Glossary
Terms as they are used in Thirty Spokes, with pointers to where each is specified.
Core mechanism
Agent — the miner's router. A single Python file defining build_agent(weights) -> agent(prompt, call_model). See Build a Router.
Artifact / bundle — the complete submission: source file + weights.bin. Public, and cryptographically bound to what ran. See Anti-Cheat.
KOTH — King-of-the-Hill. The competition format: challengers must beat the reign to enter it. See The KOTH Protocol.
Reign — the king + up to 4 recent ex-kings. Emissions split equally across whoever is currently paid (20% each when all 5 slots are full). See Economic Model.
King — the slot-1 holder. Defended by the dethrone guard.
Epoch — EPOCH_BLOCKS = 100 blocks ≈ 20 minutes. The unit of competition: one task slice, one proof, one scoring pass. See The KOTH Protocol.
Pool / spoke — the owner-pinned allow-list of provider models an agent may call. A "spoke" is one such model. Off-list calls raise UnpinnedModelError.
Hub — the empty center. Holds no model, does no inference. The design constraint that lets the network route to any provider without a conflict of interest. See Using the Gateway.
Scoring
Q_lcb — the miner's score: a cost-gated, weighted sum of per-benchmark lower confidence bounds, not raw accuracies. You are scored on the bound, so variance costs you. See The KOTH Protocol.
LCB — lower confidence bound, from a bootstrap (alpha = 0.05, boot = 1000). The conservative estimate of your accuracy.
Budget — the eligibility cost ceiling (default $0.50 per epoch slice). Cost is a gate, not a divisor: under it, cost does not enter the score; over it, you are excluded. See Economic Model.
f_min / floor — the per-benchmark accuracy floor (default 0.10). Fall below it on any benchmark and you are disqualified (below_floor:<b>) — benchmark-abandonment is not a cost strategy.
ε (epsilon) — the margin a challenger must beat an incumbent by. Decays with the incumbent's artifact age from 0.02 to 0.002 (tau = 8.0). See Economic Model.
Dethrone guard — the clamp protecting the king from noise-driven churn: a challenger needs a confident per-benchmark gain and must stay within king_cost × 1.10. Emits no_confident_gain, regression:<b>, cost_regression — a clamp, not a disqualification.
Burn — emission sent to uid 0 when a slot has no legitimate claimant (too few miners, deregistered hotkey, unmappable uid). See Economic Model.
Evidence accumulation — an opt-in scoring mode (scoring_mode="accumulate") that pools a fixed artifact's verified per-epoch results into one decayed binomial (Wilson lower bound, EWMA half_life_epochs default 200), so a noisy 32-task epoch ranks stably and decoupled validators agree. Re-committing resets the accumulator; a missing epoch counts as 0 correct.
Benchmarks
Suite — the owner-given benchmarks. The full suite (offline sim + dev kit) is four: math 0.30, MMLU 0.22, GPQA 0.23, SWE 0.25. The live loader (miner + validator daemons) currently wires two — math + MMLU at 0.5 / 0.5 — because the GPQA and SWE real loaders are still seams (GPQA HF-gated, SWE needs the Docker grader).
Task slice — the n_per_bench (default 8) tasks per benchmark that this epoch's nonce selects. Unpredictable until the epoch's start block is minted.
Grounding check — the default memorisation backstop: every scored answer must match a pool response logged in the attested trace, else ungrounded. Pure proof-inspection — the validator runs no miner code. See Anti-Cheat.
Held-out probe — the opt-in memorisation upgrade (--probe-bank, audit_mode="probe"): a secret, owner-curated bank the validator re-executes the artifact against, DQ memorization on a collapse. Its commitment (probe_commit) is on-chain; in probe mode a validator without the matching bank refuses to score.
Attestation
TEE — Trusted Execution Environment. Here, an Intel TDX confidential VM. The hardware that lets a miner run its own benchmarks and still produce an un-forgeable proof. See Attestation.
TDX — Intel Trust Domain Extensions. The specific confidential-VM technology used.
Quote — the hardware-signed attestation binding {measurement, report_data}. The root of the whole trust chain.
DCAP — Data Center Attestation Primitives. The verification path from the quote up to a pinned Intel SGX Root CA, including TCB status, CRL, and QE identity (via dcap-qvl). See Attestation.
MRTD — Measurement of the Trust Domain: the launch measurement of the confidential VM image (firmware + initial state). The load-bearing runtime anchor. Gated against the owner's on-chain set; a mismatch is unapproved_runtime.
RTMR0–3 — Runtime Measurement Registers. RTMR1/2 capture measured boot (kernel, cmdline); RTMR3 is runtime-extended by the KOTH runtime with a domain-separated digest of the runtime measurement, suite version, and pinned pool — SHA384(0×48 ‖ SHA384("koth-rtmr3-v1" | runtime | suite | pool)). All gated.
report_data — the 64-byte field the runtime binds into the quote: sha256(proof-minus-quote). It is what makes editing the proof after attestation break the signature (report_data_mismatch).
TCB — Trusted Computing Base. Its status (UpToDate, SWHardeningNeeded, OutOfDate, …) says whether the platform's firmware/microcode is current. Policy default accepts {UpToDate, SWHardeningNeeded}.
Measured image — the reproducible, owner-pinned VM image whose MRTD/RTMR1/RTMR2 the owner publishes. Building and pinning it is the hard prerequisite for secure mainnet mining. See Attestation.
Confinement / no-egress — running the agent in a network namespace with zero egress, so its only channel is call_model (IPC to the metering parent). Enabled with --confine. See Anti-Cheat.
Chain
Bittensor — the proof-of-intelligence blockchain the subnet runs on. Provides identity (hotkeys), the emission mechanism, and the commit/weights transactions. See Bittensor Integration.
Subnet / netuid — one competitive market on Bittensor, identified by its netuid. Thirty Spokes is Subnet 99 on mainnet.
Hotkey — a miner's or validator's on-chain identity. Salts the commit and binds the proof, so copying a rival's commit or proof binds nothing (bad_commit, hotkey_mismatch).
Commit-reveal — commit a hash of your artifact now, reveal the artifact after a delay. Stops front-running; the hotkey salt stops commit-copying. Format koth1|<repo>|<revision>|sha256(...). See Anti-Cheat.
Beacon / nonce — the unpredictable per-epoch seed derived from the block hash at epoch start. Seeds the task slice; kills best-of-N and pre-computation. See Anti-Cheat.
Commit window (F7) — an opt-in intra-epoch anti-grind bound (commit_window=W): the miner commits its proof's report_data() on-chain inside [epoch·blocks, +W] and must reveal exactly it, capping best-of-N grinding to W's wall-clock. DQs no_proof_commit / commit_out_of_window / commit_mismatch.
Grace window (F2) — an opt-in setting (grace_blocks=G) that scores a settled epoch — the latest whose grace deadline has passed — rather than the live one, so submissions settle and decoupled validators agree.
Owner — the subnet operator. Pins the pool, publishes the approved-measurement governance record, and curates the secret probe bank. Uses orchestra-koth-owner.
uid — a miner's slot index in the subnet metagraph. uid 0 is the burn address.
Roles
Miner — competes by publishing an attested proof each epoch. Runs orchestra-koth-miner. See Mine the Subnet.
Validator — verifies proofs and grades reports; performs no inference of its own. Runs orchestra-koth-validator. See The KOTH Protocol.
Dev kit — orchestra-koth-dev, the local scorer that runs the exact validator scoring on your artifact so you can optimise against the real Q_lcb. See Build a Router.