Campaign runs overview
Short map of what NoCap volume runs exist across campaign workspaces.
Codes: Clock-W, Claim-W1/W2, Research-W1…W7, Archive-W1…W3.
Evidence date: 2026-07-17 · volumes listed live; seals from local claim/F1 summaries.
Scale
- 13 workspaces · each has a
nocapvolume · 153runs/entries (incl. partials)
Sealed claim path (headline)
| Piece | Workspace | Outcome |
|---|---|---|
| Public F1 denom (stock AdamW) | Clock-W | SEALED · T ≈ 3.866 h |
| Same-account F1 check | Claim-W2 | SEALED · T ≈ 3.858 h |
| Claim-stack cand seed 1337 | Claim-W1 | HIT · ≈ +8.1% vs F1 |
| Claim-stack cand seed 2029 | Claim-W2 | HIT · ≈ +7.7% vs F1 |
| Claim-stack cand seed 4242 | Claim-W1 | HIT · ≈ +9.2% vs F1 (cand-only) |
| Joint-QKV stock controls | Claim-W1/W2 | Quality MISS where run (1337, 2029) |
Mean cand first-cross ≈ 3.544 h · mean ΔT% ≈ +8.3% (3/3 seeds).
Other major families
| Family | Where | Role / result (short) |
|---|---|---|
| Early Muon full claim | Clock-W | Complete · quality MISS |
| Proxy ordinals (ix/px/b3) | Research-W1, W2 | Promoted Muon+parallel; WSD best @1600 proxy |
| Systems B32 path | Research-W6, W7 | Dual-workspace ~6% step-time; NS3/z-loss killed |
| Cheap split-QKV theory | Research-W3…W6 | Multi-seed mechanism-repeat @800; persistence sealed @2400 on Research-W4 |
| Tier A hparams / EMA / Adam LR@2400 | Research-W3…W5 | Mostly NO-GO (kept defaults) |
| Curation cleaners | Research-W7 | Screened · closed for claim path |
| Historical waves / f2 | Archive-W1…W3 | Archive-complete; not claim-stack seals |
Honest limits
- Do not claim matched first-cross speedup vs joint stock (stock never crossed on paired seeds).
- Early compose claim on Clock-W is a miss, not a win.
- Research-W6’s aborted 2400 joint arm is not a seal; Research-W4’s matched 2400 pair is.
- Many mid-horizon dirs exist without a published seal line — presence ≠ win.
For operator detail (still codes-only): see VOLUME_INVENTORY.local.md (local, do not deploy).