What is Pimsy?

Pimsy is designed around a single premise: a system that can reason should also be able to finish the work. That requires a runtime, not just a model.

Model versus platform#

The Pimsy language model handles understanding, decomposition, code, mathematics and analysis. It does not, by itself, hold state across days, call a private API, run a test suite, or refuse to submit an unsimulated transaction. Those properties come from the platform around it.

Design principles#

1. 1

   Outcomes over responses

   A task is complete when its completion criteria are satisfied and verified — not when tokens stop being produced.

2. 2

   Observable by default

   Every run emits an inspectable trace: plan, steps, tool invocations, inputs, outputs, retries and verification results.

3. 3

   Bounded autonomy

   Autonomy is granted through explicit scopes, budgets and policies. Sensitive effects require approval by configuration, not by chance.

4. 4

   Failure is a first-class state

   Steps can fail, retry, escalate, or terminate with a reported reason. A partial result with an honest boundary beats a confident fabrication.

5. 5

   Uncertainty is reported

   Where evidence conflicts or is thin, Pimsy is designed to say so and attach the conflicting sources rather than resolve them silently.

What Pimsy is not#

• It is not a claim of artificial general intelligence. Pimsy is an ambitious agent architecture with measurable, bounded capabilities.
• It is not a chat wrapper with a tool list bolted on. Planning, state, verification and policy are core runtime services.
• It is not a trading system. The Markets subsystem performs research and analysis; it does not issue financial advice.
• It is not an autonomous committer. Repository capability stops at a working branch; merge and release authority stay with humans by design.