Proposal for checkpoint validators that verify finalized Solana state without slowing it down

I am proposing a second-class tier of validators that never leads blocks or votes per slot. Instead they replay every block and sign a finalized state only every 30 seconds, so they can run on home hardware with a retail ISP. As a tradeoff, they pay no VAT, receive no leader fees/tips but earn normal staking rewards.

The benefit to Solana is a public alarm if the primary validator set finalizes a wrong state, a widely distributed and verified recovery point after a network halt, and a more natural way for new validators to prove operations and grow stake before stepping up into lead validators.

The full proposal below is drafted by Claude Opus:

Motivation

Four gaps in today’s network would each narrow with a cheap, independent tier of verifiers.

  • One client family. About 84% of stake runs Agave or clients built from it, and the Firedancer family about 16%, part of which (Frankendancer) still executes with Agave’s runtime. A bug in the majority client could finalize a wrong state, and nothing today raises a public, signed alarm.
  • One region. About 73% of stake sits in Europe, 55% in Germany and the Netherlands alone. Alpenglow needs 60% of stake to make progress, so a European outage stalls the chain. A checkpoint tier cannot prevent that stall, but it keeps signed state and snapshots outside Europe for the restart.
  • Restarts. A stalled cluster restarts by hard fork at a slot operators agree on off chain. The documented hard-fork flow still has an open Alpenglow bug (agave#16359, opened 2026-10-23).
  • A shrinking long tail. Validators fell from about 2,560 in March 2023 to about 700 now. With the VAT (about 292 SOL a year) and server costs, a validator needs roughly 90k–110k SOL of stake to break even at 5% commission (estimates vary).

Stake and geography are from public RPC and IP geolocation on 2026-11-02; the client split is from Stakewiz the same day.

Design

A checkpoint validator is a staked vote account in a new checkpoint mode. It is never scheduled as leader, never votes per slot, and pays no VAT.

Joining

  • Switch a vote account to checkpoint mode and register a BLS key and a client tag (the software it runs).
  • Hold a minimum active stake, proposed at 5,000 SOL, own or delegated. Stake accounts, delegation, commission and payouts work as they do today.

Duties

  • Replay every block with its own client. Running minutes behind the chain is fine.
  • Every 150 slots, sign (slot, block id, bank hash) for that finalized slot, plus its own custody value (see Proof of replay). At each epoch boundary, also sign the next epoch’s stake-set hash and a snapshot hash.
  • Send signatures to a few rotating aggregators. They merge them into one BLS signature plus a signer bitmap and post it on chain.

Proof of replay

  • A signature alone does not prove the work. A member could copy another member’s bank hash and sign it with its own key, and the signature would still be unique.
  • So each checkpoint message also carries a custody value that only that member can produce. The member signs a seed message for the checkpoint slot with its BLS key, and the hash of that signature is its seed. BLS signatures are deterministic, so anyone can check the seed against the member’s public key.
  • The seed picks a few hundred accounts from those written during the 150-slot window, a different set for every member. The custody value is a hash of their state at the checkpoint slot.
  • Another member’s value is no help, and public RPC cannot reliably supply account state at an exact past slot. In practice the member has to replay those slots, or buy a live feed of every account update (see Risks).
  • At each checkpoint, every member recomputes the custody values of a few randomly chosen peers. A wrong value becomes a signed claim that any replaying node can check. Once enough members confirm it, the member forfeits that epoch’s rewards and is removed.
  • Custody values differ per member, so they stay out of the BLS aggregate. Aggregators post a Merkle root of them on chain and publish the full list off chain.

Alarms

  • A member whose replay gives a different bank hash for a finalized slot publishes a signed mismatch at once, without waiting for a checkpoint.
  • A public alarm needs mismatches from at least two independent clients, or a re-executable proof. One immature client cannot raise it alone.

Promotion

  • A member moves to the main set by switching to normal mode and paying the VAT. Its delegators move with it, so the tier is a working bench for the main set.
Main validator Checkpoint validator
Votes Every slot, sent to every other voter One signature per 150 slots, sent to a few aggregators
Timing Within a slot (200 ms) Minutes behind is fine
Leader slots Yes, in proportion to stake None
Admission ticket 0.8 SOL per epoch, ~292 SOL a year None
Hardware 12 cores / 24 threads and 256 GB+ RAM (Anza minimum) Desktop class plus ~1–2 TB of fast NVMe; Mithril replays mainnet on 8 cores and 64 GB in its developers’ tests
Network 2 Gbit/s symmetric minimum, 10 Gbit/s recommended ~41 Mbit/s raw block feed, or ~3.4–6.6 Mbit/s as a compressed stream
Stake to break even ~90k–110k SOL at 5% commission ~3k–6k SOL at 5–10% commission (estimate)

The block-feed figures were measured on mainnet blocks on 2026-11-11; the checkpoint break-even assumes ~$1,000–1,500 a year for power and hardware, at ~$75 per SOL.

How a checkpoint flows

The diagram at the top shows the flow. Checkpoint validators replay what the main set finalized and sign it; aggregators post one combined signature per checkpoint. A mismatch goes out as a public alarm at once, and the signed record gives operators a restart point after a halt.

Checkpoint schedule

Checkpoints run on two cadences: every 150 slots (about 30 seconds) and at each epoch boundary. Signatures for a checkpoint are accepted for about 10 minutes, so members on cheap links can lag and sign in batches.

Cadence Checkpoints per epoch Average wait until attested Verdict
Every slot 432,000 About the members’ lag No: duplicates main-set voting
Every 150 slots (~30 s) 2,880 ~1.5–5.5 min Proposed default
Every ~10 min ~144 ~6–10 min Workable if users accept the wait
Once an epoch 1 ~12 h Too slow alone, and pays for minimal participation

Waits assume members run 1–5 minutes behind. Below about 30 seconds the wait is set by member lag, not by the interval.

The epoch-boundary checkpoint serves the jobs that need a stake set or a snapshot: light clients following the chain from epoch to epoch, new nodes checking a downloaded snapshot, and defense against forged old history.

Frequent checkpoints cost almost nothing:

  • Computation: almost none. Every slot’s bank hash already commits to the full account state through the accounts lattice hash, and a custody value adds hashing a few hundred accounts per checkpoint.
  • Traffic: 4,000 members signing every 30 seconds send about 133 signatures a second, about 27 KB/s for the whole tier.
  • On chain: one aggregate per checkpoint, about 2,880 transactions a day. A 4,000-member signer bitmap is 500 bytes and fits one transaction.
  • Custody values: 32 bytes per member per checkpoint, about 4 KB/s for 4,000 members, published off chain under an on-chain Merkle root.
  • Pay: each counted signature earns a credit, so rewards follow real participation, as vote credits do today.

The 150-slot interval and the 10-minute window are design choices to tune in a pilot, not measured values.

Rewards

Checkpoint stake earns the full staking rate through today’s reward payout. A lower rate would leave the tier with no delegators. Main validators can pass on full emissions, because extra stake costs them nothing.

Credits are scaled so a member that signs every checkpoint earns the same per SOL as a main validator that votes every slot. Inflation is a fixed pool: stake moving over from the main set changes no one’s yield, while newly staked SOL dilutes all stakers slightly, as any new stake does.

At the full rate the two tiers stay close for delegators:

Where the stake sits Delegator yield per year
Main validator, 0% commission ~5.25% (5.24% inflation + ~0.01% tips)
Checkpoint validator, 1–2% commission ~5.14–5.19%
Main validator, 5% commission ~4.99%
Main validator, 10% commission ~4.72%

Operators sort themselves by size. Switching to checkpoint mode gives up leader fees, which grow with stake, and saves about 420 SOL a year of VAT and server. So switching pays below about 420 SOL divided by the leader fee rate:

Leader fees (% of stake per year) Switching pays below Likely to switch
1% (fee boom) ~42k SOL Under 1% of stake
0.21–0.31% (now) ~140k–200k SOL ~300–400 validators, ~5–9% of stake
0.1% (fee drought) ~420k SOL ~18% of stake; the main set shrinks to ~160 validators

All figures in this section are estimates. Inflation yield is from epoch 1083 (2026-11-10). Leader fees come from samples of 71 blocks on 2026-11-10 and 300 blocks on 2026-11-14, and swung 3× between samples. Jito tips were ~130 SOL a day in the 2026-11-14 sample. Tips paid to landing services (~1,030 SOL a day) are left out, because how much reaches validators is unknown.

Limits

The tier must never finalize blocks, even during a halt. Suppose Europe were cut off from the rest of the internet but still running. Its 73% of stake would keep finalizing above the 60% threshold, and a checkpoint tier finalizing elsewhere would create a second final history.

  • Checkpoint stake never counts toward main-set quorums, in any situation.
  • The tier adds no per-slot safety. A later checkpoint cannot overturn a block that is already final.
  • It makes the chain neither faster nor more censorship resistant, because members never produce blocks.
  • Its economic weight is small. At around 5–9% of stake, an attestation is backed by far less stake than a main-set certificate; its value is the independence of its operators, clients and countries.

Risks and open questions

The hardest problem is proving that each member really replays; custody values close most of it, and the rest are parameters and trade-offs.

  • Copying instead of replaying. Custody values stop members copying each other or public data. Two gaps remain: a member could buy a full account-update stream from a node operator instead of replaying, and one operator could run many members off one node. Making an early leak of a member’s seed punishable, as in Ethereum’s proof-of-custody design, discourages handing the seed to a service.
  • Spot checks. As far as we know, the lattice hash offers no single-account proofs, so peers check custody values by recomputing them. Open: how many accounts each value covers, how many peers each member checks, and how clients pick the accounts cheaply.
  • Enforcement. Only nodes holding the state can check a custody claim or a mismatch; an on-chain program cannot. Open: confirm claims by a stake-weighted vote of members, or enforce off chain during the pilot.
  • Client tags. Tags are self-reported, so the two-client rule for alarms can be gamed. Open: how to tell which client a member really runs.
  • No stake to burn. Solana has no slashing. Signing two hashes for one slot is trivial to prove, but without slashing the penalty is ejection and lost rewards. Open: should operators post a small deposit of their own, or is this a good first, narrow slashing condition?
  • Fewer small leaders. At today’s fees, ~5–9% of leader slots would move from many small operators to fewer large ones.
  • Less VAT burned. 300–400 validators no longer paying the VAT would cut the burn by roughly 88k–117k SOL a year (estimate).
  • Fee droughts. At 0.1% leader fees, about 18% of stake could switch tiers and the main set would shrink to ~160 validators. Whether checkpoint stake then needs a cap is left for a follow-up proposal.
  • On-chain verification. Open: can programs verify a BLS12-381 aggregate signature cheaply enough for one every 30 seconds? If not, aggregates are published off chain first.
  • Parameters. Open: the minimum stake, the cadence and the acceptance window and how aggregators are chosen all need settling in a pilot.

Rollout

The tier can start without any protocol change. Each phase is useful on its own, and only the last needs a SIMD.

  1. Off-chain pilot. Volunteers who already run RPC nodes sign (slot, block id, bank hash) and a custody value every 150 slots. They publish aggregates and mismatches on a public feed, funded by grants. This tests the alarm, the snapshot check and real member lag.
    • Gate: at least 30 epochs (about a month) of alarms with no false positives, with members on two or more clients.
  2. On-chain registry. A program, with no feature gate, holds member keys and client tags and receives aggregates and mismatch reports. Exchanges, bridges and new nodes can start reading them.
    • Gate: real users of the attestations, and tuned values for the minimum stake, cadence and acceptance window.
  3. Protocol integration. A SIMD and a feature gate add checkpoint mode to vote accounts: no leader slots, no VAT, credits per counted signature, custody checks, rewards at the main rate, and switching to and from normal mode.

Alternatives considered

The closest alternative, per-slot voters that never lead, costs the network as much as a full validator; the others are complementary.

Alternative Why not instead
Per-slot voters that never lead and pay no VAT Every vote goes to every voter, so each costs the network as much as a full validator. Free seats invite dust-stake voters, and deliberately late votes miss the certificate
Lower VAT for everyone (discussion #730: 0.8 → 0.4 SOL per epoch) Helps, but break-even stays ~50k–60k SOL and colocation hardware is still needed. Complementary
Geo-aware leader schedule (SIMD-0675) Speeds handovers between distant leaders, but adds no verification or recovery. Complementary
Regional criteria in Foundation and stake-pool delegation Moves stake, which is what consensus geography needs, but adds no verification or recovery. Complementary
Rely on existing RPC nodes They already replay, but a disagreement shows up as nodes stopping and posts in chat, with no signed public record

Sources

  • SIMD-0326: Alpenglow. 80% one-round and 60% two-round finality, the “20+20” security model with 40% crash resilience, votes sent directly to all validators, BLS aggregate certificates.
    github.com/solana-foundation/solana-improvement-documents/blob/main/proposals/0326-alpenglow.md
  • SIMD-0357: Validator Admission Ticket. 1.6 SOL per epoch deducted from the vote account and burned, voting set capped at the 2,000 highest-staked validators.
    github.com/solana-foundation/solana-improvement-documents/blob/main/proposals/0357-alpenglow_validator_admission_ticket.md
  • Solana Compass on the VAT halving request. VAT scaled to 0.8 SOL per epoch at 200 ms slots (SIMD-0525), discussion #730 opened 2026-11-10. Linked in the Alternatives table above.
  • Solana Compass on SIMD-0675 and CryptoSlate on the geo-aware schedule.
    solanacompass.com/news/simd-0675-analysis-recommends-a-10-stake-floor-and-a-run-length-of-3-to-7-for-solanas-geo-aware-leader-schedule
    cryptoslate.com/solanas-geographic-speed-plan-trusts-validator-locations-the-network-cannot-verify/
  • Agave issue #16359. Replay after --hard-fork checks pre-fork certificates with the post-fork shred version; open as of 2026-11-14. Linked in Motivation above.
  • Anza validator requirements. 12 cores / 24 threads, 256 GB+ RAM, 2 Gbit/s symmetric minimum.
    docs.anza.xyz/operations/requirements

Credits: most of the technical design in this draft came from Claude Opus 5.5 (Anthropic), inspired by a discussion with human. Human asked the starting question: validators that vote without leading, as a backbench and for geographic spread. He also steered three key choices: stake instead of a bond, full staking rewards for the tier, and per-member work so members cannot copy each other. Claude proposed the checkpoint tier, its cadence, custody values, the economics and the limits, and wrote the draft. Figures were checked against public sources but have not been reviewed by Solana core developers.