Pearl: why it forked four times in 27,000 blocks
Pearl (PEARL) changed its consensus rules four times in 27,065 blocks, from block 71,935 to block 99,000. Four upgrades that close together can look like a chain in trouble. It is closer to the opposite.
All four fix one problem from four sides. Each makes a mining proof pay for what it claims about the GPU's work. The last one closes the gap.
Two were hard forks and two were softforks. That changed what a miner had to do, not why the upgrades happened. Every height, rule and version here is read from the chain's own consensus source, not from an announcement.
Pearl's four upgrades in short
- Block 71,935, hard fork: proofs moved from the v1 certificate to v2, which can describe a mixture of experts.
- Block 91,630, softfork: the network accepts dense proofs only, for now.
- Block 96,251, softfork: a higher noise rank no longer buys an easier target, and a rank below 128 is refused.
- Block 99,000, hard fork: the noise seed is salted with the row and column counts.
- Each one activated at a fixed block, so updating the miner before that block was enough.
| Block | Type | What changed |
|---|---|---|
| 71,935 | Hard fork | Proofs move from the v1 certificate to v2, which can describe a mixture of experts and carries the noise rank. |
| 91,630 | Softfork | Mixture of experts proofs are rejected. Only dense proofs count. |
| 96,251 | Softfork | The target no longer eases as the rank rises. A rank below 128 is refused. |
| 99,000 | Hard fork | The noise seed is salted with the row count and the column count. |
What Pearl asks a miner to do
Pearl launched its mainnet on 27 April 2026. It is a layer one chain secured by matrix multiplication rather than hashing. That multiply is the dense linear algebra that AI training and inference spend almost all of their time on.
A miner runs a noisy matrix multiply on the GPU and hashes the result into a commitment. It then wraps that work in a zero knowledge proof. The network can check such a proof cheaply, without redoing the work.
This makes Pearl a form of proof of useful work. Underneath sits a fork of Bitcoin's btcd node. That base layer is deliberately boring, so the new idea runs on old, well understood plumbing.
Why the shape of the work matters
With a hash, there is exactly one way to do the work. One number says how hard it was.
A matrix multiply is different. The miner must also declare the shape of the work: how large the matrices are, and how much low rank noise is folded in. Noise is extra data, derived from the proof itself, that the miner multiplies against.
The chain never sees the GPU. It sees a proof and a description of the work behind it. Every upgrade is about that description, and about making each part of it cost something to claim.
Block 71,935: the mixture of experts certificate
The first upgrade was a hard fork at block 71,935. It moved proofs from the v1 certificate format to v2. A hard fork is a rule change that old node software cannot follow.
The upgrade is named for what v2 can describe: a mixture of experts. That is the shape real inference on a large model takes. Only a few experts out of many run for any given token.
A format that only describes one big dense multiply can only describe part of that. The v2 certificate also added a new field, the noise rank. That field made the third upgrade possible.
Block 91,630: dense proofs only, for now
The second upgrade was a softfork at block 91,630. A softfork only tightens the rules, so old nodes still accept the new blocks. From this block, the network rejects mixture of experts proofs and accepts dense ones only.
That looks like a reversal, but it is not. The certificate stayed at v2 and kept everything it had gained. What changed is which shapes count while the chain is young.
The project has not published its reasons for the timing. So we say that plainly rather than offer a guess.
Block 96,251: a higher rank stops paying
The third upgrade was a softfork at block 96,251. It is the one a miner can feel in numbers. The rank sets how much low rank noise is folded into the multiply.
Before this fork, the target a share had to beat was eased by the tile size times the full dot product length. Turning the rank up therefore bought a proportionally easier target. The GPU did no more real work for it.
In other words, one dial set two things. It decided how much of the computation was genuine. It also decided how many shares a miner was paid for.
From block 96,251, the bound is the tile size times the dot product length, divided by the rank times 128. Any rank below 128 is refused outright.
The old and new bounds differ by exactly 128 divided by the rank. So the effect on an old setting is arithmetic, not an estimate. At rank 256 the target is twice as hard for the same work, and at rank 1024 eight times.
A rule that tightens what counts as real work can also retire hardware. The new rule must be computed bit for bit on whatever card runs it. The chain's source says nothing about cards, so this post names none.
Whether a GPU generation lost Pearl support here is a question for each miner's release notes. Check them before you buy a card for one coin.
Block 99,000: the noise seed is bound to the shape
The fourth upgrade was a hard fork at block 99,000. It is the smallest change, and the easiest to describe wrongly. The noise is derived from the proof's two Merkle roots, single hashes that sum up a larger set of data.
Before this fork, the roots went into that derivation raw. So the noise did not depend on the matrix dimensions the proof commits to.
From block 99,000, each root is first salted, which means mixed with extra data. One root gets the row count and the other gets the column count. Only then does the seed chain run.
A given piece of noise now belongs to exactly one declared shape. The certificate layout is byte for byte the same as v2, and the wire still carries the raw roots.
Only the version number selects which derivation applies. That is why this fork asked miners for nothing but a new binary.
What the four add up to
The chain cannot watch the GPU. So everything a proof claims about itself has to cost something to claim. The four upgrades say that one sentence four ways.
- Mixture of experts widened what a proof can describe.
- Dense only narrowed what counts while the chain is young.
- The rank penalty stopped one dial in the claim from buying an easier target.
- The salted seed stopped one piece of noise from serving more than one claimed shape.
A young chain that finds these gaps in its first months, and fixes them inside 27,000 blocks, is a good sign. It is a better sign than a young chain that has found none.
What to do if you mine PEARL
All four upgrades activated at a fixed block with no transition period. All four also failed the same quiet way. An outdated miner kept showing a healthy hashrate while its shares stopped counting.
- Follow Pearl's releases on GitHub for the next activation block.
- Update your miner before that block, not after it.
- On fork day, watch your accepted share count, not the hashrate on your screen.
- If you merge mine, take extra care. An invalid Pearl share carries no auxiliary block either, so the loss doubles.
On our side, each upgrade was in place before its fork block. So the only thing a miner had to change was the miner binary. Same address, same port, same wallet.
For a later Pearl rule change, read our explainer on the FP8 fork and PIP-3.
Pearl consensus upgrades: questions and answers
Why did Pearl change consensus four times so quickly?
Each change closed part of one gap. The chain cannot see the GPU, so every claim a proof makes about its work must cost something to claim.
Which Pearl upgrades were hard forks?
Blocks 71,935 and 99,000 were hard forks. Blocks 91,630 and 96,251 were softforks.
What is the noise rank in Pearl mining?
It sets how much low rank noise is folded into the matrix multiply. Since block 96,251, a higher rank no longer buys an easier target. A rank below 128 is refused.
How much harder did the rank penalty make the target?
By 128 divided by the rank. At rank 256 the target is twice as hard for the same work, and at rank 1024 eight times.
How can I tell that my miner missed a fork?
Its hashrate still looks healthy, but its accepted shares stop counting. Watch the accepted share count on fork day.
Did I need a new address or port after any fork?
No. On our pool the address, port and wallet stayed the same through all four. Only the miner binary had to change.
Where these numbers come from
Everything here is read from Pearl's source code on GitHub. The four heights, the block time and the mainnet launch date come from the network parameters and the genesis definition.
The rank rule and the 128 divided by rank figure come from the proof verifier's sanity checks. The old and new bounds sit side by side there. The certificate versions and the salted seed come from the wire format and the seed module.
Anything the source does not state is left out of this post.