Learn the quality gate · fair-clock input

Official disjoint val

val_disjoint · bar ≤ 3.3821 · first-passage · not train CE / EMA

Speed claims need a quality gate first. The official gate is FineWeb val_disjoint — held out from train shards — crossing val_disjoint ≤ 3.3821.

One-line definition

First-passage = first logged row with val_disjoint ≤ 3.3821 · then read train_time_ms

01 Train shards 02 val_disjoint 03 Bar · first-cross ≠ train CE · ≠ EMA

FAIR_CLOCK CHALLENGE_TERMS GLOSSARY data/README

Visual · train vs val_disjoint vs bar schematic · Play / Step

Beat 01

Train

FineWeb train shards. Optimizes the model. Not the gate.

Beat 02

val_disjoint

Held-out FineWeb val. Official CE column in the log.

Beat 03

Bar · first-cross

≤ 3.3821. First row that crosses. Then read the clock.

Beat 01 · Train — FineWeb train shards optimize the model. Not the gate.

Train CE and EMA curves can look prettier. They do not define the cross. Sealed ΔT% lives on the report vs F1_seal. Play or Step through train → val → bar.

Problem 01 · train diagnostic only

Train shards — not the gate

Training uses FineWeb train shards. The column train_loss tracks that path. Lower train CE means the optimizer is moving — it does not prove challenge quality.

Not the metric. A train CE curve does not mint first-passage T or ΔT%.

Problem 02 · val_disjoint held out

val_disjoint — held-out FineWeb

val_disjoint is official validation cross-entropy on the FineWeb val set that is disjoint from train shards. Claims use this column — not sliding windows alone.

Train path FineWeb train shards

Seen by the optimizer every step. Logged as train_loss.

train_loss diagnostic

Official gate val_disjoint

Held out. Logged on a fixed schedule (claim-stack: every 128 steps).

val_disjoint quality

Related log columns: val_sliding, ema_val_* — diagnostics only. See data/README.

Problem 03 · bar · first-passage ≤ 3.3821

Quality bar · first-passage

The challenge quality bar is val_disjoint ≤ 3.3821. A hit is the first logged row that meets it. A run that never reaches it is a quality miss — no speed claim.

First-passage (fair clock)

1. Find first row: val_disjoint ≤ 3.3821
2. T_h = train_time_ms / 3.6e6
3. ΔT% = 100 × (3.866 − T_h) / 3.866 ← vs sealed F1_seal

Interactive · gate → clock → cross sealed curves · data/chart_series.json

Loading first-passage chart from data/chart_series.json…

1 · Gate

CE-style bar: first logged row with val_disjoint ≤ 3.3821. Not train_loss · not EMA.

2 · Clock

Read train_time_ms at that row → T_h. F1_seal freezes the public ΔT% denom at 3.866 h.

3 · Cross

First marker before baseline → ΔT% vs F1_seal. Stock runs that never cross = quality miss.

Loading…

Auto / Step / Reset (TensorTonic chart-transport pattern). Readout strip mirrors cross-entropy / perplexity widgets: one metric column updates per beat. First-passage is high-variance. Single seed ≠ confidence interval. Compile warmup stays inside train_time_ms.

Problem 04 · non-metrics do not publish

Not train CE · not EMA-as-speed

Several logged signals look like quality. Only first logged val_disjoint ≤ 3.3821 defines the cross for published speed.

Signal	Role	Defines cross?
val_disjoint	Official held-out FineWeb val	Yes — first row ≤ 3.3821
train_loss	Train CE on train shards	No — diagnostic
val_sliding	Sliding-window val	No — never the endpoint
ema_val_*	EMA of val curves / weights	No — EMA-as-speed is a NO-GO
Terminal CE alone	Final logged val	No — quality without first-passage T
step_avg_ms	Mean step time	No — forensic only

Late EMA-as-train-CE was banked as a closed NO-GO. Off for claim-stack and F1_seal.

Recompute from CSV

Public CSVs share one header. Validation rows carry val_disjoint; training-only rows leave it blank. Worked arithmetic: data/README.

first-cross = first row where val_disjoint ≤ 3.3821
T_h = train_time_ms / 3.6e6
ΔT% = 100 × (3.866 − T_h) / 3.866

Example sealed denom: f1_seal.csv first-cross @ step 4736, train_time_ms = 13,918,667.5 → T = 3.866 h on Clock-W.

Hedges / non-claims

Not claimed. Not an official NoCap win. Why each clause.

Published ΔT% is vs sealed F1_seal (3.866 h) only. Matched candidate-vs-stock first-cross ΔT% = N/A.

This page teaches the gate. It invents no new sealed percentage.

Glossary

val_disjoint — official held-out FineWeb val CE
Quality bar / hit — first logged val_disjoint ≤ 3.3821
First-cross / first-passage — that hit event; T from train_time_ms
ΔT% (vs F1_seal) — speed vs sealed AdamW denom 3.866 h
EMA — off for claim-stack / F1_seal; not the speed endpoint

FAIR_CLOCK — first-passage train time rules
CHALLENGE_TERMS — official BottleCapAI map
HOW_TO_READ_EVIDENCE — CSV recompute checklist
Evidence — sealed scoreboard / fig1 teach-beat
Report · Verify — downloads + formula
NUMBERS_CANON — sealed T / ΔT% table
Kill → bank → 3/3 — process after the gate is clear
Intro · Read more