Glossary — NoCap campaign terms

Evidence state: 2026-07-16 · claim-stack v1 · 3/3 candidate seals.
Numbers: prefer sealed summaries; do not invent.

Canonical sealed table: nocap-repo/results/claim_stack_v1/MULTI_SEED_CLAIM.md
Public denominator: nocap-repo/results/f1_seal/SUMMARY.md

Challenge & quality

Term	Meaning
NoCap / NoCap-Test	BottleCapAI challenge: train GPT-2 ≈124M on FineWeb to official disjoint val ≤ 3.3821 on one GPU, faster than a stock baseline. Challenge repo: BottleCapAI/NoCap-Test.
Quality bar / hit	First logged val_disjoint ≤ 3.3821. A run that never reaches this is a quality miss (no speed claim). Pedagogy: disjoint-val-explained.html.
val_disjoint	Official validation loss on the disjoint FineWeb val set (held out from train shards). Claims use this — not sliding/EMA curves alone. Primer: disjoint-val-explained.html.
CE / terminal CE	Cross-entropy (val loss). Terminal CE = val at the final logged step (usually 4768). Lower is better.
Horizon / 4768	Full claim horizon: 4768 optimizer updates. Tokens per update stay 524,288.

Time metric (fair clock)

Term	Meaning
train_time / train_time_ms	Cumulative training wall clock logged by the trainer (includes compile warmup). Primary clock for speed.
T / first-cross T	`train_time` (usually in hours) at the first step where val_disjoint ≤ 3.3821.
First-cross / first-passage	The event of first hitting the quality bar; synonymous with first-cross T when discussing speed. High variance — single seed ≠ confidence interval. Pedagogy: disjoint-val-explained.html.
ΔT% (vs F1_seal)	100 × (T_F1_seal - T_M) / T_F1_seal. Positive = faster than sealed F1. Public denom T_F1_seal = 3.866 h.
step_avg	Mean step time (ms). Useful forensic, not the published speed metric.
N^*	Step index of first-cross (sample efficiency). Sealed candidates: all 4736.

See FAIR_CLOCK.html.

Sealed denominators

Term	Meaning
F1_seal	Sealed stock AdamW baseline on Clock-W. First-cross @4736 · T=3.866 h. Sole public speed denominator for published ΔT%. Pedagogy: f1-seal-explained.html.
F1_Claim-W2 (local file: `F1_BCAROL_SUMMARY.md`)	Same-class AdamW re-seal on Claim-W2. T=3.858 h (~+0.2% vs F1_seal). Used for same-workspace ratios on seed 2029; not a replacement for public F1_seal.
Historical F1	Notebook-era stock cross (~229.1 min). Torch unlogged → provisional only. Never use for published %.

Claim-stack recipe language

Term	Meaning
Claim-stack / claim-stack v1	Full-horizon experiment family (Jul 15–16 seals): Muon B32 systems path ± timing-cheap split-QKV, three seeds.
Candidate	The claim arm: timing-cheap split-QKV (`muon_split_qkv=1`, `match_joint=0`) on the shared Muon B32 recipe.
Stock control (claim-stack)	Matched arm: joint QKV on the same Muon stack (not AdamW F1). Differs only on QKV treatment.
Matched pair	Same seed, same workspace, same pin, same shared recipe; only split vs joint differs. Seeds 1337 and 2029 are matched pairs; 4242 is candidate-only.
Joint QKV	Standard fused QKV projection (stock control). Primer: qkv-explained.html.
QKV (Query / Key / Value)	The three attention projections: Query asks, Key matches, Value mixes. Stranger primer: qkv-explained.html. Packing (joint vs split) is separate — see timing-cheap below.
Timing-cheap split-QKV	Campaign mechanism: split QKV path designed for low step-time tax vs dual-polar split. Banked @800/@2400; used in sealed candidates. Primer: split-qkv-explained.html. Letters: qkv-explained.html. Claim boundary: split-vs-ortho-explained.html.
Dual-polar split-QKV	Earlier split variant: ~+0.015 CE but time ratio ≈ 1.026 → CE edge eaten by step tax → NO-GO for claim. See split-qkv-explained.html.
Split ≠ Ortho	Ortho/NS×5 is Keller’s Muon kernel; timing-cheap split-QKV is separate Q/K/V treatment that still calls the same NS. Sealed claim needs both. Not inventing Muon via split. Pedagogy: split-vs-ortho-explained.html.

Systems & optimizers

Term	Meaning
B32×accum16	Microbatch 32 × gradient accumulation 16 (with seq 1024 → 524,288 tokens/update). Systems path ~6% faster core step-time vs B16 path (matched systems_v1). Pedagogy: b32-systems-explained.html.
Muon	Keller Jordan’s Muon optimizer (Muon). Claim recipe uses Muon on matrix groups + Adam on embed/head. Attribution required. Primer: muon-explained.html.
NS / Newton–Schulz	Orthogonalization steps inside Muon (Ortho ≈ `zeropower_via_newtonschulz5`). Claim pin: NS 5. Not the same as split-QKV — see split-vs-ortho-explained.html.
Warmup / warmdown (wu / wd)	WSD schedule lengths. Claim pin: wu 476 / wd 1192. (Proxy Tier A used shorter warmdowns; do not confuse.)
zloss	Z-loss term. Claim pin: 0.
Prior / frozen unigram prior	Frozen `prior_full.pt` logit prior on claim-stack. F1_seal uses no prior.
EMA	Exponential moving average of weights. Off for claim-stack and F1_seal. Late EMA-as-train-CE was a NO-GO.
Parallel block	Parallel transformer block (banked ~1.6% step in notebook-era family). On for claim-stack.
Compile warmup	Torch compile startup cost. Included in T; do not strip it.

Hardware attestation

Term	Meaning
A100-SXM4-40GB attestation	Exact GPU class required for sealed claims. Seed 4242 first spawn hit 80GB and was rejected by guard; seal is the relaunched 40GB run.
torch 2.13 / 2.13.0+cu130	Sealed pin. Log full `torch.__version__`. Historical notebook images used other builds → infra drift.

Workspaces (public codes)

Public aliases for ops roles (Clock-W, Claim-W*, Research-W*, Archive-W*). Pedagogy: workspace-codes-explained.html.

Code	Role
Clock-W	Host of public F1_seal.
Claim-W1	Claim-stack seeds 1337 (matched pair) and 4242 (candidate-only).
Claim-W2	Claim-stack seed 2029 + F1_Claim-W2 AdamW re-seal.
Research-W1…W6	Component bank, proxies, screens (ordinal / mechanism era).
Archive-W1…W3	Historical / archive hosts.

Campaign process (kill → bank → seal)

Process jargon — not a model knob. Pedagogy: kill-bank-seal-explained.html.

Term	Meaning
Kill	Close a dead end under the fair clock: quality miss, quality without under-baseline speed, or time-tax fail. Stop spending compute on that line. Examples: `claim_muon_par` (never crossed); `claim_muon_wsd` (~0.8% slower than F1_seal); dual-polar split; Tier A 5/5 NO-GO.
Bank	Keep a component that passed promote screens as a building block (e.g. B32×accum16, timing-cheap split-QKV). Bank ≠ full claim — component evidence only until a full-horizon seal.
Seal / 3/3	After kill + bank: run the composed recipe on three seeds vs sealed F1_seal. Candidate seals 1337 / 2029 / 4242 (mean ΔT% ≈ +8.3%). Freezes SUMMARY + CSV for claims. Not an official organizer win; matched cand-vs-stock ΔT% = N/A.

Evidence shorthand

Term	Meaning
Seal / sealed	Written SUMMARY + (when present) pulled CSV that freezes a run’s numbers for claims. Process sense: see Seal / 3/3 above.
Proxy (@800)	Short-horizon screen. May rank methods; must not produce published ΔT%.
Noise floor (~6%)	Empirical infra/step-time drift prior. Single-seed \|ΔT%\| < ~6% → do not publish as win.
Under-claim / under-baseline	Candidate T meaningfully below F1_seal (campaign bar ≥6%). Language stays bounded — not “officially won.”
Fire-and-forget spawn	GPU train via `spawn()` + `modal run --detach` so a local disconnect does not cancel A100 work.
Volume isolation	Modal volumes do not share across workspaces. Data/priors must be seeded per workspace.

Non-claims (do not redefine these as wins)

See FAIR_CLOCK.html and FAILURE_LEDGER.html. Pedagogy: not-claimed-explained.html.

Term	Meaning
Not claimed	Not an official NoCap win; ΔT% only vs sealed F1_seal; matched cand-vs-stock first-cross ΔT% = N/A.
Matched ΔT% N/A	Cannot compute honest matched first-cross speedup vs joint-QKV Muon stock: stock never crossed on 1337/2029; seed 4242 stock arm not run. Public % is vs F1_seal only; joint miss remains a quality ablation.

No official NoCap win (organizer acceptance not in local seals).
No matched first-cross ΔT% vs joint stock (stock never crossed where measured).
No published % vs historical F1.
F2b quality hit ≠ speed win (historical −2.9% retracted).