A retry that is really a poll
gh run watch 8841 exits 1 with “run still in progress” for the fourth time in six minutes. No files changed between attempts.
A retry counter would have stopped this at three. The job finished on the sixth check.
Exploring new ways to build, understand, and interact with artificial intelligence.
Apperson Labs is a small, independent research lab. We publish models, open-source tools, benchmarks and papers on how people and intelligent systems work together.
Arbiter
A family of small decision models. Given evidence and a short list of options, Arbiter says which option holds and how sure it is.
Read the model releaseLatest from the lab
Arbiter
A family of small decision models. Given evidence and a short list of options, Arbiter says which option holds and how sure it is.
Read moreNoetic
An open-source TypeScript agent framework that breaks agent patterns into eight composable step primitives.
Read moreNoetic Rules
One hook system for coding agents. It blocks on facts, steers on judgment, and never argues with itself.
Read more
At Apperson Labs, we build to understand how people and intelligent systems work together.
How Arbiter compares with models many times its size
- Arbiter M8B, choice readout93.1%91.9 to 94.2
- Arbiter S2B, choice readout89.4%88.0 to 90.8
- Frontier baselinelarge hosted model, prose answer94.0%92.9 to 95.0
- Open 8B baselineuntuned, prose answer81.2%79.4 to 83.0
- Open 2B baselineuntuned, prose answer71.6%69.5 to 73.7
Illustrative data. These figures show the shape of the evaluation and will be replaced by measured results at release.
Apperson Labs / Noetic
Noetic
An open-source TypeScript agent framework that breaks agent patterns into eight composable step primitives.
Everything in Noetic is a typed, serializable step. There are no base classes to inherit from and no hidden control flow. A loop can contain a branch, which can contain forked, spawned agents.
loopuntil done
step.llmplan
branchneeds research?
forkall, in parallel
spawnisolated context
step.tool
step.llm
spawnisolated context
step.claudeCode
step.runmerge
Research becomes experiments, experiments become releases
Nothing here starts as a product. Follow any piece of work back to the question it came from.
- Judged loops leads to Noetic Rules judge
- Choice readouts leads to Noetic Rules judge
- Noetic Rules judge leads to Arbiter S
- Noetic Rules judge leads to Arbiter M
- Stardrive crew leads to Arbiter M
- Arbiter S leads to Decisions
- Arbiter M leads to Decisions
- Decisions leads to Noetic Rules beta
Stardrive
A game whose crew are long-running agents with memory, temperament and their own opinions about your orders.
Why a lab makes a game: nothing else asks as much of an agent for as long.
- Agents that run for the length of a save file
- Memory that changes how a character behaves
- Believable responses inside a frame budget
- Inference cost a player never has to think about
Working on a similar problem? Work with the lab.