Aither World awgym

awgym

An ARC training gym — a game a world model can watch.

What It Does

Instead of trusting

a benchmark a model has never lived in

You check

an environment, six solver roles, and a world model that learns from the play

Overview

awgym/README
A world model is trained on transcripts of tasks nobody runs. A gym makes the task the teacher: ARC-AGI-3 environments (ARCAGI3Env + RHAE scoring, vendored from the NVIDIA dream-team tree), the six DreamTeam solver roles, and a LeWM training loop that plays, observes every transition, burst-trains and scores on a disjoint eval slice.

awgym

An ARC training gym — a game a world model can watch.

pip install awgym · The Aither World

A world model is trained on transcripts of tasks nobody runs. awgym makes the task the teacher: one environment, six solver roles playing through it, and a world model that learns from the play.

Phase 1 — ARC-AGI-3 environments + a LeWM training loop. The environments are ARCAGI3Env over the official arc-agi SDK with RHAE scoring, both re-exported from the vendored NVIDIA dream-team tree via awgym.vendor.dream_team. The loop plays, observes every transition, burst-trains, and scores on a disjoint eval slice.

Phase 2 — the six DreamTeam solver roles as agent roles. The six roles are reimplemented as awdk role agents with LeWM as the SIMULATOR — neural predictions replace the agent-written executable world model.

```bash
awgym --help        # the CLI
awgym serve         # the gym server (FastAPI)

The vendored tree

The dream-team source (Apache-2.0, NOTICE retained) lives OUTSIDE this package at ARC_GYM_DREAMTEAM_ROOT (default E:\AitherOS-Data\arc-agi-3\dream-team); this package never carries it.

Development

```bash
pip install -e ".[dev]"
python -m pytest tests/ -q

Licence

Apache-2.0.

The Aitherium Ecosystem

Portable tools you adopt one at a time. Each one works alone.

AitherConnectAitherConnectBrowser extension — federated AI search, page context, and the Living OS overlay. AitherZeroAitherZeroPowerShell 7+ automation framework — numbered, self-describing scripts. aitherkvcacheaitherkvcacheNear-optimal KV cache quantization for LLM inference — sub-byte compression. awaskawaskYour agent asks you a question — and acts on your answer. awbacawbacRole-based access control that fails closed and explains itself. awbrowseawbrowseA portable browser client — navigate, console, network, DOM, screenshot. awdelphiawdelphiAnonymous multi-round expert panels — a converged answer with a trace. awditawditAn append-only audit trail whose gaps are DETECTABLE. awdkawdkBuild AI agent fleets — 3 lines, any backend, local or cloud. awfindawfindA portable search client — query, results, ranking. awfocusawfocusSee, search and steer every Claude session from one command. awgitawgitSemantic version control on top of git — edit-ops and leases. awgraphawgraphA semantic code graph for agents — AST + tree-sitter, call graphs. awgymawgymAn ARC training gym — a game a world model can watch. awiamawiamWho is this caller? A directory and session store that fails honestly. awknoawknoThe man page for the Aither World — every brick, stack and law, offline. awknowledgeawknowledgeHow to run a coding agent so the result survives — the laws, with evidence. awmawmA portable, scoped agent memory. awmailawmailGive an agent an email address — send, and actually receive. awnboardawnboardA front gate you can put in front of anything, and hand someone the key to. awnestawnestProve there is a human before you let them into the nest. awnetawnetThe agentic web — agents host a mesh, and agents join one. awnixawnixA Linux you can hand to an agent — immutable base, capabilities included. awnodeawnodeA lightweight local gateway — bridges your apps to the AI backends you chose. awpackawpackFirst-party agent packs — the ones we build, versioned and installable on their own. awpredictawpredictPredict what your environment does next, and how surprised you were. awprismawprismTurn a failure into ranked hypotheses — and say what would confirm each one. awreasonawreasonA portable reasoning client — sessions, phases, thoughts, and the chain that produced the answer. awrecoverawrecoverLabelled snapshots with an all-or-nothing restore. awrecurseawrecurseAnswer a question over a context far larger than the window — recursively, with the trace kept. awrelayawrelayPortable agent messaging — findings, alerts, coordination. awrenaawrenaPut two agents head to head and get a verdict you can check. awreplawreplA REPL an agent can actually use — state that survives between turns. awresearchawresearchAsk a research question, get a cited report you can check. awriseawriseWake an agent on a schedule, let it do one thing, and put it back to sleep. awrtifactawrtifactDeliberately chunk artifacts into GitHub release assets — the productized aitherkvcache mirror lane. awrunawrunA priority-aware queue and dispatcher for agentic runs and ad-hoc CI builds. It also judges whether the runner pool is big enough for the queue it is draining, and can ask a host to grow it -- reserving capacity is zero-sum, so a saturated pool needs more of it, not a different share of it. awsealawsealSign an artifact so a stranger can verify it. awshawshYour terminal answers you -- type a question where a command would go. awshareawsharePublish an artifact and fetch it back verified. awskillsawskillsPortable agent skills — self-contained procedures an agent loads on demand. awtollawtollWhat every tool call costs you in context, measured from your own transcripts. awtunnelawtunnelReach a service that has no public address. awwallawwallSay what a workload may reach, and watch everything else fail closed. gawbbonetgawbbonetGobboNet campaigns with a real agent brain — scoped memory, graph recall. mediaforgemediaforgeThe creative studio — search the boards, render scenes, keep the character.