Ops — win-rate gate · file knobs · cluster peek

Ruled on-disk stack (knobs/base, ledger) and box gate inputs. Live fleet knobs.json is on Train. Skill signal is win-rate vs 29b; macro top-1 is not skill.

— gate poll on load · ≥200 games vs 29b before submit or scale

glance

—

planned epochs

steps ×1024 / window

—

league games

fixed league file

—

macro top-1

IL metric · not gate

—

set_hash

ledger tail

epochs bar · green ≤9ep · red >9ep

Gate detail · raw box poll

rule

Anchor: ckpt-2026-07-29b-00820000 (312M, Elo 1029.2). Promote when Δ Elo − 2·SE > 0 and win-rate ≥ 52.5% at N ≥ 200. Full procedure: notes/raw/win-rate-gate.md (AGENTS § Submit paths).

≥200 games Δ − 2SE > 0 win-rate ≥ 52.5%

status—

best macro tail—

epochs at best—

league line—

raw

—

9 layers · knobs/base + ledger tail

Declared stack from the last ledger row and current knobs/base. Treat as stamp state; re-check against disk and the box when promoting.

layer	state	hash / note
loading…

File knobs · G1–G5 editor

Writes through scripts/knobs (file + ledger). Load a path, edit one G-group at a time, preview the banner, then save under experiments/… or overwrite with care.

load path

save as

banner

file preview

Cluster peek · peer trees on gpucluster

One SSH poll of /shared/pokemon-llm and /shared/pokemon-model-free. Read-only cotenancy check; do not kill peer GPUs 0–3 when util is high.

not polled ssh · 12s timeout

pokemon-llm

summary—

v10 lines ? v8 lines ? draft 600: ?

elo log tail

—

pokemon-model-free

loss—

pg / value—

window / age—

—

raw sections

—

Layer swap · one G at a time

Helpers that write into the in-memory knobs form (see File knobs). Change one G-group per sweep; record the rest as fixed. See notes/knobs.md.

swap fields

Encoding (G1) Teacher (G2) Model (structural) Data (G4)

presets

	library → macro_lib_v003
	deck_pool.csv · no retrain
	side lineage · V-Trace

knobs/base banner