# Campaign overview — causal story

**Audience:** strangers who want the plot, not the ops manual.  
**Evidence state:** 2026-07-16 · **3/3** claim-stack candidates under F1_seal.  
**Live report:** https://nocap.think2earn.com  
**Detail timeline:** [CAMPAIGN_TIMELINE.md](CAMPAIGN_TIMELINE.md) · **Failures:** [FAILURE_LEDGER.md](FAILURE_LEDGER.md)

Workspace codes only. Numbers from sealed summaries — under-claim; no trophy language.

---

## One-line spine

Failed finals → fair clock + **F1_seal** → Muon compose misses → bank systems + cheap split → Tier A keeps defaults → claim-stack **3/3** under sealed F1 (candidate vs F1, **not** matched stock speedup).

---

## What the challenge is

Train GPT-2 ≈124M on FineWeb until official disjoint val ≤ **3.3821**, on **one** GPU, with less first-passage train time than a stock baseline. This campaign’s public comparator is our sealed stock AdamW run **F1_seal** (\(T\)=**3.866 h** on **Clock-W**), pending organizer rules.

---

## Act I — Early finals (quality theater)

| Run | Outcome | Takeaway |
|-----|---------|----------|
| **F2** (short warmdown) | Terminal ≈ **3.405** · never crossed | Quality miss. |
| **F2b** (long warmdown) | Crossed @4736 · terminal ≈ **3.377** | Quality hit. Historical wall ≈ **−2.9%** vs notebook F1 **retracted as speed** (infra/noise). |

Notebook-era **historical F1** (~229.1 min) exists but torch was unlogged → provisional only.

---

## Act II — Fair-clock reform

Unpinned torch builds drifted step time (~**6%** observed). Speed claims now require:

- GPU **A100-SXM4-40GB**
- torch **2.13.0+cu130**
- \(T\) = train_time at first val_disjoint ≤ 3.3821
- Public denom = **F1_seal** only

**F1_seal** (Clock-W): stock AdamW · cross @4736 · val 3.381538 · \(T\)=**3.866 h**.

---

## Act III — Full claims that did not under-baseline

| Run | Quality | Speed vs F1_seal |
|-----|---------|------------------|
| **claim_muon_par** | Miss (final **3.386130**, never crossed) | N/A |
| **claim_muon_wsd** | Hit (~3.376 terminal) | \(T\)≈**3.899 h** · ~**0.8% slower** |

Board-v3 800-step screens: **no promote.** Proxies are not claims.

---

## Act IV — Component bank (research workspaces)

On **Research-W\*** accounts, levers were screened and banked:

- **B32×accum16** systems path: ~**6%** faster core step vs B16 (tokens/update fixed).
- **Timing-cheap split-QKV**: multi-seed mechanism repeat @800 · persistence @2400 (Research-W4 seed 5151).
- Many closed NO-GOs: curation cleaners, Adam LR 0.003 alone, dual-polar split, muon_lr 0.025, NS6, warmdown 300, late EMA as train CE, Tier A mom/WD/warmdown variants.

Through 2026-07-14 evening: bank existed; **0** under-baseline full recipes yet.

---

## Act V — Claim-stack seals (the present)

Recipe: Keller Jordan **Muon** + B32 path + **timing-cheap split-QKV** candidate, vs joint-QKV stock control on the same stack, full 4768.

| Seed | Workspace | Candidate \(T\) | vs F1_seal | Stock control |
|-----:|-----------|----------------:|-----------:|---------------|
| **1337** | Claim-W1 | **3.553 h** | **+8.1%** | matched · **never crossed** |
| **2029** | Claim-W2 | **3.567 h** | **+7.7%** | matched · **never crossed** |
| **4242** | Claim-W1 | **3.511 h** | **+9.2%** | **not run** (candidate-only) |

All candidates first-cross at step **4736**. Mean \(T\) ≈ **3.544 h** · mean ΔT% ≈ **+8.3%**.

Same-workspace AdamW check **F1_Claim-W2** on Claim-W2: \(T\)=**3.858 h** (~+0.2% vs F1_seal).

**Hard honesty:** do **not** claim matched speedup vs joint stock — stock never crossed where measured.

---

## What strangers should remember

1. Multi-seed **under-baseline vs sealed F1** is sealed and documented.
2. That is **campaign evidence**, not an organizer medal.
3. Failures stay visible ([FAILURE_LEDGER.md](FAILURE_LEDGER.md)).
4. Attribution: **Muon → Keller Jordan**; timing-cheap split-QKV → this campaign’s banked mechanism.

Next step in the public narrative: package the writeup / submission — not more GPUs unless an operator asks.