Workspace map — public roles

Workspaces are ops roles, not performance claims. Each workspace has its own volume — volumes do not share across workspaces. Data, priors, and run artifacts must be seeded / pulled per workspace.

Public pages name workspaces by code (Clock-W, Claim-W1, Claim-W2, …). No tokens, secrets, or app credentials belong on the public page.

Code scheme

Prefix	Meaning
Clock-W	Sealed stock denominator
Claim-Wn	Claim-stack seal hosts
Research-Wn	Component / proxy / screen hosts
Archive-Wn	Historical / archive hosts

Clock and claim (Jul 15–16 seals)

Code	Role in sealed evidence
Clock-W	Clock. Public F1_seal (stock AdamW, T=3.866 h). Also hosted early full claims (`claim_muon_par`).
Claim-W1	Claim-stack seeds 1337 (matched pair) and 4242 (candidate-only).
Claim-W2	Claim-stack seed 2029 (matched pair) + same-workspace F1_Claim-W2 AdamW re-seal (T=3.858 h).

Research / bank (component era)

Code	Role
Research-W1	Historical proxy / ordinal research (early FAIR_CLOCK dual-workspace design). Not a Jul 15–16 claim-stack host.
Research-W2	Horizon screens + `claim_muon_wsd`.
Research-W3	Cheap-split mechanism repeats; Tier A variant; late-EMA NO-GO.
Research-W4	Cheap-split @800 + 2400 persistence seal (seed 5151).
Research-W5	Early cheap-split @800 mechanism; Adam-LR persistence work.
Research-W6	Cheap-split repeats; p2400 attempt incomplete / archived — not a seal.
Research-W7	Systems-v1 replication peer (with Research-W6). Not listed on the public HTML table.

Archive / historical

Code	Note
Archive-W1	Often local CLI default; not a claim-stack seal host for this narrative.
Archive-W2	Historical waves; preemption failures. Archive.
Archive-W3	Historical finals host (e.g. F2b pull). Archive.

Exact credit balances and retired-account dollars are ops detail — omit from public narrative unless a sealed summary already states them.

Creating a volume on workspace A does not expose workspace B’s checkpoints or FineWeb shards.
Before any claim-stack launch on a new workspace: volume put data + prior, smoke, then train.
Pull CSVs with the same workspace that owns the run (see local map for CLI profile names).

Fire-and-forget ops note

Early claim-stack work died when the local Modal client awaited nested .remote() train and hit ClientClosed / deadline — Modal canceled the GPU input.

Fix (sealed runs): GPU train via spawn() fire-and-forget + modal run --detach. Orchestrator does preflight, spawns the GPU arm, returns. Local disconnect must not cancel A100 work.

Seed 4242 also required a 40GB relaunch after an 80GB GPU-guard miss.