Keshi is independent community infrastructure for Pearl Network. It runs its own pearld full node and its own Blockbook indexer, ingests every block itself, and computes every figure it publishes from that index. Nothing is scraped from other explorers, and no number ships without a documented source and formula.

Keshi is not affiliated with Pearl Research Labs. Where Keshi’s figures differ from official or community explorers, the difference is usually the estimator window, which is why the window is always displayed next to the number it shaped.

Every metric Keshi displays is defined in docs/metrics.md with its source, formula, window and caveats, and each API payload carries the anchor it was computed from. If a number has no entry there, it does not ship.

The small i beside every figure opens that record: the source it was read from, the formula anchor it was computed with, the moment it was computed, and any caveat that shaped it.

pearld (our node)

Chain ingest: blocks, transactions, mempool, peers

Mempools are per-node; ours is labelled as exactly that.

Blockbook (our instance)

Address summaries, cross-checked heights

Address pages are a verbatim pass-through until Keshi’s own address index ships.

Keshi index (PostgreSQL)

Every chart and metric: work rate, difficulty, intervals, fees, issuance, pools

Money is integer grains end to end; floats never touch a monetary value.

CoinGecko

Market ticks only

Collected on its own schedule, never merged into chain metrics, and clearly framed as external data.

Pearl secures its chain with proof of useful work. Instead of hashing a nonce until a header falls below a target, a Pearl miner performs tiled INT8 matrix multiplications with deterministic low-rank noise, then searches for a tile whose BLAKE3 commitment satisfies the consensus predicate. The winning work is wrapped in a recursive zero-knowledge proof and carried in the block as a certificate, versioned rather than nonced. Blocks arrive on a 194-second target and difficulty adjusts with every one of them.

The appeal is obvious: the same GEMM kernels that secure the chain are the kernels that run neural networks, so the electricity could in principle buy two things at once. That is a statement about protocol capability. Whether it also describes what miners are actually doing is a separate question, and it is the question Keshi refuses to blur.

Nothing on-chain proves that any particular block’s matrices served a real workload. The predicate accepts any matrices that satisfy it: an acknowledged property of the design, examined empirically in arXiv:2606.04819 and unrebutted as of 2026-08-02. So Keshi labels the security throughput “PoUW work rate” and footnotes every EH/s figure as matmul attempts per second: a difficulty-derived consensus measure, not a measure of AI computation performed. If verifiable attestation ever exists, we will publish the methodology first and the number second, and we will name precisely what it attests: recomputing a block’s weight commitment proves which operand was committed, never that inference ran or that a customer was served.

EH/s on Keshi means matmul attempts per second: a difficulty-derived consensus measure, not a measure of AI computation performed.

docs/metrics.md: language policy, binding on every surface

Genesis

2026-04-27

Target block interval

194 s

Timestamps

Always UTC, never the viewer’s timezone.

Money

Integer grains end to end; floats never touch a monetary value.

This is Keshi v0.1. Explorer depth, an own address index, mining intelligence and a public API programme arrive in later phases.