LLM Council
BuiltSignature capability
Instead of one sycophantic model, a panel of named personas with explicit priors debates each decision, and a mandatory devil's-advocate forces the uncomfortable hypothesis into the open.
What it is
The council reviews high-stakes decisions in my trading system (a refit rejection, a calibration-drift lesson, a promotion block, a new-strategy discovery) by running the same structured context past several named personas, each with an explicit prior and a distinct failure-mode lens. It aggregates their verdicts into one recommendation: promote, hold, reject, or investigate. A devil's-advocate persona is mandatory, because it is the only reliable way to surface the hypothesis the other personas avoid. Verdicts are strictly advisory: the council is a gate input and never opens or closes a trade. Each persona call is routed through the Pulse LLM router and fails open to a deterministic “investigate” verdict if the endpoint is unreachable, so a network blip never silently passes a bad decision.
Highlights
- One persona per failure mode (stat-arb, macro, behavioural), each with explicit priors, to force reasoning the single-LLM path skips.
- Mandatory devil's-advocate persona to surface the self-deception hypothesis the others avoid.
- Verdicts are advisory-only — the council is a gate input and never places an order.
- Fail-open by design: a transport error degrades to a deterministic “investigate” verdict and logs the cause.
- Two implementations behind one Protocol: a deterministic NullCouncil for offline/testing and an HTTPCouncil routed through Pulse.
- All persona prompts live in code, so adding a persona is a reviewed change that goes through tests.
Tech · Python (typed Protocol design), Pulse LLM router (HTTP), named-persona prompt panel, deterministic NullCouncil for offline/testing, fail-open transport, strict-TDD test suite.