Mining
Pool attribution is by coinbase address labels with stated confidence. Blocks with no label stay in one “unattributed” row, and every concentration figure below is computed over attributed blocks only: a lower bound, not the whole picture.
Attribution by coinbase output address against Keshi's label table.
Blocks whose coinbase address carries no label are reported as one “unattributed” row. They are never hidden and never redistributed across known pools.
The windowed figures above use pool labels; this series bins the whole chain by height and counts coinbase miner addresses, labelled or not, on two accounting bases. Gaps are undefined measures, never zero.
Per-bin concentration of block production by coinbase miner address, with the unattributed fraction always in view. All-blocks basis: shares divide by every canonical block in the bin. The Nakamoto crossing is undefined when addressed blocks never exceed half the bin.
Unit of analysis: coinbase miner addresses per height bin, over canonical blocks; pool labels play no part. A measure with no value is null, never 0: an empty basis, or a Nakamoto crossing that does not exist because addressed blocks never exceed half the bin. Such bins render as gaps, never as zero. The unattributed fraction (blocks with no decodable coinbase address over all canonical blocks in the bin) is always shown, never hidden, never redistributed. One operator can span several addresses, so address-based top-N shares are lower bounds and the address-based Nakamoto coefficient is an upper bound on operator concentration. Effective operators (inverse HHI) is defined on the attributed basis only. The newest bin is still filling and is flagged partial.
Gross protocol revenue per unit of consensus work: Σ(subsidy + fees) / Σ work over canonical blocks in the window.
Gross protocol revenue per unit of consensus work, not miner profit: hardware, power and payout policy are off-chain. Per-block work is 2^256/(target+1), summed with the work-rate accounting, so the unit is matmul attempts: a difficulty-derived consensus measure, not a measure of AI computation performed. Denominator: canonical blocks in the window that carry a work sample; the window’s full canonical count is shown alongside, and the figure is flagged partial when the two differ (backfill) or the window spans under two blocks.
Instantaneous per-block rate over the 200 window.
matmul attempts per second, a difficulty-derived consensus measure, not a measure of AI computation performed.
Header-to-header spacing against Pearl's 194 s target.
Header timestamps are miner-set with a ±5 minute tolerance, so single-block spacing is noisy by construction; the rolling mean is the readable signal.
Decoded from the proof certificate every block commits. Keshi parses the full corpus back to genesis. Declared dimensions are what the miner claims was multiplied; the ZK proof covers at most tileSize ≤ 256 output entries of the result, never the full declared matmul, and never a measure of AI work performed. Research notes on this corpus →
Per-block noise rank over the 200 window.
Noise rank is a miner-chosen proof parameter, not throughput. The v1.3.0 softfork (height 96,251; no PIP documents it) rejects rank < 128 and multiplies the difficulty bound by 128/rank.
Per-block proven/declared. State which form is quoted; the windowed aggregate is MAC-weighted, not this series' mean.
Share of the declared multiplication the ZK proof actually covers, MAC-weighted over the window (Σproven / Σdeclared). Per block the ratio is purely geometric: tileSize / (m·n).
Declared shapes are what miners claim. Rank 0 is the genesis empty certificate, shown rather than hidden. Canonical blocks only. Orphaned blocks’ certificates stay in the corpus but are excluded here.
The (m, n, k, rank, tile) combinations miners declare most, by block count.
| m | n | k | Rank | Tile | Blocks |
|---|
Coinbase-address attribution; the unattributed remainder is always shown, never redistributed.
| Pool |
|---|
Blocks whose coinbase address carries no label are reported as one “unattributed” row. They are never hidden and never redistributed across known pools.
Certificate eras
Below the MoE fork height (71,935); V1 certificates.
Between the MoE hard fork (71,935, PIP-2) and the dense-only softfork (91,630); V2 certificates, MoE proofs accepted.
At or above the dense-only softfork height (91,630); MoE proofs are rejected again. No PIP documents this fork; it exists only in pearld's source.
The certificate class in force changed three times on mainnet: V2 certificates with MoE support at 71,935, a dense-only softfork that rejects MoE proofs again at 91,630, and V3 certificates with a salted noise seed at 99,000. The strip above is split at the first two, which decide whether MoE proofs are accepted; V3 keeps the same wire layout and changes how the noise seed is derived, so it does not move an era boundary here. Other consensus changes altered certificate validity without changing the class, and the consensus-change register carries all of them with their evidence. A block’s era tells you which certificate rules it was mined under. Read the register.