ai os 1.stub

---

id: ai-os-1 name: AI-OS-1 repo_path: ~/Documents/sw30labs/repos/AI-OS-1 remote_url: https://github.com/sw30labs/AI-OS-1.git primary_language: Markdown framework: n/a (research corpus — no runtime, no src/) category: local-inference-dgx stacks: [dgx, agentic] ingested: 2026-08-18 last_commit_date: 2026-08-14

AI-OS-1

Purpose

A model-aware distributed inference OS for bandwidth-bound MoE clusters, published as a research corpus rather than a runtime. The first lab target is DeepSeek V4 Flash across N ≥ 2 NVIDIA DGX Spark boxes with DwarfStar (antirez/ds4) as the native engine; a two-Spark pair is stated as the minimum viable lab, not the design point. PAPER.md is the source of truth — there is deliberately no src/, and implementation is gated on Phase 0 measurements from the paper's lab plan. Drafting name was Keel; the public project name is AI-OS-1.

Thesis

Decode on CX7-connected Sparks is dominated by sequential activation hops when a model that already fits is pipeline-split, and by cold-expert traffic when unified memory is oversubscribed. AI-OS-1 sees phase, layer kind, and expert ids, then admits, places, and prefetches so decode avoids DwarfStar's pipeline-parallelism tax. Memory bandwidth (273 GB/s per box) is treated as the remaining bound; inter-box ConnectX-7 (200 Gbps) is first-order; on-die NVLink-C2C is explicitly not an inter-box link.

Architecture

No runtime — the corpus is structured as frozen interfaces plus an append-only knowledge graph:

• PAPER.md — canonical design; workitems.json enforces exclusive section ownership.
• schemas/ — frozen JSON Schema interfaces: forum-post, budget-vector, affinity-graph, session-slice-assembly, plus a confidence rubric.
• forum/ — append-only Decision / Claim / Loop posts (forum/2026/*.json) with a forum/index.jsonl topic index. Challenges fork rather than delete.
• ontology/ — Tier-0 core types (Agent, Loop, LoopInstance, Topic, Hypothesis, Experiment, Observation, Evidence, Claim, Decision, Challenge, OpenQuestion, Artifact, ConfidenceAssessment, OntologyTerm) over 13 closed relation verbs, plus Tier-1 extensions.
• research/ — independent critic rubric scores (critic/iter1.json, iter2.json), pruned inner-loop keepers, a writing-process ledger, and a Porter five-forces note.
• docs/diagrams/ — editorial HTML+SVG figures with PNG snapshots; original mermaid archived alongside.

Process objects

Session (isolation: one conversation, one KV quota, one quality class) → Slice (placement: layers and experts on one node) → Assembly (execution: a typed set of slices plus a transport that can finish one phase). The research plane (Ralph outer loop, Karpathy-style auto-research) shares the forum schema with the inference plane but not a mutable world state, and cannot write a live Slice.

Keepers already pruned

• Placement family: phase-split hybrid — Flash q2 decode stays on one box; PP is prefill and capacity; gather KV if cheaper than remaining hop tax; q4 needs a TP pair.
• Affinity update: last-window coactivation counts (2048 tokens).
• Admit: phase-split budget vector, p50 reservation, p95 hard cap.
• Forum grain: publish Decisions and Experiments; roll Observations up.

Confidence discipline

Closed C1–C5 rubric (paper §10.1). New mechanisms start at C2 and cannot reach C4/C5 without a lab, which keeps design claims separable from measured ones.

CI/CD

None — the corpus is prose, schemas, and JSON. CITATION.cff supports academic citation. Apache-2.0; the vendored diagram-design skill is MIT (Cathryn Lavery).

Entities and topics

• bonsai-ternary-27b-dgx — sibling DGX Spark workload characterization
• driftlab-dgx — DGX-side companion of a dual-platform pair
• ralph-dgx — the Ralph outer loop the research plane assumes
• sst-autoresearch — Karpathy-style auto-research loop of the same family
• wiki-vs-rag — related question of structured corpus vs retrieval
• Tags: [cuda, vllm, local-inference, llm]