Stake-Weighted F3 Finality·Discussion on the FIPs repo

Beyond the chain: operationalizing stake-weighted finality

Stake-Weighted F3 Finality lays out the changes required to the chain actors and node clients: the power table for F3, the fast-finality protocol Filecoin’s committee runs, is derived from staked FIL instead of quality-adjusted storage power. This document describes the additional off-chain tasks that will make stake-weighted finality a successful program. They all serve one goal, the activation threshold: about 80M FIL staked and dispersed, the level at which certificates become authoritative.

What the protocol ships

On chain, the FIP ships a staking actor recording a staked balance and a registered BLS signing key per participant. The actor supports deposit, withdrawal behind an unbonding delay and delegation, and holds the activation state the finality switch reads. Power-table construction changes in one input. A reward is paid per included signature and carried on chain by a block producer who submits each certificate as an ordinary message for a smaller includer share. Equivocation slashing scales with how much stake equivocates in the same window.

The FIP names the parties this work falls to: delegated staking operators, custodians, exchanges, wallets and dashboards.

Filecoin settles on Expected Consensus today. Recruitment speed sets when fast finality starts. Certificates stay advisory until the threshold is met, the state the network already lives in, so safety holds at today’s level.

The participation API

Certificate signers are a bitset over the instance’s power table, so reading “who signed” means reconstructing the ordered table for that instance. Per-participant inclusion accounting, and an operator-level view of it in the CLI, are new work in go-f3 v0.8.14, lotus and forest as of 2026-08-05.

The participation API is the feed that turns a signer bitset into “this participant signed.” Payment checking reads it, since the reward grows only when a signature lands in a canonical certificate, denominated in the epochs that certificate newly finalizes. So do pool operators, whose revenue is per included signature, and concentration monitoring, since the dispersion constants are re-checked against who controls signing weight rather than against address counts, which are cheap to manufacture.

It is also how a staker diagnoses earning nothing: a participant holding less than 1/65535 of total stake (roughly 1,526 FIL at a 100M-FIL pool) scales to zero power, still listed in the power table as a positive-power entry and silently unable to sign.

The feed is new work in every client, built in two languages and held identical by shared conformance vectors.

Delegated staking operators

A depositor buys one product: FIL in, a share of the participation reward out, without running a node or holding a signing key. The FIP calls this expected and healthy, since the operator performs the real signing work. The certificate condition disallows collecting on idle balances. The operator’s obligations, as the FIP states them:

BLS key registration and rotation. Rotation must be delayed so a certificate signed under an old key stays attributable, on the model of the miner actor’s rotation delay; without it an operator could repudiate a signature. Today a single non-BLS entry aborts power-table construction in both lotus and forest, so admitting a non-BLS address is a one-registration denial of fast finality for everyone.

Running a participant whose revenue is per included signature. Signing alone earns nothing; the gate is inclusion in a canonical certificate, which makes inclusion monitoring the operator’s revenue reporting, depositor accounting and uptime alerting at once. In lotus the signing key must sit hot in the node daemon’s wallet, and remote signing is marked in the code as future work.

Reward accounting and a share back to depositors. Claims are pull-based, so the posted rate is gross of the includer share and of the gas the staker pays to claim, and any net rate an operator quotes to a depositor is the operator’s own arithmetic on top of the posted one.

A stated slashing exposure. Isolated equivocation draws a small base slash (working figure: 1–5%) plus committee ejection, and the slash scales with the total stake equivocating in the same window, reaching a total burn as that fraction approaches the third needed to threaten finality. Because the scale grows as further proofs arrive, slashes sit in escrow until the window closes and execute at the final scale. A mass event, which looks like a network partition or a defective release, is the other reason to wait.

Which stake burns. Under delegation the signing key is the operator’s and the slashed capital is largely the depositors’. Which burns (the operator’s bond, the depositors’ pro-rata share, or both in some order) follows from the committee identity model. The FIP records it as unsettled, and more open items depend on it than on any other.

Unbonding terms. Withdrawal sits behind a delay set by two conditions: it must strictly exceed the fault-proof submission window so an equivocator cannot withdraw ahead of the slash, and it must be long enough that an exit is priced against the posted rate rather than traded intraday. Working figure: 90 days.

What dispersion means for an operator that grows

Activation requires distribution as well as depth, because a deep pool concentrated in a few hands is cheap to tip. The conditions are that the smallest coalition of distinct participants controlling a third of stake exceed N_min participants, and that no single participant hold more than max_share of the pool. Both are measured over delegated operators, the unit that actually signs, rather than over raw addresses, and the concentration analysis sets both constants.

For a successful operator that is a ceiling on the share of the pool it can hold while the network still activates. Lido reached roughly 23% of all staked ETH through the route this proposal opens (its own February-2026 tokenholder update). The conditions are necessary but not sufficient, since on-chain identities are cheap to multiply: they force a deliberate attacker to at least manufacture apparent distribution. Pool depth and correlation-scaled slashing are the economic backstop. Reversing activation after a concentration finding takes a FIP and a network upgrade.

Custodians and exchanges

Staked FIL leaves the holder’s account balance and lives in actor state, which changes what a custodian’s ledger has to carry.

A staked-and-accruing balance concept. The balance lives in actor state and grows only when a signature lands in a canonical certificate. Ledgers, statements and reconciliation have to carry a balance that is committed and accruing.

Pull-based claims. Rewards arrive only when someone claims them. Someone runs the claim and pays the gas, and the ledger has to account for both, against a posted rate that is gross of them.

The unbonding delay in withdrawal terms. A custodian offering FIL staking is offering an instrument a customer exits on the unbonding delay. That is a disclosure question.

The dispersion rules constrain participants, and nothing in them constrains jurisdictions. Delegated stake will tend to sit with exchanges and custodians, regulated companies a government can lean on. Jurisdictional concentration is a separate measure from the dispersion rules. The damage is bounded by the payoff: compelled custodians can at worst stall finality back to today’s network.

Wallets

A staker needs deposit, key registration, withdrawal and claim against the staking actor’s methods. All four are new wallet work, and a first-class staking client starts as a new repository. Lotus binds enrollment to a miner address today, so verbs that start and stop participation for a plain FIL holder are part of the same work. Key registration carries invariants the wallet has to enforce at deposit and registration time: BLS only, subgroup validation, a duplicate-key rule, the rotation delay.

Finality consumers

Exchanges, bridges, verifying contracts and snapshots carrying F3 state all track the certificate chain. The certificate format, its verification and the two-thirds quorum rule are unchanged, and the chain restarts: the committee begins a fresh instance chain under a new network name, with instance numbering starting over. That retires the old instance chain cleanly, so verifiers built against it reject stake-era certificates instead of mistaking them for authoritative.

Re-initialization. New bootstrap epoch, new initial power table, new network name, which changes the peer-to-peer topics and the datastore prefix and invalidates every published F3 snapshot, so snapshot producers are affected too.

Reading the authority mark. Certificates carry no authority bit today. An external verifier needs a way to tell a binding certificate from an advisory one, including the trivially-quorate first stake-era certificates, since the existing smart-contract verification pattern will accept an advisory certificate as final. The FIP answers with an append-only activation record in the staking actor: a certificate is binding exactly when the record shows an entry opened at or before its instance and not closed by it. A node reads the record directly. A verifier holding only certificates reads the latch state out of the committee’s supplemental-data commitment, where a committee attests to its own authority. An integrator that takes one-way activation at its word can instead pin a single epoch, the way an upgrade height is published. The commitment’s encoding is open.

Certificates read as binding (certificate head final within tens of seconds), advisory, or halted. Advisory and halted report finality on the 900-epoch rule at about 7.5 hours, and only binding and advisory pay signers. The floor under all three is today’s network, so the fallback an integrator engineers for is the one it runs on now.

One jump in reported finality at activation. A node reports as finalized the later of the Expected-Consensus finality and the latest certificate head. With advisory certificates suppressed beforehand, that tag jumps forward by up to about 900 epochs in one step when certificates bind, and the derived safe tag moves with it, so anything computing confirmations off those tags sees the discontinuity once. What an integrator gets in return is settlement in tens of seconds against today’s 7.5 hours. On the modeled scenarios certificates bind from years 2–4 and the pool holds the security level for 12–14 years. The schedule sunsets at year 30.

Dashboards and supply reporting

Circulating-supply classification is mandatory, in both lotus and forest, in the same release. The state-traversal circulating-supply RPC hard-errors on any unclassified actor, so the reward pool and the staking actor must both be added to the classification or the call fails outright. The pool’s whole unpaid balance is netted out of circulating supply, reported as a single balance.

Staked FIL stays inside circulating supply, and dashboards carry a liquid-float line. On the protocol ledger the locked figure for recent network versions is the power actor’s total locked only. Extending it to staked FIL would be another consensus-critical change, and it would push pledge down as stake accumulates. The funding analysis leaves staked FIL inside circulating supply for that reason and reports the staking lock as a liquid-float line, circulating supply minus staked FIL. Circulating supply itself differs by source, since the protocol figure nets out pledge: 903.13M FIL against 819.16M at the aggregators at the same snapshot (epoch 6,275,948, 2026-08-13).

Concentration monitoring is a standing off-chain task. The activation conditions call for an analysis of realized ownership and delegation. That analysis sets the dispersion constants and runs entirely off chain. The dispersion requirement is new for the committee: its Nakamoto coefficient at the one-third threshold reads 7, down from 84 in January 2023, with the top ten owners holding 38.5% of block rewards (Filecoin Data Portal series behind the public L1-health dashboard, 2026-07-13). Whether the staking actor keeps publishing distribution statistics after activation, as a monitored signal with no automatic consequence, is open, and the FIP argues it should.

The recruitment campaign

The threshold is crossed by persuading holders of tens of millions of FIL to stake. That work happens entirely off chain, and the FIP calls it arguably the larger endeavor.

The target. About 80M FIL staked and dispersed. At the modeled arrival rate of 5M FIL a month, that takes sixteen months of uninterrupted inflow. The scenarios assume 130M FIL of reachable idle capital, 14.6% of the aggregators’ circulating supply, bracketed from two ends: the low end is the roughly 32M FIL parked in GLIF, the revealed on-chain preference for yield (3.6% of circulating), the high end what a cross-chain staked share of 30–60% of circulating supply would imply, 267M to 534M FIL. Both the 130M and the 5M a month are modeling assumptions, placed between the two readings by judgment. The FIP calls uptake its largest open variable. Its first test is whether staked FIL reaches 80M within two years of the upgrade epoch. That is where the model puts it, and depth far short of that falsifies the case. Thin demand is also expensive: at 40M FIL of reachable capital the pool never activates, every certificate paid for over thirty years is advisory, and the bill is 97% of what full demand costs.

Who holds idle FIL. The measured base is holders already accepting yield on chain: roughly 40M FIL, about 5% of circulating supply, as of 2026-07-05 (45M including one failed protocol’s stranded funds; CoinGecko circulating 795M at that date). GLIF dominates it at about 32M FIL across 3,590 iFIL holders (total value locked, 2026-07-05). FIL-denominated deposits held flat through an 85% drawdown in dollar terms (1.8M FIL July 2023, 34M July 2024, a 35.6M peak July 2025, 31.9M at the 2026-07-05 reading), but the base plateaued in 2024 and is bleeding about 10% a year, GLIF itself is diversifying into non-Filecoin tokens, and the stFIL failure burned a cohort that never re-parked. About 10% is where on-chain FIL products bearing credit risk have attracted supply.

Recently exited storage providers are a second population, and their holding decays with time since exit: of 15 sampled top-50 miners that exited in 2025, 67% still held at least 1,000 FIL in provider-controlled actors, against 27% for 2024 exiters, 7% for 2022H2–2023 and approximately zero for the 2021 cohort (n=15 per cohort for the 2022H2–2025 windows, top-50 miners only, sampled 2026-07-05; the 2021 reading is a separate 99-miner top-100 sample). Roughly 2.2M FIL sits in or near provider control across 30 recent exiters, counting every untraced withdrawal as sold.

Attributing FIL balances to exchanges, custodians, ETFs and corporate treasuries is research the campaign needs. The largest single holding is 155M FIL, about 19% of the aggregators’ circulating supply, untagged and created November 2020, with its address history still to be traced.

The posted rate falls as stake arrives. It is a published function of pool depth and released balance, opening under a 12% ceiling (working figure) and settling at 4.0–5.0% while staked FIL holds the security level. At launch, on the FIP’s worked configuration: 12.00% at zero staked, 8.25% at 25M, 6.29% at 50M, 4.89% at 80M, 4.26% at 100M, 3.57% at 130M staked FIL. Early stake is paid the most, so the first 80M FIL is recruited at a declining price. Whether the includer share comes out of that rate or sits on top of it is open.

Sequencing against deployment. Staking opens at the network upgrade, when the staking actor, the submission path and slashing activate; no committee runs and finality behavior is unchanged. Staking is open but unpaid through that stage, since the reward starts with the first signature inside a certificate. The committee forms at a coordinated client release triggered by registered stake passing a formation floor set far below the activation threshold, after which certificates are produced, submitted on chain and paid for, with slashing live; signers earn through that stage even though certificates are advisory, because payment turns on signatures rather than on bindingness. Activation is data-triggered with no release of its own. Neither later stage carries a date, so recruitment plans key off the stake figure.

The year-3 review. It fires at year three if the activation conditions have never been met, and annually after, reporting what the pool did over the trailing year and carrying the amendment path. It keys on activation rather than on the pool balance, because the case it catches is a pool that recruits stake, pays for advisory certificates, and never reaches the level at which any of them binds. A schedule that has bracketed the market’s required return wrongly is amended there.

Decisions that product terms wait on

The FIP records these as open, and settling each one releases work somebody has to ship.

Block-producer software has to submit certificates, and storage-provider F3 participation software stops being eligible and needs a deprecation note as participation moves to stake.

Further reading: the FIP discussion, funding the finality reward: options and a recommendation, how the finality reward is paid for, storage-only attacks, two classes of stake, all supporting documents.