Liquidity / Execution
Tests executable depth, route impact and whether the proposed size can clear without moving the market beyond the committed bound.
Each attacker has a specialist method, a separate run context and a typed tool grant. They share the committed input, not each other's work.
Resolved against the live OpenRouter and Groq catalogues. Substitutions are recorded, not silent.
| ID | Method | Provider | Model | Family | Open weight |
|---|---|---|---|---|---|
| A01 | Liquidity / Execution LIQUIDITY_EXECUTION | openrouter | z-ai/glm-5.3-flash | glm | Not claimed |
| A02 | Asset Safety ASSET_SAFETY | groq | openai/gpt-oss-120b | gpt-oss | Yes |
| A03 | Logic / Coherence LOGIC_COHERENCE | openrouter | google/gemini-2.5-flash | gemini | Not claimed |
| A04 | Correlation / Portfolio CORRELATION_PORTFOLIO | openrouter | dots-studio/dots-3-note-preview:free | dots | Not claimed |
| A05 | Thesis Inversion THESIS_INVERSION | openrouter | qwen/qwen3.8-27b | qwen | Not claimed |
| A06 | Data Provenance DATA_PROVENANCE | openrouter | recursive/flint-v1.0:free | flint | Not claimed |
| A07 | Timing / MEV TIMING_MEV | openrouter | recursive/flint-v1.0:free | flint | Not claimed |
| A08 | Base Rate BASE_RATE | openrouter | particle-ai/particle-1:free | particle | Not claimed |
Specialists differ by method and source diet, not by temperature.
Tests executable depth, route impact and whether the proposed size can clear without moving the market beyond the committed bound.
Reviews contract permissions, token behavior and asset-specific failure modes that the thesis may have treated as background.
Checks whether the thesis follows from its evidence, whether terms are used consistently and whether the conclusion is overclaimed.
Looks for hidden exposures, shared dependencies and correlated drawdowns across the proposed position and the wider book.
Builds the strongest case against the thesis, then asks what observable evidence would make that inversion correct.
Compares the source, timestamp and derivation of the numbers the thesis relies on. A different source name is not independence by itself.
Examines the execution window, transaction path and adversarial ordering risk around the proposed action.
Tests the thesis against historical base rates, survivorship effects and the reference class the agent implicitly chose.
Enforced at the run boundary.