Seven adversarial drill families run against your agent before it meets size: prompt injection, regime fabrication, drawdown spikes, stale-data traps, compromised-host scenarios, and more. The agent that passes has earned its envelope.
§1Capabilities
Stress-tests the decision layer with hostile inputs an agent will eventually meet: instructions embedded in market commentary, fabricated regime signals, data feeds that quietly stop updating.
Verifies the safety primitives actually bind — that a triggered killswitch really produces hold, that envelope breaches really refuse, under adversarial pressure rather than happy-path testing.
Produces a drill report an allocator or risk committee can read: what was thrown at the agent, what it did, where it bent.
§2Design
Seven families. D1 injection through D7 compromised host — each family is a corpus of payloads plus a scoring rubric, run in-process against your agent’s real prompts.
Pre-capital gate. Designed to run before deployment and after every material prompt or strategy change — conformance as a repeatable check, not a one-time review.
Custom drills. Institutional licensees get quarterly corpus updates and drills built for their specific strategy surface.
§3Availability by tier
| Open | D1-LITE sample via the Prompt Library (MIT). |
| Pro | Not included — the harness is an Institutional product. |
| Institutional | Internal license, full corpus, quarterly custom drills. |
Indicative until launch.
§4 · Stated plainly
Passing drills reduces failure modes; it does not certify profitability or eliminate risk. No audit can.