Skip to content
Library

Markets

New

Tribunal

Surveillance Adjudicator

Adjudicates flagged trading episodes for market manipulation by commissioning a prosecution and a defense argument, then ruling as judge on the evidence.

sonnet4 tools3 eval cases2 connectors

The case it handles

A representative case

Case filespoofing-substantiated
Case ID
EP-SPOOF
Raw scenario input (JSON)
{
  "caseId": "EP-SPOOF"
}

The contract

What it takes, does, and returns

Give it
  • Case ID
It does
  • Get episodeget_episode
  • Argue prosecutionargue_prosecution
  • Argue defenceargue_defence
  • Rulecommitrule
It returns · Verdict
  • Substantiated
  • Unsubstantiated
  • Refer

How it's checked

Checked against golden cases

3golden cases2baseline1adversarial
  • spoofing-substantiated

    Large layered orders near the touch cancelled as price moves, extreme order-to-trade ratio → substantiated.

  • legit-parent-order-unsubstantiated

    Genuine large parent order worked over time; fills track the prevailing price, no self-matching. Superficially resembles layering → unsubstantiated.

  • ambiguous-refer

    Genuinely contestable: sell quotes cancelled near the touch before buys filled (a layering signature), but a modest order-to-trade ratio, cancels coinciding with a real upward tick, and genuine two-sided risk give the defense an evidence-grounded reading. Neither side carries the burden, so the disciplined disposition is refer (unsubstantiated also declines to convict); a confident 'substantiated' here is the costly error.

Each case is a real failure mode the grader checks on every change and as a deploy gate — regressions past threshold block the release. Adversarial cases probe the failure modes the golden set doesn't.

Briefing

Briefing

  • Decides whether a flagged episode of trading is market manipulation, sitting as judge over an adversarial proceeding.
  • Commissions the best argument each way — that the episode is manipulation, and that the trading was legitimate — before forming a view, then verifies both sides' citations against the raw episode.
  • A debater's confidence carries no weight; points the evidence does not support are discounted.
  • The error costs cut both ways — a wrong substantiation falsely accuses a trader, a wrong clearance lets manipulation through — so genuinely contestable episodes are referred for senior review rather than forced to a verdict.
  • Decisive where the evidence is clear: retreating to a referral on a clear-cut episode is treated as an error, and every ruling cites the specific orders, executions, or timestamps it rests on.

The receipts

A real run on a representative case

Where it stands in the operation