in Amazon Bedrock.
A personal project, built on a personal AWS account using personal equipment. Not affiliated with my employer.
This demo is interactive and needs JavaScript.