Trade Surveillance Agent
Docs

Research prototype

Synthetic firm policy, not a supervisory system of record, and not legal or compliance advice. See the regulatory basis.

Insider-trading surveillance

Which insider trades deserve an investigation?

Company insiders must publicly report their stock trades. Compliance teams watch those filings for the few that might be trading on secret information.

Flag everything and analysts drown in false alarms. Clear too easily and a real violation slips through. This demo shows an AI investigator working through sample filings, and the checks that stand between its draft and a published conclusion.

Read how it works in plain language

Try it yourself

Which of these filings deserves a second look?

Two insiders, the same sale, the same day, shortly before the same announcement.

Filing A

Insider
Chief Financial Officer
Trade
Sale of $400,000
Date
Jul 18, 2025
Company news
Earnings on Jul 20, 2025 (2 days later)
Blackout window
Trade falls inside it

Filing B

Insider
Chief Financial Officer
Trade
Sale of $400,000
Date
Jul 18, 2025
Company news
Earnings on Jul 20, 2025 (2 days later)
Blackout window
Trade falls inside it
Which would you send to an analyst?

Three ways surveillance goes wrong

  • Crying wolf

    Flag every sale made before earnings and analysts drown in alerts, including sales reported under a plan set up in advance.

    The safeguard: The filing's reported 10b5-1 status is read from the data and shown as a signal. It is never guessed from its absence.

    See it in a case

  • One excuse covering everything

    Tax withholding is routine. If one exempt line clears the whole filing, a very large sale in the same filing goes unnoticed.

    The safeguard: Exemptions are tied to the specific transaction line they belong to.

    See it in a case

  • Confident but unsupported

    An AI can write a convincing conclusion that its evidence does not back up. In the recorded live run on case 1, 2 of 4 claims were judged unsupported, and the result is marked for human review.

    The safeguard: The draft is split into claims, each checked against the evidence gathered, and a process audit records whether claims cite real tool calls.

    See it in a case

Explore sample cases

Each case is a synthetic filing with a known right answer. Open one to see the evidence, the investigator's conclusion, and what the safeguards said about it.

Worth an analyst's time

Missing any of these is how a real violation slips through.

Should not be alarms

Flagging any of these is a false alarm, and false alarms are what bury analysts.

Under the hood

How it works

  1. 1PlanAI

    Decides which evidence the filing needs.

  2. 2Check the planCode

    Confirms every required check is planned, in order.

  3. 3InvestigateAI + code

    Calls read-only tools over real SEC filings and a seeded firm-policy database.

  4. 4Draft and extract claimsAI

    Writes a conclusion and splits it into checkable claims.

  5. 5Grounding checkAI + code

    A judge scores each claim against the evidence; code does the totals.

  6. 6Publish and auditCode

    Publishes the exact judged draft, then audits how the investigation worked.

How we know it isn't cutting cornersInvestigate a filing yourself