Thirty Spokes — Subnet console

subnet 99 · v3window 17tasks/arm 250benchmarks hle · livecodebenchpower gate separatesfeed store.thirtyspokes.ai

the throne

5CkC…9hwq

a miner's conductor

0.506finalquality 0.557

benchmark	quality	$/task	final
hle	0.250	$0.0016	0.196
livecodebench	0.864	$0.0015	0.816

emission share 85.0%

king 0.85 · the five most recent ex-kings 0.05/0.04/0.03/0.02/0.01 · an unfilled slot burns

king final · every window that measured it

window 2 → 17 · 3 of 13 published windows measured the king · filled dot = a crown changed handsfinal = quality − λ·spend/C · the king's arm is re-run on every window's fresh slice

reigns · the throne and what it pays

0.85 to the king · a five-deep decaying pension · every unfilled slot burns

king	uid	reign	windows	role	share now
5CkC…9hwq	18	window 2–17 · still reigning	12	reigning	85.0%
thirtyspokes-genesisKing₀ · the cascade policy	0	window 1	1	dethroned	—

15.0% of this window burns at uid 0 · share is this window's slate as set on chain — an ex-king that no longer resolves to a uid is paid nothing, and its slot burns rather than moving to the king

queue · committed, not yet judged

3 waiting · 3 duelled per window · cap 6

#	uid	hotkey	committed	manifest	state
1	—	5DrD…aMRC	#9083824	eed6d904d8…	queued
2	—	5Ew7…XrFi	#9087101	cd39dbfbea…	queued
3	—	5F9X…ABek	#9099280	0ec93eb7ac…	queued

published by the validator at block 9133797 · window 50 — republished whenever the queue changes

duel history · every challenge judged

4 duelled · 1 crowned

window	uid	hotkey	model	verdict	Δ	lcb	king	challenger	when
17	21	5G6r…M4e8	hidden until promotion	king	+0.036	+0.010	0.506	0.542	this windowcommit #9077021
10	20	5Ex4…aZ9y	hidden until promotion	king	+0.022	+0.008	0.428	0.450	~39h agocommit #9065965
2	19	5Dcw…EByH	hidden until promotion	kingcleared ε	+0.084	+0.051	0.359	0.442	~3d agocommit #9053397
2	18	5CkC…9hwq	unnamed	challenger	+0.087	+0.045	0.359	0.430	~3d agocommit #9053265

Δ > ε (0.05) · bootstrap lower bound > 0 · leave-one-out > 0 · median benchmark > 0 — click a row for the arithmetic · age is derived from the window number, so the row carries the window and the commit block too

audit · the rest of what this window published

power gate

can this slice tell two known-different policies apart?

+0.338separatescascade over always-cheapest

measure	value	needs
spread	+0.338	> 0.05
bootstrap lcb	+0.243	> 0
tasks per arm	60	—
hleper benchmark	+0.200	—
livecodebenchper benchmark	+0.475	—

two fixed policies, once per window · if the slice cannot separate them, no duel is scored and no challenger spends its one shot

weights · as set on chain

the only part of this page the chain also holds

uid	hotkey	share	role
18	5CkC…9hwq	85.0%	king
0	—	15.0%	burn

an unfilled or unresolved slot burns to uid 0, never to the king

the slice · freshness and the seen/unseen split

tasks are reusable; freshness comes from unpredictability, not scarcity

tasks per arm: 250
categories: 2hle · livecodebench
recycled: 100%hle · livecodebench
graders that failed: none

arm	unseen final	unseen tasks	seen final	seen tasks
5G6r…M4e8	0.516	134	0.505	116
5CkC…9hwq	0.342	134	0.406	116
always-cheapest	0.293	19	0.087	41
cascade	0.358	19	0.464	41

the slice is drawn by a nonce committed before the window opened · a miner that memorised earlier windows shows up as a gap between its seen and unseen columns

the window's outcome table

one live buy per (task, model, attempt), replayed for every arm

rows: 760
bought live: 760the provider was paid
replayed: 335answered from the table
delegates asked: —across all arms
provider paid: $0.46
arms charged: $0.000.00× the bill
rows the provider failed: 16121.2% of rows

hatched = the share of all delegates answered from the table rather than bought again

a hit is charged to the arm at the row's own price and calls no provider and no grader · identical policies tie exactly, so what separates two arms is their routing and not their luck

table re-test

a sample of rows bought a second time and compared

this window ran no re-test

the workers are sampled, not deterministic, so a disagreement is the pool's own variance rather than a fault — it is published because a table nobody checks is a table nobody can trust

metering · every arm reconciled against its receipts

what the arm was charged, what the provider was paid, and which endpoint served each fill

arm	payer	delegates	replayed	bought	charged	provider paid	recorded	audit

charged = every delegate at the table's price · provider = the fills alone · a miner pays for its own arm on its own key (D18); the owner pays the reference arms and the king's · resumed = tasks restored from a checkpoint at zero new calls

Thirty Spokes · subnet console