The modelsunder the agents

Four tracks

One model per hard problem

We kept them separate on purpose. Generating a brain, keeping it current, routing work, and gating risk are different problems with different failure modes, and each gets better faster when it can be measured on its own.

1. 01

   Genesis

   Brain generation

   In production. Improving coverage on supplier and compliance facets.

   The question

   Given one or two lines a founder writes about their business, how accurately can a model reconstruct the structure of that business — products, market, channels, inputs, intent?

   Why it matters

   Every agent reads the brain. If the brain is wrong on day one, every agent is wrong on day one. Genesis decides how much the owner has to correct before the platform is useful.

   What we measure

   • Facet accuracy — fraction of generated facets the owner accepts without edit
   • Coverage — fraction of the facets an agent will need that were generated at all
   • Correction cost — owner edits required before first task, and where they cluster
   • Prompt sensitivity — how much output quality moves with prompt length and detail
2. 02

   Calibra

   Continuous calibration

   In production. Active work on thinly-evidenced changes — when to wait for more before flagging.

   The question

   As agents complete work, how should the brain update — which completed tasks are evidence that a fact has changed, and how much should one piece of evidence move a belief?

   Why it matters

   A brain that doesn't move becomes a confident description of a company that no longer exists. Calibra is what keeps the platform accurate in month six without asking the owner to maintain anything.

   What we measure

   • Drift detection — how quickly a changed fact (a supplier rate, a payment term) is reflected after it shows up in completed tasks
   • False update rate — facets changed that the owner then reverts
   • Flag precision — of the changes surfaced for review, how many the owner confirms
   • Confidence calibration — whether a facet's stated confidence matches how often it turns out right
3. 03

   Relay

   Task routing and handoff

   In production for the three live agents. Extending handoff contracts as agents ship.

   The question

   Given an instruction in plain language, which agents own it, in what order, and what does each have to hand to the next — without a person brokering the sequence?

   Why it matters

   Owners describe the job, not the agent. Relay is the difference between a chat box that works and one that asks you to pick from a menu.

   What we measure

   • Routing accuracy — first agent chosen matches what the task needed
   • Handoff completeness — in multi-agent tasks, whether each agent received what it needed from the previous one
   • Clarification rate — how often Relay has to ask before starting, and whether asking was warranted
   • Time to first action — from Assign to the first agent reading context
4. 04

   Sentry

   Approval and risk

   In production. Working on card design for multi-step actions.

   The question

   For every action an agent is about to take, does it fall inside the owner's rules — and when it doesn't, how do we present the decision so a person can make it in thirty seconds?

   Why it matters

   Nothing consequential should execute without a person's yes. Sentry is the part of the platform that makes that true, and makes saying yes cheap.

   What we measure

   • Policy adherence — actions that executed outside a rule: the target is zero, and it's measured, not assumed
   • Decision time — seconds from card shown to approve or hold
   • Card sufficiency — approvals made without the owner opening anything else
   • Stop precision — how often a gated action was one the owner was glad was gated