A research-os instance — durable, version-controlled research state for autonomous AI-systems research. Runs on the research-os engine. First-class objects, not session logs.
registry/
claims/ CLAIM-xxxx.yaml — first-class claims (seed/active/weakened/killed/promoted)
experiments/ EXP-xxxx.yaml — registered experiments (no compute without one)
verdicts/ VERDICT-xxxx.yaml — committee verdicts
cemetery/ DEAD-xxxx.yaml — killed ideas (resurrection-guarded)
academic_map.yaml · baselines.yaml · projects.yaml
experiments/ EXP-xxxx/ — experiment artifacts (results, analysis)
prior_art/ CLAIM-xxxx/ — prior-art packets
sessions/ raw execution logs (forensic only — NOT source of truth)
learning/ crash postmortems + operational lessons
projects/ project lifecycle
Config (research-os.config.yaml) is git-ignored here (Meta-internal: nodes, committee backends, cert glue)
— it lives in the private research-os-meta repo.
Theme (PROJ-0001): GPU CUDA-VMM is the wrong abstraction for agentic KV branching.
- Promoted (6/6 GREEN): CLAIM-0001 (HW CoW dominated), 0002 (NVIDIA mapping-ceiling cliff), 0003 (compounding throughput collapse), 0004 (mapping-budget wall), 0006 (prefix-cache invalidation law + CDC repair), 0007 (attention-visible write-after-share, bit-identical).
- Active/YELLOW: CLAIM-0005 (injection penalty — sub-quadratic, capped), 0008 (error-class→recovery).
- Cemetery (8 killed): HW isolation primitive, handle-attestation, VMM decision-procedure, super-linear artifact, segmented-hash, KV-divergence model, layer-stratified reusability, interior KV-repair.
Sessions under sessions/5_30_2026/ are the raw forensic logs that produced this state. The registry —
not those logs — is the source of truth. New work goes through the engine (ros seed/exp/...).