Company

How we build, and how we got here

PaxLabs is an applied research lab for reliable agents and the distributed systems they run on.

Core principles

The rules we hold ourselves to

Honest partials, never fabricated success

When a system runs out of budget or evidence, it says so. Reporting incomplete work honestly is a precondition for trusting a machine with anything consequential.

Memory must be provable

A claim about the past should be checkable, not taken on faith. Append-only logs with tamper evidence turn recollection into an auditable artifact.

Consequence gets a receipt

Conversation can be fluid. Actions that move money, sign, or change irreversible state pass through a compiler and come back signed, attested, and replayable.

Publish the failures too

We measure the alignment tax, benchmark games, and reproducibility gaps — including in our own systems. A lab that only publishes wins is doing marketing.

Warm material, machine rigor

Paper, ink, and light wrapped around kernels, compilers, and ledgers. The machine should feel like a well-made instrument, not a slot machine.

Our story

From forgetful agents to provable ones

The problem

Agents were fluent and forgetful. Every serious deployment we saw failed the same way: no durable memory, no evidence trail, and no safe boundary between talking and acting.

The turn

We stopped treating memory as a prompt trick and rebuilt it as infrastructure — one typed event log per actor as ground truth, with deterministic projections rebuilt from that log.

The stack

Neo for the agent surface, Neocortex for provable memory, MCL for consequential execution, and Paxeer for metering and settlement between machines.

Now

PaxLabs operates as an applied research lab: we publish the science, ship the runtime, and hold both to the same standard of evidence.