Glossary

Terms as they are used in Thirty Spokes under the v3 mechanism (v3), with pointers to where each is specified. Repository paths refer to docs/WHITEPAPER.md (§) unless stated.

The mechanism

Conductor — what a miner submits: full model weights on the pinned Qwen3.6-35B-A3B architecture. Given a rendered state it emits one action; it never writes an answer and never touches the worker's input. (§1–§3, Build a Conductor)

Worker — the catalog model a DELEGATE or RETRY sends the task to. It answers; its output is graded; its cost is metered. A worker erroring or refusing is an outcome scored against the conductor that chose it. (§8b.3)

Scaffold — the owner's pinned loop: render the state, parse one action, call the worker with the original task, record cost and outcome, grade at STOP. Identical for king and challenger, so its quality cancels in the duel. (§3)

Action — REASON <text> (internal, capped at 512 tokens), DELEGATE <model>, RETRY <model>, STOP. Anything else is a parse failure; three in a row force STOP. MAX_STEPS = 12. (§2)

Invariant — the only things that reach a worker are the pinned task text and a validated model ID. The property that makes public benchmarks safe to score on. (§2.1)

Window — one scoring period: a fixed span of blocks with a committed slice, a chain-beacon nonce, a catalog snapshot, two reference arms, one king's arm and up to six challenger arms. (§6.2, §8)

Slice — the window's tasks, drawn stratified across the admitted benchmarks from the nonce. Tasks are reusable across windows; freshness comes from the draw being unpredictable. (§6.3b, D12)

Schedule root — one chained-manifest root covering every window's draw and parameters, committed on chain under the owner's hotkey before launch (schedule:v1:<sha256>). The validator refuses to start unless the root it computes matches it. (§6.3)

Nonce / beacon — the chain block hash at the block a window opens; the slice is a pure function of the committed entry and the nonce, so nobody chooses it. (§6.3)

Arm — one policy run over a slice: the king's, a challenger's, or a reference policy's. Every arm is metered and reconciled against its receipts.

Power gate — two fixed reference policies run on a subsample before any artifact is scored. If they cannot be told apart on this slice, the window cannot rank anyone: no duel, no shot spent. (§5.6)

King₀ — the best fixed routing policy found by the admission sweep (today a cheap-then-escalate cascade). The baseline every challenger must beat; while it reigns its 0.85 burns. (§5.5, §8b.8)

Duel — king versus challenger on the same slice, same scaffold, same catalog, each paid on its own key. (§5.2)

Verdict — the four conditions a challenger must clear: paired Δfinal > ε, bootstrap lower bound > 0, leave-one-out > 0, median per-benchmark Δ > 0. Ties break on earliest commit block. (§5.2)

ε (eps) — 0.05 on the final scale; the anti-noise margin on the headline comparison. Behaves like a 7–8 point bar at ~250 tasks per arm. (§5.2b, D10)

Leave-one-out — breadth condition: the win must survive dropping the challenger's single best benchmark. (§5.3)

Median condition — breadth condition: the median per-benchmark delta must be positive — the typical benchmark moved the challenger's way. At two benchmarks it equals the aggregate, a recorded limit. (§5.2, D16)

Coronation — the largest winning delta in a window takes the crown. A reversion to King₀ is not a coronation and never enters the lineage. (§5.5, §8 step 4)

Pension / pensioners — the five most recent ex-kings, paid 0.05 / 0.04 / 0.03 / 0.02 / 0.01 by recency. Unfilled or unresolvable slots burn. (§5.7)

Burn — emission sent to UID 0. The king's 0.85 while King₀ reigns; any pension slot that is empty or whose hotkey no longer resolves. Never redistributed to the king.

Scoring

quality_b — mean graded score over a benchmark's tasks in the slice; quality is the equal-weight mean over benchmarks. (§5.1)

final_b — quality_b − λ_b · spend_b / C_b; final is the mean over admitted benchmarks. The number every verdict runs on. (§5.1b)

λ_b (lambda) — the measured exchange rate between quality and dollars on a benchmark: the slope between always-cheapest and King₀ from the fixed-policy sweep. Never chosen. (§5.1b, reference.fit_pool)

C_b — the mean per-task spend of the λ pair on that benchmark; the unit spend is priced in.

Dominator — a benchmark where the best fixed policy buys no quality over the cheapest model; excluded from scoring because spending there cannot be priced. (§5.1b, §6.1)

Outcome table — the window's shared record of real worker outcomes, one row per (task, model, attempt). The first arm to ask buys it live on its own key; every later arm is handed the same row at the same price and no provider or grader is called again. It makes identical policies tie exactly, so a duel scores routing rather than the luck of separate draws. Per window, never carried across windows. (§5.2c)

Charged against paid — an arm is charged for every delegate it made, hit or miss, and that is the spend_b the score prices; the provider is paid only for the rows that arm was the first to ask for. (§5.2c)

Exhaustion (budget_exhausted) — an arm's key or cap ran out mid-slice; remaining tasks score zero and are published as such, distinct from a stalled model and from the validator's own clock. (§4, §5.8)

Reconciliation — every arm's metered spend must equal its recorded receipts, or the arm is unscoreable and the reveal says so.

Entry

One shot — a hotkey gets exactly one submission, ever; the on-chain ready signal at the end of submit spends it. (§7)

Admission — the gate a tree passes before it is served: manifest against commitment, every file hashed as it lands, config.json field by field, tokenizer by digest, tensor inventory, safetensors only, no bundled code. (§1.1, v3/admission.py)

Manifest — the signed list of a submission's files with sizes and SHA-256s, uploaded last; its digest is what the ready signal commits to. (v3/store.py)

Mailbox — the prefix on the subnet's public store (store.thirtyspokes.ai) where a credential envelope, sealed to the miner's ed25519 hotkey and signed by the owner, is published at a key only the miner's registration derives. (v3/access.py)

Credential — a prefix-scoped, expiring write token for the miner's own upload prefix. A rotation is more time for the one submission, never a second one.

Allowance — what one window may spend on a miner's registered OpenRouter key: the miner's own cap, bounded by what the key reports it can still spend, debited per metered call against a signed receipt. Read at window open. (§4, Economic Model)

Queue — challengers in (commit block, hotkey) order; six judged per window; depth twelve, beyond which commits are refused. Immunity must cover three windows. (§8b.1)

Dev kit — thirtyspokes-dev: the validator's own admission gate, scaffold and scorer, run locally against your conductor. (v3/devkit.py)

Reveal — the owner-signed record of a window, published at v3/reveal/<n>.json: slice, gate, king's arm, every verdict, metering, weights, and the decision traces of the arms the owner paid for. (§5.8, §6.4, dashboard)

Decision traces — per step: rendered state, action, worker outcome, cost — published for the king's and reference arms every window so the training set is free. (§6.4, D15)

Infrastructure

Sandbox host — the machine whose Docker daemon grades: it applies a model's patch and runs a repository's tests. Declared explicitly at validator launch; separate from the box holding the chain key by design. (§8b.9)

Served name — v3-<hotkey>@<manifest digest>: the only channel a served artifact's identity has onto the wire; the validator refuses an endpoint reporting anything else.

Measured image — a reproducible confidential-VM image whose boot chain and rootfs are bound into TDX measurements (MRTD, RTMR1–3); inherited from the first-generation mechanism, the basis of confidential serving.