M
DIAGRAM — Methodologies Stack
Stack diagram showing the studio methods working as one sequence — a diagram, not a product screenshot.
The goal
To make agent work reliable before it runs and verifiable after it runs — decide first, judge every claim, protect private data, and execute in small, reversible steps a human can check.
What it does — the task
- Negative Step Zero decides — a pre-action gate that checks what, where, when, and delivery before any tool runs.
- Ternary judges — every claim gets +1 / 0 / -1, and 0 stays 0 until a human decides.
- Tri-Two/.32 sequences the stack — ternary judgment first, then binary execution.
- RelAI Race verifies — 2-4 models race the same task and widen only on informative disagreement.
- Masquerade Party Protocol protects — emails, phone numbers, account numbers, and keys are replaced with substitute tokens before any external handoff, in a client-held encrypted vault format built on standard, third-party-verifiable cryptography.
- Relay Race executes — work is decomposed into small segments separated by reversible gates, each run by a fresh agent instance.
- Operationalised today in the PreFlight Checklist tooling, which is in active use.
Estimated savings
Savings come from the mechanisms, not a promise: cheapest-capable routing sends each job to the least expensive model that can actually do it, prompts are tightened before they run, and gates stop wasted work before tools execute. Estimated savings — varies by workload, model mix, and volume
Status: Working methods in daily use across the studio's products — this page is their showcase.