Repository navigation
FIP Draft: Stake-Weighted F3 Finality #1288
Replies: 8 comments 1 reply
|
Funding the finality reward: options and a recommendation. The FIP leaves the funding source open. Where the ~100M FIL of pool income could come from, what each option costs, who pays, and which one to adopt. |
|
How the finality reward is paid for. The posted rate, the release schedule, and what a staker and a storage provider each see, in plain language, with an FAQ. |
|
Analysis of storage-only attacks under Expected Consensus, QAP Fast Finality, and Stake Finality. What an attacker holding storage power alone can do under this FIP, compared with today (F3 off, Expected Consensus alone) and with F3 back on weighted by QAP as FIP-0086 designed it, and what each share of QAP costs over the next twelve years. |
|
Two classes of stake: an idea for exploration. Exploratory; nothing in the FIP depends on it. A locked class of stake pledged behind a provider's sectors would let providers borrow pledge from the finality pool at about 5% instead of the 15% they pay today, and cut the pool's outlay 12–24%. It costs security, and one constant, the cap on locked stake's share of finality weight, is unpriced. |
|
Beyond the chain: operationalizing stake-weighted finality. What has to exist outside the protocol for the mechanism to work and reach its threshold: the participation API, delegated staking operators, custodians and exchanges, wallets, finality consumers, dashboards and supply reporting, the recruitment campaign, and the open items that block product terms. |
|
Thank you @hannahhoward for starting this thread, for those who missed Core Devs 94, where this was presented, recording can be found here: https://youtu.be/it1ft03pI4A?si=LY77jG8EMIVyZJMe |
|
AI 跑的分析,供各位参考。(拱手) FIP-1288《Stake-Weighted F3 Finality》深度解读
0. 三分钟速览一句话这是一个用「质押的 FIL」替换「存储算力」来给快速最终性(F3)提供安全基础,并为此动用约 1 亿 FIL 资金池来发放参与奖励的草案——它只改最终性,不改出块权、不改 PoRep、不改存储市场。 它做什么 / 不做什么
最关键的 5 个数字
最关键的 3 个判断(详见 §6、§7)
|
| 规则 | 机制 | 确认时间 | 实际状态 |
|---|---|---|---|
| Expected Consensus(EC)回退 | 基于存储算力加权的分叉选择,900 epoch 内的分叉会被拒绝 | 约 900 epoch ≈ 7.5 小时 | ✅ 今天实际在用 |
| F3 快速最终性(FIP-0086) | GPBFT 委员会,⅔ 权力签名产出证书 | 数十秒 |
F3 在 2023 年作为 FIP-0086 上线,设计上由质量调整存储算力(QAP)加权的委员会运行。它的问题不在于技术实现,而在于经济学:
FIP-0086 刻意没有给参与者发奖励,理由是"签名很便宜,参与者自己会从更快的最终性中获益"。
这个假设在实践中站不住。委员会失去法定人数(quorum),F3 停止产出证书,网络退回到 900 epoch 规则。这个提案的全部动机就从这里来。
这里有一个对理解全局至关重要的推论:因为 F3 其实已经是"零成本就能干掉"的状态(不参与就行),所以网络今天的安全容忍度就是 EC 的 ~20% 算力,而不是 BFT 标准的 33%。
1.2 提议者的前提:存储基础在萎缩
FIP 明确把这个当作给定前提(taken as given),而不是需要论证的假设:
- Filecoin 没有原生质押机制——FIL 今天唯一"锁定"的场景是绑定到存储扇区的 pledge 抵押品
- 流动 FIL 没有任何协议原生收益
- 共识完全依赖存储算力,而存储算力在下降
提案的逻辑链因此是:
存储算力下降 → 共识安全基础变薄 → 需要一个新的、不随算力萎缩的安全基础
↓
可销毁的质押 FIL
↓
但没人有理由质押 → 需要协议原生收益 → 需要资金池
↓
需要约 1 亿 FIL → 需要选一个资金来源
作者的诚实之处:她明确说"如果你不接受这个前提,这个工具就是错的"——并给出了更便宜的替代方案(见 §4 维度 D4)。这是一个很少见的、主动给反对者递刀子的写法。
1.3 提案者的四个问题(我建议你沿用这个框架)
FIP 开头列了四个必须按顺序回答的问题,这本身就是很好的分析框架:
| # | 问题 | 如果答案是否 |
|---|---|---|
| 1 | 我接受这个前提吗?(存储基础在萎缩,FIL 需要持久的安全与持有激励) | 更便宜的替代方案存在(#1106 恢复 QAP 委员会法定人数、vote escrow 锁定),到此为止 |
| 2 | 质押加权的最终性委员会是正确的第一步吗? | 应该提出别的干预方式 |
| 3 | 机制站得住吗? | 应该指出哪里要收紧、哪个 strawman 常量该改 |
| 4 | 这个价格值不值?(不管钱从哪来,约 1 亿 FIL) | 直接反对整个提案 |
我的补充:这四个问题里,问题 1 和问题 4 才是真正的分歧点,问题 2、3 是技术细节。但对持币人和矿工来说,决定切身利益的其实是问题 4 的答案——也就是"谁买单",而这恰恰是 FIP 正文刻意留空、推给配套分析的那一块(见 §4 维度 D2)。
1.4 对前提的外部核实(2026-09-17)
结论先行:提案的核心前提基本成立,而且比它自己说的还要严重。但提案里几个关键数字无法独立核实,另有一个数字已经过时。
① 存储基础在萎缩 —— ✅ 成立,且幅度惊人
用 Filecoin Data Portal 的日频数据独立核实:
| 指标 | 峰值 | 2025-09-15 | 2026-09-15 | 12 个月变化 |
|---|---|---|---|---|
| QAP(质量调整算力) | 25.10 EiB(2024-02-16) | 23.60 EiB | 12.00 EiB | −49.2% |
| RBP(原始算力) | 16.90 EiB(2022-08-25) | 3.116 EiB | 1.337 EiB | −57.1% |
| 全网总 pledge | 176.12M FIL(2023-12-19) | 128.28M FIL | 64.03M FIL | −50.1% |
| 有算力的 SP 数量 | 4,100(2022-06-24) | 1,356 | 521 | −61.6% |
四年内存储提供者数量缩水 87%。 这不是"增长放缓",这是生态规模的量级收缩。
② F3 已停止出证书 —— ⚠️ 没有独立公开确认,但旁证非常强
这是本次核实中最需要谨慎处理的一条。
草案反复声称"委员会失去法定人数,F3 停止出证书"、"快速最终性已经离线"、"F3 已经因零成本的冷漠而停滞"。但我没有找到任何官方声明确认主网 F3 停机。具体查证:
go-f3的 README 至今仍写 "🚀 Live on Mainnet / successfully activated on Filecoin mainnet on April 29, 2025 on epoch 4920480"(已过时或名义上仍在运行)status.filecoin.io没有针对 F3 停摆的事故公告- Filecoin 官方 newsletter 没有报道
但旁证链条非常强,而且指向同一个结论:
| # | 时间 | Instance | 证据 |
|---|---|---|---|
| 1 | 2025-05-03→06(激活后第 4 天) | 6017 | lotus #13094 [SEV3] F3 Instance: 6017 not reaching a decision。"finalization has reverted to EC"。F3 observer 只收到 ~200 个节点的消息,而 quorum 约 1500 个节点 |
| 2 | 2025-07-11→14 | 183760 | lotus #13208。Aider 进程故障 + SP 未设 initial power table CID、重启后丢状态退出 |
| 3 | 2025-09-24→25(NV27 后) | 313457 | lotus #13359。归因于升级期参与率跌破 66.6% 阈值:"70% 参与 → 5% 未升级 → 66.5% < 66.6%" |
| 4 | 2026-01-22→23 | 446663 | lotus #13473。一组共享基础设施的 SP 出现基础设施级宕机,观测到 ~500 PiB 掉电。事发时参与率"hovering near the protocol's minimum threshold (~66%)" |
| 5 | 2026-03-07 起,至今 open | 465951→466454 | lotus #13544。"F3 is not finalizing. … Commit3 participation is only 63%"。2026-03-10 更新:"we were at 68% now we are at 66.8%" |
根因已经很清楚:
参与率长期贴着 66.6% 的法定人数阈值
↓
任何扰动(一次共享基础设施宕机、一次网络升级)都能把它压到阈值以下
↓
而 GPBFT 的指数退避没有上限(go-f3 #1056 原文:
"We've been impacted by the lack of a cap multiple times in the
last 6-7 months since F3 has been activated. There isn't a good
reason to exponential backoff for days.")
↓
一旦卡住,要恢复得等数小时甚至数天
关键语境:F3 上线前,passive testing 阶段报告的参与率是"约 77%–79% 的网络算力",看起来相当安全。核心开发者当时没有预见到"不给奖励 → 参与率会慢慢滑到阈值边缘"这条路径。 FIP-0086 原文甚至明确预见到了"参与不足 → F3 liveness 受损 → 掉回 900 epochs"这个后果,但判断"成本增量不足以改变节点运营成本结构",因此决定不给激励。
这是一个"预言了自己的失败但低估了概率"的案例。
deltaBackOffMax)、FIPs PR #1226(2026-01-16 开,至今 open)——即为了修这个问题去改 FIP-0086 的 PR。FIP-0086 的规范文本本身也还在被回填修补。
对你的判断意味着什么:草案的前提(F3 事实上不可用)是可信的——有 5 次主网停摆的 SEV3 记录、有公开承认的根因、有至今未关闭的修复 issue。但"永久停机"和"反复停摆后长期处于停滞"是两种不同的状态,草案的措辞倾向于前者。准确的表述是:F3 反复停摆,最近一次从 2026-03-07 持续至今未恢复。
③ FIL 价格与流通量 —— ✅ 准确(但已过时)
| 草案说法 | 核实结果 |
|---|---|
| FIL 价格 $0.77 | 草案基准日 2026-08-23 日均 $0.7609 ✅ 准确 |
| CoinGecko 流通量 795M(2026-07-05) | 实测隐含聚合方供应 795.32M ✅ 完全吻合 |
| 协议口径 903.13M vs 聚合方 819.16M(2026-08-13) | 实测 903.37M vs 819.18M,epoch 日期亦吻合 ✅ |
| 缺口稳定在 84–89M FIL(约供应的 9-10%),最新(9/15)为 912.76M vs 827.81M |
④ 无法独立核实的两项(都影响安全论证)
| 草案说法 | 核实结果 |
|---|---|
| Nakamoto 系数:⅓ 阈值处为 7,2023-01 为 84 | ❌ 未查到任何独立来源。Data Portal 的 62 个字段中无集中度指标;生态仪表盘与 Starboard 均无此指标;学术论文只有定性结论 |
| 前十大所有者获得 38.5% 区块奖励 | ❌ 未查到任何独立来源 |
方向性判断可以做:SP 数量从 1,356 跌到 521,且 QAP 高度集中(GLIF 一家绑定的 pledged miners 就有 212–219 个、对应 QAP 约占全网 18–28%)。在 521 个有算力的 actor 中达到 ⅓ 阈值需要的数量级与"7"不矛盾。但"84(2023-01)"这个历史值在直觉上偏低——2023 年 1 月有算力的 SP 数以千计。建议标注为待核实。
⑤ 与草案存在出入的三处引用(技术层面)
| 草案/常见转述 | 官方原文 | 影响 |
|---|---|---|
| 「FIP-0118 是区块奖励时间表」,15.2/9.5/6.8/6.0/5.4/4.8M FIL 是它的数字 | ❌ FIP-0118 全称是「(Solstice) Deprecate Fil+ and Fund Services via Block-Reward Split」,通篇没有任何年度释放表。那 6 个数字是 #1288 自己用 FIP-0118 的权重机制推算的(草案表格标题写的 "FIP-0118 as passed",但它是推算值不是发布值) | 中等:不影响结论方向,但引用时要注明出处 |
| 「FIP-0081 规定了 9% 的流通供给上限」 | ❌ FIP-0081 原文说的是 30%(PledgeLockTarget = 0.3 × CirculatingSupply)。9% = 30% × 30% 的推导,在配套分析里写出,但不在 FIP 原文中 |
中等:需准确表述为"下限为穿越前总量的 30%,其渐近结果约为流通供给的 9%" |
| 「FIP-0081 设了 32 FIL/QAP-TiB 的上限」 | ❌ 该上限真实存在,但来自 miner actor 里 space-race 时代就有的常量(InitialPledgeMaxPerByte),FIP-0081 未修改也未提及它 |
轻微 |
另外一处需要标注的技术缺口:
草案说"所需的 BLS12-381 运算已在 FIP-0105 下提供,不需要新 syscall"。
但 FIP-0105 的标题与定位都是「FEVM Precompiles」,其原文只声明"在 FEVM 中引入预编译函数"。
而草案 §3a 假设的是在内建 actor(reward-pool actor)内直接验证证书。
FIP-0105 原文并未声明 built-in actor 可直接调用这些运算。 考虑到草案自己就把"每证书 gas 未测量"列为早期交付物,并准备了客户端验证的退路,这个衔接点值得标为待验证。
1.5 提案的真实成熟度:⚠️ 仍在 Ideation,未进入 nv29
这一条在 FIP 正文里看不到,但对判断"这件事会不会发生"至关重要。
2026-08-26 的 Core Devs 94 会议上,作者正式presented了这个提案。官方会议纪要的决议是:
"Stake-Weighted F3 Finality remains in Ideation."(仍处于构想阶段)
"It is not scoped for NV29."(未纳入 nv29 范围)
没有正式决定将其推进到 Draft,也没有绑定到任何网络升级。
也就是说:截至 2026-08-26,这个提案还没有 FIP 编号,没有进入 Draft,被明确排除在即将到来的 nv29 升级之外。
会议上与会者提出的关键质疑(这是目前唯一的实质性外部反馈):
| 谁 | 质疑 |
|---|---|
| Andy Jackson | 轻量级 F3 参与软件已经存在,只需要 < 200 MB RAM 和 < 1 GB 磁盘空间 → 这实际上在暗示:参与率低不是硬件门槛问题 |
| Jennifer Wang | "约 80M FIL 的参与目标"是否可达?如果永远达不到、或之后跌破阈值怎么办?需要多高收益?参与是否会被少数运营商垄断? |
| Jennifer Wang(追问锁定期) | 警告快速进出无法提供预期安全收益。Hannah 确认"将要求最低锁定期",且应长于 EC 最终性时间,但具体时长尚未参数化 |
| Jennifer Wang | 提到 6 个月锁定期在之前已被部分潜在参与者认为过长 |
| Jennifer Wang | 建议基于与大型 FIL 持有者和 SP 的实际沟通做更多建模——他们最关心的是预期收益、最低锁定时长、流动性、参与要求、以及达到目标的概率 |
| Andy Jackson | 建议阶梯式接近目标以缓解冷启动 |
资金:Hannah 在会上表示倾向于用挖矿储备作为资金来源——这与配套分析的推荐一致。
另一个值得注意的观察:该 Discussion 帖发布一周后,除了作者自己和一位 collaborator,没有任何实质性第三方讨论。6 条回复里 5 条是作者自己贴配套分析链接,1 条是致谢。没有任何核心开发者提出技术质疑或表示支持。 反应数只有 1–2 个。
这改变了你怎么读这份材料:它是一份质量极高但尚在早期、且连"要不要做"都还没达成共识的构想。FIP 正文里那种"六个部件 + 11 个开放项 + 三阶段部署"的完备感,会让人高估它的成熟度。 配套分析做了约 54,000 种配置的穷举——工程严谨度很高,但商业和政治可行性完全未验证。
2. 机制拆解
2.1 总览
flowchart TB
subgraph 资金来源
R["挖矿储备金<br/>282.93M FIL(推荐来源)<br/>一次性拨 91M"]
S["FIP-0118 区块奖励切分<br/>(备选,会削减 SP 收入)"]
end
R --> RP["奖励池 Actor<br/>unreleased / released 两本账<br/>无 controller"]
S -.-> RP
RP -->|"固定释放表<br/>不读链上状态"| REL["released 余额 P"]
subgraph 质押侧
H["FIL 持有者"] -->|"存入 + 注册 BLS 密钥"| SA["Staking Actor<br/>§1"]
D["委托者"] -->|"委托给运营方"| SA
SA -->|"Stake_i / TotalStake<br/>10 实例 lookback"| PT["GPBFT 权力表<br/>§2"]
end
PT --> C["F3 委员会<br/>⅔ 签名"]
C -->|"证书"| SUB["链上提交<br/>§3a"]
SUB -->|"验证签名"| RP
RP -->|"signer share + includer share α≈10%"| PAY["奖励发放"]
C -.->|"双签"| SLASH["故障证明<br/>§4 相关性罚没"]
SLASH -.->|"销毁质押"| SA
SA --> LATCH["激活闸门<br/>§6 深度 + 分散度<br/>sticky latch"]
LATCH -->|"打开后"| BIND["证书 binding<br/>数十秒最终性"]
LATCH -->|"未打开"| ADV["证书 advisory<br/>仍按 900 epoch 上报"]
2.2 六个部件逐个说
§1 质押 Actor(新增)
链上质押的唯一真相来源。记录每个参与者的余额 + BLS 签名密钥,支持存款、延迟提款、委托,并维护激活状态。
关键设计:解绑延迟有两个下界
- 必须严格长于故障证明窗口——否则作恶者可以先提款再被罚
- 必须在经济上有意义——退出要按公布利率定价,而不是被当作日内交易
strawman 值:90 天。它是延迟,不是固定期限(at-will stake),所有模型都假设随时可取。
这个部件的影响远超技术层面:90 天的解绑期意味着你质押的 FIL 不是"随时能拿回来"的,这对钱包、交易所、托管方的产品条款都是硬约束。
§2 权力表换源(唯一"平凡"的改动)
这是整个提案技术上最简单的部分——只是换了个输入:
FIP-0086 原公式: ParticipantPower ← truncate(0xffff * MinerQAP / TotalQAP)
本提案公式: ParticipantPower ← truncate(0xffff * Stake_i / TotalStake)
其余全部不变:同样的 10 实例 lookback(CommitteeLookback)、同样的 16 位归一化、同样的证书机制。
但资格条件换了一整套:
- 旧的(矿工侧):10 TiB 算力下限、零 fee debt、无未结共识故障
- 新的(质押侧):最低质押额 + staking actor 自身的故障状态
两个编码限制决定了最低质押额:
| 限制 | 后果 |
|---|---|
| 低于总量的 1/65535 会被归一化为 0 权重 | 在 100M FIL 的池子里约 1,526 FIL |
| CBOR 把权力表和每个证书 delta 上限设为 8,192 条 | 最低质押额必须让委员会规模落在限制内 |
⚠️ 一个具体的静默失败模式:质押量低于 1/65535 的参与者会被缩放为 0 权力,但仍以"正权力条目"出现在名单里——本人无从得知自己静默地签不了名、赚不到钱。这是配套文档明确点出的 UX 陷阱。
奖励在权力表余额之外累积,避免每 epoch 复利导致每个参与者都进入每个证书 delta。
§3 证书门控奖励(提案的经济核心)
规则一句话:质押的 FIL 只有当它的签名出现在一个有效的、规范的最终性证书里、且证书新最终化了某些 epoch 时才赚钱。
由此推出的性质(这些设计很讲究):
- 重复提交已最终化的 head → 赚不到钱
- Epoch 每 30 秒推进一次,与委员会行为无关 → 加快出证书不能提高支付
- 拖延、闲置质押、F3 停摆 → 赚不到钱
- 一个签名者的奖励不会减少另一个人的 → 审查同伴没有收益,拒绝提交则损失自己的 α
奖励怎么分:
- signer share —— 给签名者
- includer share
α(约 10%) —— 给链上提交者。实践中是区块生产者,因为只有收集到 α 的生产者有动机包含提交。
公布利率公式:
r(S, P) = spend_rate × P / (S + S_offset)
S = 总质押 FIL
P = 奖励池的 released 余额(只有 released 可花费)
S_offset = 让利率在低质押时保持有限,并控制下降速度
worked configuration:spend_rate = 0.22,S_offset = 55M,启动利率上限 12%(strawman)。
| 质押总量 | 启动时公布 APR | 第 6 年公布 APR(池子 34.15M FIL) |
|---|---|---|
| 0 | 12.00% | 13.66% |
| 25M | 8.25% | 9.39% |
| 50M | 6.29% | 7.16% |
| 80M | 4.89% | 5.57% |
| 100M | 4.26% | 4.85% |
| 130M | 3.57% | 4.06% |
| 200M | 2.59% | 2.95% |
验算:启动时
0.22 × 30M / (0 + 55M) = 12.00%——精确落在天花板上,上方没有任何余量。
两个进一步的调度性质:
- 池子缩放:利率随 released 余额移动——发钱降低它,计划释放提高它。未花的钱会留存,所以慢启动不会饿死后来的质押者;如果释放快于支出,利率会上升直到质押响应。
- 年支付在任何深度都低于
spend_rate × P。worked 情景下第 1 年是 6.60M FIL(对应 30M released 池)。不支付任何 epoch 都没有补发。
具体算例(来自 FIP 正文):
假设你在签名开始的第 10 个月质押 10,000 FIL,全网质押 50M FIL → 利率 6.29% → 一年约 630 FIL。
当全网质押到达 100M(模型路径的第 6 年,34.2M FIL 已释放未支付)→ 同样 10,000 FIL 赚 4.85%,约 485 FIL/年。注意:这些是毛值,未扣 includer share α 和领取 gas。领取是 pull-based(需要你自己发起)。
§3a 证书上链提交与验证(最大的新协议面)
今天 F3 证书完全在链下——存在节点本地存储和 P2P 协议里。§3 要求链能接收并验证它们,这是提案中最大的新协议组件。
流程:
- 区块生产者把每个新证书作为消息提交给奖励池 actor
- actor 维护 per-instance watermark
- 第一个有效提交给签名者记账并支付 α
- 重复提交会廉价地失败
- 同一 instance 的冲突证书被接受为证据(但不给奖励)→ 为 §4 提供链上证据
- 提交是无许可的——任何账户都能提交。要压制一个有效证书,需要每个生产者都放弃自己的 α,违背自身经济利益
技术可行性:聚合验证需要一次系数加权的多标量乘法 + 一次 pairing。所需的 BLS12-381 运算已在 FIP-0105 下提供,不需要新 syscall。
三个未定价的工程风险:
| 风险 | 状态 |
|---|---|
| 每证书 gas 成本 | 未测量,是早期实现交付物。若过高 → 退路是客户端验证 + 系统证明摘要,但那会给每个客户端加一条区块验证规则,并依赖 FIP-0107 |
| 64 KiB 消息上限 vs 权力表 delta | 证书格式允许更大的 delta,提交必须分块,或让权力表按有界节奏读质押快照 |
| 历史委员会快照的保留窗口 | 必须覆盖故障证明提交窗口 |
§4 相关性罚没(安全性的真正来源)
这是与存储算力最本质的区别:
攻击者获得的存储算力可以无损转移,其 pledge 是可回收的抵押品(可证明的双分叉挖矿罚款仅约 4 FIL);
而作恶的质押 FIL 会被销毁,且销毁比例随作恶规模上升。
两种举证路径:
| 路径 | 机制 |
|---|---|
| 单点双签 | 需要新的故障证明子系统:节点保留并 gossip 冲突消息对,任何人可提交,actor 重建 FIP-0086 payload 并验证两个签名 |
| 相关性作恶(更简单) | 同一 instance 的两份有效证书本身就是证据——出现在两个 bitset 交集里的签名者都双签了,而这些证书已经在链上(§3a) |
罚没曲线的形状(这是设计精髓):
孤立作恶(可能是运维事故) → 基础罚没 1-5% + 委员会除名
↓ 罚没随同一窗口内作恶的质押比例上升
作恶比例逼近威胁最终性的 ⅓ → 约 100% 全额销毁
三个关键机制细节:
- 除名在 lookback 生效(故障记录后 10 个实例),因为在途实例的权力表已经固定
- 金额先进入 escrow 托管,等窗口关闭后按最终规模执行——因为最终规模取决于可能稍后到达的更多证明
- 报告者获得奖励(上限为销毁额),剩余部分销毁。为什么不还给池子?因为还给池子会提高
P和公布利率,等于让攻击补贴其他质押者
为什么"孤立作恶只罚 1-5%"是对的:作者的理由是"运维事故必须能被承受,否则惩罚会同时吓阻攻击和参与"。
为什么 escrow 特别重要:大规模事件里,"⅓ 委员会作恶"看上去既像攻击,也像网络分区或缺陷版本发布。
必须成立的硬约束:
- 罚没从第一张证书起就生效(早于激活)
- 解绑延迟必须严格长于故障证明提交窗口 ← 否则整个威慑不成立
- 攻击成功的代价 ≈ 攻击质押本身(约池子的 ⅓,激活时是数千万 FIL)。作者明确说:这个数额必须超过快速最终性所保护的双花价值,并建议随结算规模增长重新检查激活带
⚠️ 未决且影响最大的一项:委托模式下,签名密钥属于运营方,而被罚没的资本主要属于委托者。到底烧谁的——运营方保证金、委托者按比例分摊、还是两者按顺序——取决于"委员会身份模型",而 FIP 把它记录为未决。这意味着在定案之前,委托质押者根本无法评估自己的真实风险。
§5 资金池的四个要求
无论资金来源是什么,机制要求它满足四条:
| 要求 | 内容 |
|---|---|
| 无 controller 的奖励池 actor | 内置,两本账:unreleased / released。只有 released 可花费,且 r(S,P) 只读 released |
| 固定释放表 | 余额按预定表从 unreleased 移到 released,不读任何链上状态——这让公布利率保持为可观测值的公开函数 |
| 供应量核算处理 | 池子的全部未付余额都要从流通量中排除(不是只排除某个来源的子账本)。两个新 singleton(奖励池 + staking actor)都必须加入流通量遍历,否则 StateCirculatingSupply 会报错 |
| 复核与日落钩子 | 第 25 年资金复核 → 第 30 年日落(停止支付、处置余额);第 3 年未激活复核(若从未激活,之后每年重复)——覆盖"付了建议性证书却从未 binding"的池子,并提供修正路径 |
§6 激活闸门(最有意思也最容易读漏的部分)
问题:升级时质押量为零。F3 需要非空的初始权力表,而一个几乎空的委员会比没有委员会更糟——单独一个参与者就是 ⅔ 法定人数。
解决方案是三重保险:
① 全新实例链,不是旧链交接
质押加权 F3 以全新的实例链启动:新的 F3 网络名、新的 bootstrap epoch、固定的初始权力表。
为什么不在旧链内交接?因为那需要旧 QAP 加权委员会的 ⅔ 签名——而这个法定人数正是本提案要替换掉的东西。
新链还有一个干净的好处:旧时代的证书流和新时代的证书流被彻底隔离——针对旧流构建的验证器、桥、快照会拒绝新时代证书,而不是误认为它们有权威性。
② 深度门槛
total_staked ≥ stake_activation(strawman:80M FIL)
在这之前,网络继续使用今天的 900 epoch 规则,质押加权证书只是建议性的(advisory)。
这避免了一个叫 "thin-but-trusted" 的失败模式:一个廉价就能获得的法定人数变成权威。
③ 分散度要求(容易被忽略但很重要)
深度不够——如果深池握在少数人手里,推翻它的成本很低。
两个 strawman 条件:
- 控制 ⅓ 质押的最小不同参与者联盟必须包含超过
N_min个参与者 - 任何单一参与者不得控制超过
max_share的池子
两者都以委托"运营方"(真正签名的实体)度量,而不是以原始地址度量。
作者对这两个条件的诚实评价(值得赞赏):
这些测试是必要但不充分的,因为链上身份可以廉价地增殖。它们能防止意外的集中被激活,并迫使蓄意攻击者至少制造出分散的表象。真正的经济底线仍然是池子深度 + 相关性罚没。
④ Sticky Latch:故意不对称的开关
打开条件: total_staked ≥ stake_activation AND 分散度条件
→ 连续保持 W(strawman:3 个月)才打开
(单个深度足够的 epoch 无法激活最终性)
关闭条件: total_staked < stake_deactivation(strawman:50M FIL)
→ 连续保持 W′(strawman:12 个月)才关闭
关键:激活之后,分散度失败永不关闭 latch
为什么"激活后分散度失败不关闭"? 否则攻击者可以通过制造集中来获得一个关闭开关——比 stall 更便宜、且不需要 ⅓ 质押或作恶。激活后的集中由罚没机制处理,必要时留给后续 FIP。
⑤ 权威规则
实例 i 的证书是 binding,当且仅当 ActivationRecord 中存在
一条在 i 之前或同时打开、且未被 i 关闭的激活条目。
ActivationRecord 是 staking actor 中的只追加列表,记录 (opened_at_epoch, opened_at_instance, closed_at_epoch|null, closed_at_instance|null)。
2.3 三阶段部署
| 阶段 | 触发 | 状态 | 学到什么 |
|---|---|---|---|
| 1. Actor 部署 | 网络升级 | 质押开放,没有委员会运行,最终性行为不变 | 在真实链上真实余额上演练迁移和资金逻辑 |
| 2. 委员会组建 + 建议性累积 | 注册质押超过组建下限(远低于激活阈值)后的协调客户端发布 | 证书产生、上链提交、有报酬;罚没已上线;证书仍是建议性的 | 在真实经济下运行所有没有生产先例的子系统;在这里测量每证书 gas |
| 3. 激活 | 数据触发,无独立发布 | 深度 + 分散度条件持续满足 W 后,证书变成 binding | — |
每一阶段只依赖前一阶段已演练过的组件。 部署风险与委员会运营风险分离,委员会运营风险与 binding 最终性风险分离。后两阶段没有固定日期——这是一个顺序,不是日历。
从第一张质押时代证书起,质押者就开始赚取参与奖励——奖励累积在激活之前就开始了。这是有意为之:必须给累积阶段付钱,才能招募到质押跨越激活门槛。
2.4 三种状态(Appendix B)
在所有三种状态下,区块生产和链排序完全相同——这是理解这个提案的关键:
| 状态 | 什么为真 | 客户端上报的最终性 | 奖励 |
|---|---|---|---|
| Binding | Latch 打开,法定人数活跃 | F3 证书 head,数十秒内 | ✅ 按证书支付 |
| Advisory | Latch 关闭(激活前,或去激活后),法定人数活跃 | 900 epoch 规则,约 7.5 小时 | ✅ 按证书支付(§3 门控的是签名,不是 bindingness) |
| Halted | 无法定人数 | 900 epoch 规则,约 7.5 小时 | ❌ 无,§3 的支付自动停止 |
注意 Advisory 也给钱——这是一个容易被误读的点。付费条件是你的签名进入规范证书,而不是证书是否有约束力。
三者的地板都是今天的网络——三重状态里最差的那个也等于现状,所以不会比现在更糟。
3. 关键数字总表(速查)
3.1 参数与常量
| 类别 | 项目 | 值 | 状态 |
|---|---|---|---|
| 质押 | 最低有效质押(编码限制) | 总量的 1/65535(100M 池中约 1,526 FIL) | 计算得出 |
| 权力表条目上限 | 8,192(CBOR) | 固定 | |
| 权力表 lookback | 10 个实例 | 沿用 FIP-0086 | |
| 解绑延迟 | 90 天 | 🟡 strawman | |
| 奖励 | spend_rate |
0.22 | 🟡 strawman |
S_offset |
55M FIL | 🟡 strawman | |
| 启动利率上限 | 12% | 🟡 strawman | |
includer share α |
约 10% | 🟡 strawman | |
| 奖励公式 | r(S,P) = spend_rate × P / (S + S_offset) |
设计确定 | |
| 罚没 | 基础罚没(孤立作恶) | 1–5% + 委员会除名 | 🟡 strawman |
| 相关性罚没上限 | 作恶比例逼近 ⅓ 时 ≈100% 销毁 | 设计确定 | |
| 解绑延迟 vs 故障证明窗口 | 解绑必须严格更长 | 硬约束 | |
| 激活 | stake_activation |
80M FIL | 🟡 提议值,最终值开放 |
N_min / max_share |
未定价 | 🔴 开放 | |
stake_deactivation |
50M FIL | 🟡 strawman | |
打开维持窗口 W |
3 个月 | 🟡 strawman | |
关闭维持窗口 W′ |
12 个月 | 🟡 strawman | |
| 资金 | 30 年总需求 | 90M–97M FIL(100M 为 strawman 口径) | 建模 |
| 推荐方案包络 | 91M FIL(挖矿储备金的 32%) | 🟡 推荐 | |
| 最便宜再平衡方案 | 48M FIL(35%/25%) | 🟡 备选 | |
| 释放表 30 年合计 | 18.4392M FIL + 升级时 30M = 48.4392M | worked 情景 | |
| 日落销毁 | 14.362M FIL(worked 情景) | worked 情景 | |
| 时间 | 第 3 年未激活复核 | 若从未激活则每年重复 | 设计确定 |
| 第 25 年资金复核 | 默认动作是"无结论则停止支付" | 设计确定 | |
| 第 30 年日落 | 停止支付、处置余额 | 设计确定 |
3.2 安全与攻击成本
| 项目 | 数值 |
|---|---|
| 今天 EC 的容忍度 | ~20% 算力(来源:Wang, Azouvi, Vukolić, AFT 2023) |
| BFT 最终性的标准容忍度 | ~33% 质押 |
| 激活时 ⅓ 质押背后的可销毁资本 | 26.7M FIL |
| 激活时 ⅓ 算力背后的可回收 pledge | 21.5M FIL |
| 获取否决权(⅓ 池子)的成本 | 30–43M FIL(高于 ⅓,因为攻击者自身质押会推高分母) |
| 持有否决权的成本(放弃质押收益) | 约 1.42M FIL/年 ≈ 0.08% 仓位/周 |
| 当前委员会 Nakamoto 系数(⅓ 阈值) | 7(2023 年 1 月为 84) |
| 前十大所有者区块奖励份额 | 38.5% |
| Lido 参照(ETH 质押占比) | 约 23% |
3.3 逐年对比(FIP 正文表)
⚠️ 出处说明:下表的"SP 共识奖励"一列,草案标注为 "FIP-0118 as passed",但它不是 FIP-0118 发布的数字——FIP-0118 原文("Solstice: Deprecate Fil+ and Fund Services via Block-Reward Split")通篇没有任何年度释放表。这 6 个数字是本草案用 FIP-0118 的权重机制推算出来的(consensus stream = w1 × 发行量)。引用时请注明。
| 距最终性启动年数 | 1 | 2 | 3 | 4 | 5 | 6 |
|---|---|---|---|---|---|---|
| ⅓ 质押 FIL(M) | 20.0 | 26.7 | 28.3 | 31.7 | 31.7 | 33.3 |
| SP 那年共识奖励(FIP-0118 已通过版,M) | 15.2 | 9.5 | 6.8 | 6.0 | 5.4 | 4.8 |
读法:两组数字在签名第一年就交叉,之后差距扩大。作者用这个论证"到证书真正 binding 时,可罚没预算已经大于模型里的战利品"。但要注意表里 SP 奖励在快速下降——这个对比的说服力相当程度上来自分母的萎缩。
4. 资金从哪来 —— "谁买单"的完整拆解
这是全文最重要的一节。 因为同一个 FIP,换一个资金来源,对矿工可以从"零成本"变成"直接减收 30-65%",而 FIP 正文刻意把这一块留空,只在配套分析里回答。
4.1 首先要记住的一件事:不新增铸币
所有资金方案都不改变每 epoch 铸币量,不动 simple minting、baseline minting、baseline 目标、以及 2 亿总供应上限。这是提案一个重要的道德优势——它不是通过通胀来给质押者发钱。
但"不新增铸币"不等于"没有成本"。成本是从别的地方挪过来的,只是挪的是谁的钱,取决于选项。
4.2 两个候选来源
| 来源 | 内容 | 关键事实 |
|---|---|---|
| ① 挖矿储备金(必需项) | 创世分配的 300M FIL,用途是"未来尚未确定的新矿工类型",今天约 282.93M FIL,从未被提取过。提取路径走 FIP-0085:keyless account,需要 FIP + 升级 | 推荐用它单扛。包络 48M–91M = 储备金的 17%–32%,一次性划转 |
| ② FIP-0118 区块奖励切分中的 finality stream(可选项) | 在 Solstice 的区块奖励切分里插一条流,service 份额按"stream 不存在"计算且永不因它被削减,stream 只吃 residue | 需要两处修订:W2_CAP 从 0.50 → 0.35/0.40;SP 份额继续下探到 50% floor 之下。成本直接落在存储提供者身上 |
两者几乎精确地一对一互相替代。 stream 在第一个十年交付的每一 FIL,都是储备金不必释放的一 FIL。
4.3 选项集与成本(关键表)
统一口径:从第 2 年起、以 5% 要求回报连续质押 12 年 80M FIL。
| Service cap | SP floor | 储备金包络 | 第一个十年高于今天既定曲线的量 | 谁受损 | 30 年 SP 共识奖励 |
|---|---|---|---|---|---|
| 50%(不变)——推荐 | 50%(不变) | 91M FIL | +38.8M FIL | 无人(只花储备金) | 84.2M(不变) |
| 40% | 35% | 67M | +21.1M | SP | — |
| 40% | 25% | 53M | +15.5M | SP | — |
| 35% | 30% | 55M | +13.5M | SP | — |
| 35% | 25% | 48M FIL(最便宜) | +10.9M FIL | SP(最重) | 68.0M(截 20%) |
关键取舍的精确形状:
储备金用得越多 → 第一个十年多发的 FIL 越多(+10.9M → +38.8M)
→ 但区块奖励完全不动,SP 不受影响
区块奖励切得越深 → 储备金花得越少(48M vs 91M)
→ 但 SP 的每 EiB 共识收入被压到:
35% floor → Solstice 已通过版的 70%
30% floor → 60%
25% floor → 50%
4.4 作者为什么推荐"只用储备金"
作者给的三个理由:
- 这笔储备金创世时就是为"尚未确定的新矿工类型"预留的,从未动用——而逐证书的最终性委员会确实是一项新的、已验证的服务。用途匹配。
- 再平衡方案会从升级 epoch 起就制造 SP vs 质押者的对立阵营,并加深共识切分。更关键的是:费用义务不会随收入同步下降,于是"终止扇区"变成理性选择,从而移除质量调整算力——作者明确承认这个循环没有建模。
- 储备金收入在升级时就固定,不依赖 FIP-0118 是否激活、不依赖交易量门槛、不依赖发行衰减。
代价也很直白:这是选项集里花钱最多的一条——第一个十年 +38.8M FIL,是"最便宜选项 +10.9M"的三倍多;全网共识总支出 148.9M vs 123.3M。
4.5 一个必须知道的会计细节:净额处理
不做一项修正的话,"发行量逐年下降"这个卖点会直接破掉:
- 储备金 FIL 已经铸出、在发放前不在流通中
- 今天的客户端会把"发放"当作储备金余额下降处理,等于把整笔划转立刻计入流通
- 修正案:从流通量中减去资金池的全部未支付余额,并按网络版本门控(做法同 FIP-0100)
为什么要"全部"而不是只减一个来源的子账本?否则每笔 stream 入账都会重复计算。
还有一个共识层面的细节:FilCirculating 是共识输入——miner actor 把初始 pledge 绑定到它的 30%(LOCK_TARGET_FACTOR_NUM/DENOM = 3/10)。所以:
质押的 FIL 仍然计入流通量。把它减掉会降低每扇区 pledge,把最终性质押市场和存储抵押市场耦合在一起。质押锁定改为作为流动浮增(liquid-float)报告行呈现 = 流通量 − 质押量。
对照:以太坊 34.9% 质押、排除量为零;ETH、SOL、ADA、ATOM 都只把它当报告项。
4.6 这笔钱对矿工到底意味着什么
如果采纳推荐方案(91M 纯储备金):
- ✅ SP 在九季度斜坡期内(到 2029-03)收入完全不变,斜坡之后也不变
- ✅ SP 作为 FIL 持有者同样有资格领取参与奖励(统一利率,不触碰扇区、pledge 或区块生产)
- ✅ SP 的区块生产者额外赚取 includer share(提交链上证书)
- ✅ 模型显示 SP 的 per-EiB 收入在所有选项下都逐年下降(因为区块奖励本身在衰减),但这个 FIP 推荐方案不额外加一笔下降
如果采纳再平衡方案:
- ❌ 35% floor:30 年 SP 共识奖励 56.05M / 67.99M(两处口径),vs Solstice 已通过版的 72.32M / 84.20M
- ❌ 稳态 per-EiB 收入降到 Solstice 的 70%(35% floor)/ 60%(30%)/ 50%(25%)
⚠️ 作者警告:更深的削减把边际扇区推向终止,从而移除网络安全所依赖的存储算力——这个循环未建模,且作者说"在任何更深的削减被提出之前,建模是欠着的"
给你的操作性结论:在看这个 FIP 的时候,第一个要问的问题不是"机制好不好",而是"最终采用哪个资金来源"。 如果采用推荐方案,矿工几乎不用反对;如果转向再平衡,矿工的反对理由非常充分。
5. 我建议的分析维度(怎么读才读得懂)
你问"从哪些方面分析能帮助更好理解"。以下是我认为最有用的七个切入角度,按重要性排序。前三个缺一不可,否则很容易被提案的叙事带走。
维度 1:把「安全机制置换」和「代币需求工程」拆开看
这是读这个 FIP 最重要的一件事。 提案实际上捆绑了两件性质完全不同的事:
| A. 安全机制置换 | B. 代币需求工程 | |
|---|---|---|
| 做什么 | F3 委员会权重从 QAP 换成质押 FIL | 用协议原生收益创造质押需求 |
| 论证方式 | 技术性、可验证——"可销毁的资本比可回收的抵押品是更强的威慑" | 经济学、高度不确定——"市场需要多少回报?会有多少人来质押?" |
| 可证伪性 | 高(形式化分析、攻击成本可计算) | 低(依赖 uptake,作者自己说是"最大开放变量") |
| 成熟度 | 相当扎实 | 薄弱 |
为什么要拆开? 因为 A 的说服力会"传染"给 B。当你读到"攻击成本 30–43M FIL"这种精确论证时,很容易顺势相信"所以这个提案是值得的"——但攻击成本只论证了 A,完全没有论证 B(值不值 1 亿 FIL)。B 的部分作者自己给出的证据其实是偏悲观的(见维度 6)。
具体怎么用:读的时候不断问自己——"这段话在论证哪一个?如果只采纳 A 不采纳 B 会怎样?"(答案:A 单独存在是可能的——即纯用罚没机制、不发奖励,但那会回到 F3 没有法定人数的老问题。所以 B 是 A 的必要条件,这个捆绑不是任意的,但它的成本效益需要单独评估。)
维度 2:「谁买单」——决定所有人得失的元变量
这个在 §4 已经展开。核心洞察是:这个 FIP 的成本承担有至少三种截然不同的路径,而对各方的利弊完全相反。
| 买单方式 | 谁受损 | 谁受益 |
|---|---|---|
| 储备金(推荐) | 无人直接受损;代价是未来新矿工类型的预算被用掉 | 质押者、SP(可赚奖励 + α)、整个网络 |
| FIP-0118 切分 | 存储提供者(每 EiB 收入降 30–65%) | 质押者、整个网络 |
| 增发(已被明确排除) | 所有持币人(稀释) | 质押者 |
为什么这是"元变量":它决定了你对"这个提案好不好"的回答方向。如果你只看机制不看资金来源,你会得出"这是个中立偏好的提案"——但这个结论在两种资金来源下都是错的。
操作建议:在任何关于这个 FIP 的讨论里,先问"钱从哪来"。FIP 正文刻意回避了这个问题("资金来源仍是开放的,本线程先测试机制"),这是一个策略性的回避——把最有争议的部分隔离出去,先让大家接受机制。你要意识到这个结构。
维度 3:安全会计——「可销毁 vs 可回收」到底意味着什么
这是提案安全论证的核心,也是最容易被过度简化的地方。
提案的论证:存储算力是可以无损转移的资产(连同扇区和 pledge 一起,无费用、无延迟),其 pledge 到期全额返还、主动终止返还 91.5%+;而作恶的质押 FIL 是被销毁的。所以同样规模的 ⅓,一个是"押金"、一个是"真金白银"。
但完整的安全会计必须加上三个修正项:
修正 1:攻击者的现金成本远低于名义金额
| 口径 | 20% QAP | 33% QAP | 51% QAP |
|---|---|---|---|
| Pledge(今天) | 13.1M FIL | 21.5M FIL | 33.3M FIL |
| Cash at risk(真金白银) | 1.3M FIL($1.0M) | 2.1M FIL($1.6M) | 3.2M FIL($2.5M) |
Cash at risk = 8.5% 终止费 + 约 1.2%(90 天、5% 融资成本)= pledge 的 9.7%。
更关键的一句(来自配套分析):如果矿工是低于其 pledge 折价买入、并任其到期,返还的 FIL 会多于成本——所以在困境价格下,现金数字可以为负。
这意味着"获得存储算力很便宜"这个论断比表面看起来更强——它削弱了"存储算力不如质押"的对比力度。
修正 2:质押侧的威慑依赖一个尚未规格化的子系统
罚没要真的能执行,需要:故障证明的消息格式、提交窗口、escrow 语义、委员会身份模型(烧谁的)、以及防止"解绑在罚没落地前完成"的时序保证。这些全部在 Appendix A 里标记为未决。 作者自己也说:"罚没子系统与最终罚没计划决定了威慑是否成立。"
修正 3:「必须超过双花价值」这个条件没有被验证
作者写得很清楚:
"那个数额必须超过快速最终性所保护的双花价值,且激活带应随结算规模增长重新检查。"
但配套分析里没有做这个计算。 我们知道 Filecoin 上的结算规模(Filecoin Pay、USDFC、桥的 TVL)在增长,但26.7M FIL 的可销毁资本是否足以覆盖潜在双花价值,这个判断没有给出。 这是一个应该自己补上的分析。
综合结论:安全论证的方向是对的(可销毁确实强于可回收),但幅度被高估了——因为存储侧的实际现金成本比名义 pledge 低一个数量级,而质押侧的威慑还没被证明可执行。
维度 4:反事实——不做会怎样、做别的会怎样
任何提案都应该在反事实里评估。作者给了还算完整的对照:
| 方案 | 内容 | 成本 | 能解决什么 | 不能解决什么 |
|---|---|---|---|---|
| 现状(什么都不做) | F3 继续停摆,网络跑 900 epoch | 0 | — | 容忍度停留在 ~20% 算力;FIL 无原生收益 |
| Discussion #1106(Kubuxu, 2025-01) | 用最多约 10% 的区块奖励给 F3 参与发钱,按证书证据分配,signer/producer 分成 | 对 SP 是净转移(total SP rewards 不变,只在提供者之间重新分配) | 修复 F3 法定人数 | 不创造新的安全基础(仍基于 QAP);不解决代币需求 |
| Vote escrow | 用锁定创造锁定,不改共识 | 低 | 代币锁定 | 不提供可罚没的最终性安全基础 |
| QAP→stake 渐变(blend) | 存储算力淡出、质押淡入 | 中 | 可立即恢复存储担保的快速最终性 | 引入 QAP 进最终性表;渐变期还要付费给存储签名者 |
| 本提案 | 纯质押加权 + 激活门槛 | 约 100M FIL | 关闭 900 epoch 深度回滚;创造可罚没安全基础;创造代币需求 | 见 §7.3 |
关键判断:#1106 是真正的竞争对手,因为它更便宜且能达成"恢复 F3"这个直接目标。作者对它的反驳是:
- 它把法定人数恢复到一个持续萎缩的安全基础上
- 它的抵押品是退出后返还的
- 它不解决代币需求目标
这三条你都要自己掂量。第 1 条在"存储确实在萎缩"的前提下成立;第 2、3 条是从提案自己的目标出发的论证(如果你不认同"需要代币需求"这个目标,第 3 条就不成立)。
一个更锋利的反问:如果目标只是"让 F3 恢复运行",#1106 用现有区块奖励的一小部分就能做到,为什么需要一个新的 1 亿 FIL 计划? 作者的答案本质上是"因为存储会一直萎缩,而我们需要一个不萎缩的基础"——这个答案的说服力完全依赖于对存储萎缩趋势的判断。
维度 5:时间与不可逆性——这个机制建成后很难拆
这是对长期持有者最重要的维度,但很少被讨论。
不可逆性的三个来源:
| 机制 | 效果 |
|---|---|
| sticky latch | 打开需要连续 3 个月满足条件,关闭需要连续 12 个月不满足。激活后分散度失败永不关闭 |
| 参数治理 | 所有参数(利率常量、激活阈值、罚没常量)都是协议常量,只能通过 FIP + 网络升级修改。作者刻意避免运行时可控的调参 actor,理由是这样会"给最终性经济创造额外攻击面" |
| 既得利益者 | 一旦有 80M+ FIL 质押进来并开始按 4-5% 收钱,降低或取消这个收益需要一次网络升级,并得罪一个已经组织起来的利益集团 |
还有一个"期权"性质的考量:提案明确说"采纳本提案不是决定把 Filecoin 迁移到 PoS"。但它同时承认,它为 PoS 转型建好了四个组件:
✅ 链上质押登记处(存款、委托、解绑)
✅ 作恶罚没 + 故障证明 + 相关性缩放
✅ 一个有已知深度和可测分布的可罚没质押池
✅ 运行一个委员会权力来自质押的 BFT 组件的实操经验
❌ 还缺的:基于质押的领袖选举(作者称之为"任何 PoS 转型中最大的未解组件")
怎么用这个维度:如果你是长期持有者,这个提案的真正价值可能是"它把 Filecoin 转向 PoS 的期权成本降低了一半"。这个期权是好事还是坏事取决于你的立场:
- 如果你认为 PoS 转型是 Filecoin 的未来 → 这是一个非常有价值的隐性收益,100M FIL 很便宜
- 如果你认为 PoRep 是 Filecoin 的立身之本 → 这是"滑坡的第一步",而 sticky latch 和协议常量治理意味着建成后很难拆
维度 6:可证伪性——哪些观测能推翻它
作者给了非常明确的证伪条件,这在这个行业里很少见,值得赞赏:
"第一个经验测试因此是:升级 epoch 后两年内,质押 FIL 是否达到 80M FIL 的安全水平,正如模型假设的那样。一个远低于该深度的池子将证伪 uptake 案例。"
而且作者自己的数据是偏悲观的(这是全文最大的内在张力):
| 指标 | 数值 | 方向 |
|---|---|---|
| 模型假设的可触达闲置资本 | 130M FIL(流通量的 14.6%) | 🟢 乐观 |
| 已接受链上收益的实测基数 | 约 40M FIL(流通量的 5%,2026-07-05) | 🔴 |
| GLIF(主导产品) | 约 32M FIL,3,590 个 iFIL 持有者 | 🔴 |
| FIL 计价存款趋势 | 2023-07 1.8M → 2024-07 34M → 2025-07 峰值 35.6M → 2026-07 31.9M | 🔴 2024 年见顶 |
| 年流失率 | 约 10% | 🔴 |
| GLIF 动向 | 正在向非 Filecoin 代币多元化 | 🔴 |
| 经验上限 | 约 10% 是承担信用风险的链上 FIL 产品能吸引到的供应比例上限 | 🔴 |
| 区间下沿 / 上沿 | 32M FIL(3.6%) / 267M–534M FIL(30–60% 跨链质押份额) | 🟡 极宽 |
作者自己的压力测试结果是:
在 40M FIL 可触达资本的情景下,资金池永远无法激活,三十年间支付的每一张证书都是建议性的,而账单是全需求成本的 97%。
这是我在整份材料里看到的最重要的一个数字。 因为它意味着:在最悲观但作者认为合理的情景下,这个机制会花掉几乎全部预算,却不产生任何安全收益。
同时还要注意:作者用 130M FIL 作为建模输入,但明确标注这是"靠判断力放进去的建模假设",不是实测值——而她自己列出的实测数据(约 40M,且见顶、流失、产品方正多元化)恰恰指向区间下沿。
这个张力值得你亲自判断:作者是用乐观假设跑模型、但诚实地把悲观证据一并列出来;还是假设本身就不够保守?
维度 7:未决项 → 风险传导(Appendix A 怎么读)
Appendix A 列了 11 个"决定"类开放项。不要把它当作待办清单读,要把它当作风险传导图读——每一项都阻塞某些人的具体工作。
| 未决项 | 阻塞谁 | 后果 |
|---|---|---|
| 委员会身份模型(worker key / 专用密钥 / 委托运营方密钥) | 下游未决项比任何其他问题都多 | 决定委托作恶时烧谁的质押;决定权力表是"每运营方一条含池化质押"还是"每存款人一条"。定案前,运营方无法写罚没披露,托管方无法写产品条款 |
| 每证书 gas 成本(未测量) | 架构决策 | 若过高 → 验证移到客户端 + 系统证明摘要(改变架构,且依赖 FIP-0107) |
N_min / max_share |
集中度分析(完全在链下) | 决定"分散"的定义;一个成功的运营方在激活前就被设了份额上限 |
| 故障证明格式与窗口 | 解绑延迟的取值 | 窗口变长 → 解绑延迟必须更长 |
| 基础与相关性罚没常量 | 罚没披露 | 决定实际风险敞口 |
α 在公布利率内还是之上 |
报价的基准 | 决定你看到的"净收益率"是什么意思 |
| 消息容器与批处理 | 工程边界 | 受 64 KiB 上限约束 |
| 公布利率是否应追踪链上 pledge 收益率 | 非续约通道的响应性 | 若 pledge 收益率下降,利率曲线不会自动调整 |
stake_activation 最终值 |
激活时点 | 80M 是提议值,最终值开放 |
W / W′ 计 epoch 还是实例 |
规范精度 | 实例是自然单位(分布统计按委员会计算) |
| inactivity leak(不活跃惩罚) | 明确不包含在本提案 | 作者给了重新审视的触发条件:若 stall 真被行使,或所需回报溢价跌破 0.25 个百分点 |
6. 多视角影响分析
读法提示:每个视角我都拆成「✅ 好的 / ❌ 坏的 /
⚠️ 取决于未决项」三档。⚠️ 那一档往往才是真正重要的——因为很多影响在当前文本下无法确定。
6.1 矿工 / 存储提供者(SP)
总体判断:在推荐方案下接近中性偏利好;在再平衡方案下明显受损。这是一个"看资金来源决定立场"的群体。
✅ 好的影响
| 事项 | 说明 |
|---|---|
| 推荐方案不削减区块奖励 | 91M FIL 全部来自从未动用的挖矿储备金,SP 的共识奖励完全不变 |
| 额外收入来源 | SP 作为 FIL 持有者同样有资格以统一利率质押赚取参与奖励——不触碰扇区、pledge 或区块生产 |
| includer share | SP 的区块生产者提交链上证书可赚 α ≈ 10%,这是一个新的、与其现有工作(出块)自然结合的收入流 |
| 可能获得更便宜的 pledge 融资 | 配套的探索性文档提了一个"两类质押"想法:让矿工以约 5% 借入 pledge(对比今天 GLIF 的 15%),能把融资成本从"两年都亏损"翻转为"有毛利"。但这不是本 FIP 的一部分,作者明确说"草案不依赖它" |
| 单位算力收益可能改善 | 若部分资金真的从存储流出(见下),留存的矿工的单位算力收益反而上升——这对坚持留下来的人是利好 |
❌ 坏的影响
| 事项 | 说明 |
|---|---|
| ❌ 若采用再平衡方案:直接减收 | per-EiB 共识收入降到 Solstice 已通过版的 70%(35% floor)/ 60%(30%)/ 50%(25%)。30 年 SP 共识奖励从 84.2M → 68.0M FIL |
| ❌ 失去 F3 参与资格 | 委员会资格从"≥10 TiB 算力的存储提供者"变成"≥最低质押额的质押身份"。现有 SP 的 F3 参与软件不再具备资格,需要弃用说明 |
| ❌ 新增运营负担 | 区块生产者必须提交证书(否则放弃 α),这是一个新的软件义务 |
| ❌ 潜在的资金外流通道 | 质押利率把 pledge 资金吸引出去有两条路(见下),可能加速算力下降 |
⚠️ 取决于未决项 / 需要自己判断
这里有一个反直觉的传导链,值得单独拆开:
质押收益率吸引存储侧资金 → 两个通道:
通道 A:提前终止(early termination)
- FIP-0098 收初始 pledge 的 8.5%
- 模型显示:公布利率超过这个门槛的时间只有 30 年中的 4 个月
- 且单次提前终止在最高 5 年的剩余扇区寿命内都无法收回这笔费用
→ 结论:这条通道基本关闭 ✅
通道 B:不续约(non-renewal)
- 让扇区自然到期,再把返还的 pledge 拿去质押,无终止费
- 只需打败存储提供者在承诺 pledge 上的净回报:模型值约 3.5%/年(当前价格)
- 模型显示:公布质押利率在全部 30 年都高于这个水平
→ 结论:这条通道是敞开的 ⚠️
但作者给了一个重要的反向论证:
当质押池足够深时,边际质押收益降到约 4.3–4.9%,低于价格高于现货时的承诺 pledge 回报($1.00 时 9.5%、$2.00 时 19.6%)。
而且高初始利率是容量受限的——闲置资本进场比存储承诺资本解绑快得多。
同时作者也承认:现有证据表明存储算力趋势主要由存储硬件和交易经济学驱动,参与收益是次要因素。
- "存储算力由硬件和交易经济驱动"这个判断如果错了,通道 B 的实际影响会被显著低估。作者的建模"建立了从共识奖励到存储侧回报的反馈方向,但尚未建立其量级"。
r(S,P)只响应质押深度,不响应 pledge 收益率。 作者自己把这一条标为[needs precise spec]——如果 pledge 收益率因为所采用的区块奖励时间表而下降,质押利率曲线不会自动调整。这意味着矿工可能同时面对"pledge 收益下降"和"质押收益不变(相对更有吸引力)"的双重挤压。- "两类质押"想法如果被采纳,对矿工是重大利好(5% vs 15% 融资),但它降低池子的攻击成本——配套分析承认:攻击者为一个 ⅓ 算力攻击所必须购买的浮存会从 26.7M FIL 降到 20M(20M 锁定)或 13M(40M 锁定),而且攻击者已经持有的约 21.5M FIL pledge 可能自锁进池子,反向计入它本来要被防御的同一个 ⅓ 权重。作者的原话是:"这个上限没有被任何人定价"。
6.2 持币人(普通 FIL 持有者)
总体判断:这是提案最直接瞄准的受益群体,但"收益"的成色需要仔细看——它不是免费的钱。
✅ 好的影响
| 事项 | 说明 |
|---|---|
| 第一次有协议原生 FIL 收益 | 此前流动 FIL 没有任何协议原生收益。这是一个真实的产品空缺被填补 |
| 不需要跑节点就能参与 | 通过委托质押(delegated staking)路径,把 FIL 交给运营方即可——FIP 明确说这是"预期行为,不是 gaming" |
| 起始利率很高 | 12% 天花板,前 80M FIL 是"在一个不断下降的价格上被招募的"——早期参与者拿得最多 |
| 不新增铸币 | 所有资金来源方案都不改变铸币曲线、2 亿上限不动。不存在"用通胀给质押者发钱"的稀释 |
| 不锁定(at-will) | 它是延迟,不是固定期限——不是 lock-up,只是 90 天解绑等待 |
❌ 坏的影响
| 事项 | 说明 |
|---|---|
| ❌ 收益率是毛值 | 未扣 includer share α 和领取 gas。领取是 pull-based,需要你自己发起并付 gas |
| ❌ 90 天解绑 | 你质押的 FIL 不是随时能拿回来的。这不是日内交易工具 |
| ❌ 罚没风险 | 你的质押可能被销毁——而且烧谁的在当前文本下未定(见 |
| ❌ 利率必然下降 | 从 12% → 4.0-5.0% 是设计的一部分,不是暂时的。到 100M 质押时只有 4.26% |
| ❌ 存在静默零收益门槛 | 质押低于总量 1/65535(100M 池中约 1,526 FIL)会被归一化为 0 权重,但仍以正权力条目出现在名单里——你无从得知自己签不了名、赚不到钱 |
| ❌ 如果采用增发方案(已被排除但值得知道) | 会被稀释。作者已明确排除这条路,这是加分项 |
⚠️ 取决于未决项
| 事项 | 为什么重要 |
|---|---|
| 委托模式下,签名密钥属于运营方,而被罚没的资本主要属于委托者。到底烧运营方保证金、委托者按比例分摊、还是两者按顺序——取决于"委员会身份模型",而它未定 | |
| 公布的净费率不是协议承诺 | |
早质押不锁定任何东西;但奖励支付会降低 P、计划释放会提高 P、新质押会提高 S——在模型范围内深度项占主导。所以"等更高利率"通常不是好策略,但也不是无风险的判断 |
|
| 大质押者可以压着存款不让进以维持高利率,代价是放弃那部分奖励;把仓位拆成多个身份不会增加奖励,因为利率是统一的 |
💡 一个需要注意的结构性事实
质押 FIL 仍然计入流通供应量。 因为
FilCirculating是共识输入(决定每扇区 pledge 锁定目标)。质押锁定只作为流动浮增(liquid-float)报告行出现 = 流通量 − 质押量。对你的意义:质押不改变市值口径,所以"质押率上升"不会让 FIL 看起来更稀缺。如果你期待的是"锁仓→减少流通→推高价格"的叙事,这个机制不提供。
6.3 委托质押者
总体判断:这是风险/收益最不对称的群体——你买的是一个产品,但产品条款还没写完。
你买的是什么:FIL 进,参与奖励的一部分出,无需运行节点、无需持有签名密钥。FIP 认为这是预期之中且健康的行为。证书条件明确禁止对闲置余额收取奖励(所以运营方必须真的在签名)。
✅ 好的影响
- 低门槛获得协议原生收益
- 无需运维、无需密钥管理
- 运营方承担签名工作,与你分享奖励
❌ 坏的影响
- 奖励是 pull-based,需要有人执行领取并付 gas(运营方或你自己)
- 退出只能按 90 天解绑,不能日内交易
- 公布的净费率是运营方自己的算术,不是协议承诺
⚠️ 最大的风险点:罚没的承担者未定
委托模式下:
签名密钥 → 属于运营方
被罚没的资本 → 主要属于委托者
→ 到底烧谁?FIP 记录为未决
这个未决项的连锁影响最大——配套文档原话是"依赖它的下游未决项比任何其他问题都多"。定案之后:运营方才能写罚没披露,托管方才能写产品条款。
给你的一句话结论:在这个未决项定案之前,你无法计算自己作为委托者的真实风险敞口。 如果看到有产品在当前阶段就开始招揽委托,要意识到它的风险披露不可能是完整的。
6.4 交易所 / 托管方
总体判断:工作量明确、不性感,但不做就无法支持这个产品;结构性风险在于"政府可施压"。
需要新建的东西
| 需求 | 说明 |
|---|---|
| "已质押并累积中"余额概念 | 质押 FIL 离开持有人的账户余额,存在于 actor state 中。账本、对账单、对账(reconciliation)都必须能承载它。而且这个余额只在签名落入规范证书时才增长——不是线性的 |
| 拉取式领取的执行方 | 奖励只在领取时到账,需要有人执行领取并付 gas |
| 提款条款中的解绑延迟披露 | 托管方若提供 FIL 质押,就是在提供一个按 90 天解绑才能退出的金融工具。这是披露问题 |
⚠️ 一个作者明确点出、但没有解决的结构性风险
分散度规则约束的是"参与方",不是"司法辖区"。
委托质押会趋向交易所与托管方——而它们是政府可以施压的受监管公司。
司法辖区集中度是一个独立指标。 提案"不测量也不解决这个风险"。
但作者给了一个重要的损害上界:
被强制的托管方最多只能把最终性拖延回今天网络的状态(900 epoch 规则),而无法破坏安全性——因为它们还需要存储算力才能重组。
这是一个相当有力的辩护:这个提案把"被胁迫的托管方"的最坏后果限制在"回到现状",而不是"破坏链"。对照一下:如果今天某个持 ⅓ 存储算力的实体被胁迫,后果会更严重。
6.5 项目方 / 基金会
总体判断:获得了叙事和基础设施,但背上了一堆非协议的、没人承诺要做的工作。
获得什么
- 一个可讲的故事:FIL 第一次有了协议原生收益,可以吸引此前无法触达的资金
- 一个不随算力萎缩的安全基础:如果存储萎缩的判断成立,这解决了协议层的一个长期隐患
- PoS 转型的四个组件(如果生态未来选择那条路)
- 社区讨论的抓手:作者做了一整套配套分析,讨论质量很高
需要承担什么
| 事项 | 说明 |
|---|---|
| 出资约 91M FIL | 从储备金拨付,需要 FIP + 网络升级(走 FIP-0085 的 keyless account 路径)。这确立了"FIP 可以花储备金"的先例——作者强调该先例被单次事件、上限、付款速度上限、日历化复核、日落销毁所限定 |
| 完全链下的招募活动 | 作者称之为"可以说更大的那一摊工作"。需要在 16 个月内(按 5M FIL/月)招募约 80M FIL 并实现分散 |
| 持续的集中度分析 | 设定分散度常量 N_min/max_share 的分析完全在链下运行,是常态化任务 |
| 追溯未知持仓 | 最大的单一未标记持仓为 155M FIL(约占聚合方流通量的 19%,创建于 2020-11),地址历史仍有待追溯。把余额归因到交易所、托管方、ETF、公司金库也是待做的研究 |
| 推动未决项定案 | 每一个未决项都"释放某个人必须交付的工作",而委员会身份模型是根节点 |
⚠️ 最尴尬的可能结果
在 40M FIL 可触达资本的情景下:资金池永远无法激活,三十年间支付的每一张证书都是建议性的,账单却是全需求成本的 97%。
即:花了几乎全部预算,什么都没买到。 这是项目方需要认真对待的下行情形,而作者自己的实测数据(GLIF 约 32M、基数 2024 年见顶、年流失 10%)恰恰落在这个区间附近。
作者的对冲:第 3 年未激活复核(之后每年一次)+ 清盘(wind down)路径——可以移除 stream 条目、停止计提、保持索赔开放、提前销毁。这是负责任的设计,但你仍然付掉了 3 年的钱。
6.6 开发者 / 客户端团队
总体判断:工作量真实且分散,且有几个"必须做对否则炸"的点。
| 需求 | 说明 |
|---|---|
| 所有客户端都要写新代码 | go-f3、lotus、forest,用两种语言实现,用共享一致性向量(conformance vectors)保证逐位一致 |
| actor bundle 是唯一共享交付物 | 含 staking actor、奖励池、证书提交与验证路径、罚没逻辑 |
| 其余全部各客户端独立实现 | 权力表派生、最终性门控、参与方 API、供应量核算、面向质押者的工具、迁移 |
⚠️ 三个"必须做对"的点
| 风险 | 后果 |
|---|---|
| 一个非 BLS 条目会中止整个权力表构建 | 今天在 lotus 与 forest 就是这样。这意味着一次非 BLS 的准入即构成"对所有人的快速最终性的一次注册否决" |
| lotus 与 forest 必须在同一版本加入流通量分类 | 状态遍历 RPC 对未分类 actor 硬报错。奖励池与质押 actor 都必须被加入分类 |
| lotus 中密钥必须热放在 daemon 钱包里 | 远程签名被标记为未来工作。这是重大的运维含义——运营方必须把签名密钥放在联网节点上 |
其他工程负担
- 快照生产者必须全部重做:新网络名导致 P2P topic 与 datastore 前缀变化,所有已发布的 F3 快照失效
- 每证书 gas 未测量:在部署阶段 2 测量。若高到不可接受 → 退路是把验证移到客户端,actor 消费系统证明摘要——这会改变架构,且依赖 FIP-0107
- 消息容器、连续提交的批处理方式未定,64 KiB 上限是边界条件
- 参与方 API 需要把证书的 signer bitset 解析为逐参与方的纳入情况——好消息是这部分已经部分存在(go-f3 v0.8.14 新增;lotus 与 forest 截至 2026-08-05 已具备)
6.7 集成方(交易所、桥、L2、Filecoin Pay 等最终性消费者)
总体判断:必须改代码,不改会出事;但改完之后拿到了"7.5 小时 → 数十秒"的真实收益。
✅ 收益
| 事项 | 说明 |
|---|---|
| 结算时间从 ~7.5 小时降到数十秒 | 如果 binding 证书被当作结算依据 |
| 可观测性 | 集成方可以从链上状态判断快速最终性是否权威——能工程化地实现自己今天已经在用的回退方案 |
| 地板保证 | 三种状态中最差的那个也等于今天的网络,不会比现状更差 |
❌ 必须做的工作
| 事项 | 说明 |
|---|---|
| 重新初始化 | 新 bootstrap epoch、新初始权力表、新网络名 → P2P topic 与 datastore 前缀变化,影响快照生产者 |
| 否则现有的智能合约模式会把建议性证书当作最终接受——这是配套文档明确点名的安全缺口 | |
| 承受一次性标签跳跃 | 节点上报的 finalized = EC 最终性与最新证书 head 的较晚者。由于此前建议性证书被抑制,在 binding 那一刻这个标签会一次性向前跳最多约 900 个 epoch。派生 safe 标签的系统会观察到这个不连续 |
⚠️ 一个重要的信任假设
读取权威标记有两条路径:
- 节点:直接读 staking actor 的只追加激活记录 ✅
- 只持有证书的验证器:从委员会的补充数据承诺读取 latch —— 即委员会为自己的权威性背书
作者自己指出了这个局限:
这个标记只有在委员会已经具备权威性之后才可靠——而这正是激活所建立的东西。它本身无法阻止一个激活前的委员会声称自己有权威性。
缓解方案:集成方可以像钉住网络升级高度一样钉住一个已发布的激活 epoch。但因为 §6 也允许后续去激活,这个集成方仍然需要一个可信的方式来学习关闭事件。
🔴 一个被低估的风险:短回滚在所有制度下都存在
配套的攻击分析给出了一个很硬的结论:
| 深度 | 集成方举例 |
|---|---|
| 0 | Filecoin Pay settleRail(结算到 head);USDFC 清算与赎回(无确认等待);Pyth pull 更新;lotus StateWaitMsg |
| 1 | lotus latest;Hyperlane 的 Filecoin 路由(confirmations: 1, reorgPeriod: 0);synapse-sdk 同步调用;Curio 的 FEVM watcher(MinEthConfidence = 1) |
| 3 | Axelar 部署工具 |
| 5–10 | Curio 的 Filecoin-message watcher;lotus MessageConfidence(5);Boost sector-committed(6);PublishStorageDeals(10) |
| 60+ | Coinbase、Binance、Bitget、HTX、Bithumb(60);KuCoin、Coinone(120);Kraken(300);提现在 900 |
结论:1–6 个 epoch 的短回滚在三种制度下都是 20% QAP,成本不下降。
因为 Filecoin Pay 在 head 处结算 rail,在同一个 epoch 内对结算采取行动的收款方正处在"持有 4 个 owner 矿工的攻击者"可达的窗口内。
好消息:没有任何交易所会在这个窗口内入账。
给你的判断:这个 FIP 对集成方的价值是分层的——交易所(depth ≥60)今天就是安全的,收益主要是"更快";而 depth 0–3 的集成方(Filecoin Pay、USDFC、Pyth、Hyperlane)今天就在 20% 攻击者的射程内,而本提案不改变这一点。
6.8 整个网络
总体判断:关闭一条攻击路径,打开一个新的否决权,引入一套新的复杂度和集中度风险。净效果取决于你对"存储萎缩"的判断。
✅ 好的影响
| 事项 | 说明 |
|---|---|
| 关闭 900 epoch 深度回滚 | 这是唯一靠质押加权能关闭的攻击路径。今天需 20% QAP,QAP-F3 下需 33% QAP,质押加权下仅凭存储算力根本执行不了 |
| 容忍度从 ~20% 提到 ~33% | 从 EC 的实际水平提升到 BFT 标准(但只在证书 binding 时成立,见 |
| 创造了一个不随算力萎缩的安全基础 | 如果存储萎缩的判断成立,这是结构性改善 |
| 可罚没资本确实是一个新的安全预算 | 模型里 80–130M FIL 的质押在第二年到第一个十年末之间就位,这在今天是不存在的 |
| 激活门槛用协议规则强制了最低深度 | 而不是交给市场的存储算力水平 |
❌ 坏的影响 / 新增攻击面
| 事项 | 说明 |
|---|---|
| ⅓ 的注册质押可以拒绝签名而不作恶——§2 计的是注册质押而非参与质押,所以 §4 的罚没永不触发。后果是回退到 900 epoch 规则 | |
| 获取⅓ 池子的成本 30–43M FIL,但持仓成本只有约 1.42M FIL/年(约 0.08%/周) | |
| 作者的原文:一个单一假名买家可以替代八到十个可识别的在位者,"这些在位者的协调难度是今天重要的威慑因素"。同时 carry cost 下降了大约五倍 | |
| 见 §6.7 | |
| 区块生产在所有制度下仍由存储算力加权 | |
| Lido 通过类似的聚合路径达到了 ETH 全部质押的约 23%。分散度规则约束参与方而非辖区;供应链身份的廉价增殖使这些规则"必要但不充分" | |
| 新 actor、新参数、新复核机制、新供应量核算 | |
| "一个集中度的发现不会关闭 latch"——修正需要一次 FIP 和网络升级 |
⚠️ 对"安全性提升"这个卖点的三条修正
修正 1:增益只在证书 binding 时存在,且 stall 可以零成本关闭它。
作者的辩护(我认为站得住,但改变了性质):
三件事限制了否决权的价值:
- 已经最终化的一切保持最终——证书不能被"拒绝签下一张"召回
- 重组仍然需要存储算力(回退规则下约 20% QAP),和今天完全一样,不更容易
- 第二把锁不能被悄悄撬开——证书停止本身就是警报,几小时内可见于每个交易所和桥,它们会回退到慢确认,把大部分战利品拿走
作者的结论很克制:
"对于逐利者,算术不闭合;对于只想造成伤害的攻击者,下面的 carry 和可见性就是威慑——本节不做更多声称。"
这意味着"安全性提升"的性质应该被重新定性:它不是"防止攻击",而是**"给集成方争取可见的反应时间 + 让已经最终化的历史不可撤销"**。这是有价值的——但和"阻止攻击"是两回事。
修正 2:停摆是自我限制的。
没有人能在 stall 期间赚取参与奖励,诚实质押可以退出,池子会向 §6 的去激活底线滑落——到那时 binding 最终性回到 advisory,而不是继续被停摆的委员会挟持。
否决权的价格因此随池子深度下降,在模型路径中到第 25 年降到 9.1M FIL。
修正 3:集中质押的权力比集中算力小。
持⅓ 质押的实体可以让证书 stall。即使持⅔ 质押,它也不能出块、不能审查交易、不能单独重写状态、不能铸造 FIL——因为领袖选举仍由存储算力加权。
⅔ 时额外获得的能力是"把被俘获的最终性与存储算力攻击结合,为一次重组背书",这需要大约两倍于否决权的质押资本 + 单独的存储算力仓位,而质押集中度在链上表中是可见的。结论:集中的质押对协议的控制力小于集中的存储算力,因为存储算力选举每一个区块。
这条论证我认为是提案最有力的部分。 它诚实地说明了这个机制的能力上界。
📊 一张总结矩阵
| 群体 | 推荐方案下的净效果 | 再平衡方案下的净效果 | 主要不确定性 |
|---|---|---|---|
| 矿工 / SP | 🟢 中性偏利好 | 🔴 明显受损 | 资金来源;协议收益率挤压 pledge 的通道 |
| 持币人 | 🟡 中性偏利好(有收益但递减、有风险) | 🟡 类似 | 罚没烧谁的;净收益率的真实数字 |
| 委托质押者 | 🟡 风险收益不对称 | 🟡 同 | 委员会身份模型(决定风险敞口) |
| 交易所 / 托管方 | 🟡 工作量明确,结构性风险有上界 | 🟡 同 | 司法辖区集中度(提案不解决) |
| 项目方 / 基金会 | 🟢 拿到叙事,负担不小 | 🟡 同 | uptake 是最大变量;招募能力 |
| 开发者 / 客户端 | 🟡 工作量真实,三点必须做对 | 🟡 同 | 每证书 gas(未测量) |
| 集成方 | 🟢 depth ≥60 的获益;depth 0-3 的没变化 | 🟡 同 | binding/advisory 区分必须做对 |
| 整个网络 | 🟢 关闭深度回滚 + 🟡 新增 stall 否决权 | 🟡 同,但可能加速算力流失 | 存储萎缩的判断是否成立 |
6.9 与现实数据的对照 —— 这几个数字会改变你的判断
以上分析大多基于提案的模型。把模型放回现实,有几个数字值得单独拎出来。
🔴 最重要的一条:现有竞品已经在付更高的利率
| 项目 | 利率 |
|---|---|
| GLIF(iFIL)实测存款 APY | 5.583% |
| 提案在 80M FIL 目标深度时的公布利率 | 4.89% |
| 提案在 100M 时的公布利率 | 4.26% |
这是我在核实过程中发现的最尖锐的一处矛盾。
作者论证招募可行性的核心是"约 40M FIL 已经接受了链上收益,说明需求存在"。但:
- 最大的那个产品(GLIF,32.9M FIL)现在付的是 5.583%——比提案到达目标深度时愿意付的 4.89% 还高
- 而且 GLIF 的利率是市场定价的(借贷池利用率驱动),提案的利率是被释放表固定死的、只会往下走
- 提案自己的模型说"超过 77M FIL 质押会让利率公示到 5% 以下,于是流入停止"——但现实中已经有人在付 5.583% 了
作者对此的可能反驳(需要你自己权衡):
- GLIF 的 5.583% 是承担信用风险的(Hindenrank 对 GLIF 的评级是 C/C+,提到"抵押不足的 SP 借贷"和"iFIL 赎回可能延迟数周"),而协议原生质押是无信用风险的
- 5.583% 是当前利用率下的瞬时值,会波动;提案的利率是可预期的
- 两个池子的风险性质不同,不能直接比利率
但无论如何,这个对比说明"5% 左右就能招到 80M FIL"这个假设比草案呈现的更脆弱。 而且要注意:GLIF 的存款利率越高,说明它借出去的价格(草案说 15%)越真实——这个利差正在被一个未受监管、有信用风险的产品捕获,而提案想用一个更低、更安全的利率把它抢过来。
🟡 其他改变判断的现实数据
| 数据 | 值(2026-09) | 对判断的影响 |
|---|---|---|
| FIL 价格 | $0.79–0.87(草案用的是 $0.77) | 攻击成本的美元口径上升约 10%;但仍然极低——33% 算力的现金风险敞口约 $1.6M |
| 全网 QAP | 12.0 EiB(峰值 25.1 EiB,2024-02) | 存储萎缩的幅度比提案呈现的更严重 |
| 全网 pledge | 64.03M FIL(峰值 176.12M) | 提案说"65.3M 且在下降"✅ 准确,现在更低 |
| 有算力的 SP 数 | 521(峰值 4,100) | 四年缩水 87%。委员会基数被严重削弱,这与 F3 停摆同源 |
| GLIF TVL | 32.9M FIL | 提案说"约 32M"✅ 准确 |
| GLIF 存款人数 | 6,459(第三方口径) | 提案说的"3,590 个 iFIL 持有者"无法独立核实(口径不同,iFIL 持有人 ≠ 池内存款地址) |
| Filecoin Pay 年化付费量 | 约 $140K | |
| Grayscale FILG 持仓 | 2,633,200 FIL,AUM 约 $1.56M | 机构持仓基本可以忽略。没有美国现货 FIL ETF |
| FIL 24h 成交额 | 口径冲突:$481.6M(数据门户聚合)vs $92–134M(Binance)vs $24.8M(TokenInsight 现货) | 流动性集中在 Binance(约 35%)+ OKX(约 21%) |
🟢 两个对提案叙事有利的宏观背景
| 事实 | 为什么对提案有利 |
|---|---|
| 2026-10-15:PL 与 Filecoin Foundation 的归属期结束 | 此后最大的新增 FIL 来源(约 66.7M FIL/年)消失,只剩区块奖励(约 21.7M FIL/年 ≈ 流通量的 2%),总发行量减少约 75%。模型显示 2027 年底可能转为净通缩。"FIL 需要一个不靠发行的需求来源"这个叙事因此更有说服力 |
| 2026 Filecoin Network Strategy(2026-02-19) | 明确从"Storage Capacity Network"转向"Service Economy Network","从扩张供给转向放大需求",并写入"改善存储提供者经济学"。提案的方向与官方战略一致 |
🔴 但官方叙事回避了提案的核心前提
这是一个值得注意的政治现实:
Filecoin Foundation 与 Protocol Labs 没有任何公开声明直接承认或讨论"存储提供者 pledge 下降""算力萎缩"这一具体现象。搜索 "pledge decline"、"storage provider exit"、"network economics" 只找到第三方风险机构(Hindenrank 把 "storage provider exodus" 列为 Filecoin 首要风险项)和社区提案(Daybreak / Discussion #1238 说 SP "basically bleeding money and running at a loss")的说法。
官方口径始终是"转向服务经济"的正面叙事,没有正面回应算力/质押萎缩。
这有两层含义:
- 这个 FIP 等于把一个官方尚未承认的结构性问题摆上了台面——这是它的价值,也是它推进的政治阻力所在
- 如果你在评估"这个提案会不会被采纳",要考虑它触动了官方叙事的一个禁区
7. 批判性评估
7.1 站得住的部分
| # | 判断 | 理由 |
|---|---|---|
| 1 | 诊断是对的 | F3 失败的根本原因是"签名不给钱",这是可观察的事实,不是推测。提案直接对症 |
| 2 | "可销毁 vs 可回收"的方向性论证成立 | 存储算力确实可以无损转移、pledge 确实可回收、质押确实可烧。这个对比的性质是对的(幅度需要打折,见 7.2) |
| 3 | 递减利率 + 无下限 + 30 年日落的设计很优雅 | 作者对以太坊「EIP-8363 所说的『没有一刻质押激励会关闭』」的批评是切中要害的。一个会把利率棘轮立起来、需要后续治理才能关掉的设计,确实比一个"设计上就会自己熄灭"的设计更差 |
| 4 | 证书门控解决了搭便车问题 | 给所有注册质押发钱会重现 F3 的搭便车困境。要求"签名进入有效的规范证书"是为实际服务付费,且是自我限制的——没有证书就没有支付 |
| 5 | 分三阶段部署,风险分离 | 每阶段只依赖前一阶段演练过的组件。部署风险 / 委员会运营风险 / binding 最终性风险被清晰隔开。这是很成熟的工程思维 |
| 6 | 分散度要求承认了"深度 ≠ 安全" | 且作者主动承认 N_min/max_share 是必要非充分条件(链上身份可廉价增殖)。没有夸大自己机制的能力 |
| 7 | 激活后分散度失败不关闭 latch 的理由充分 | "否则攻击者可以用制造集中来获得一个关闭开关,比 stall 更便宜"——这个反事实推理是对的 |
| 8 | 明确排除增发 | 所有资金方案都不动铸币曲线。这是一个重要的道德优势 |
| 9 | 对"stall 否决权"的处理非常诚实 | 作者没有掩盖这个弱点,反而用一整段论证它的上界,并明确说"本节不做更多声称"。这种克制在协议提案里很少见 |
7.2 存疑的部分
| # | 存疑点 | 我的分析 |
|---|---|---|
| 1 | "存储算力优于质押"的对比幅度被高估 | 提案拿 26.7M FIL 可销毁 vs 21.5M FIL 可回收做对比。但配套分析的"Cash at risk"口径显示:存储侧的真金白银风险敞口只有 pledge 的 9.7%(8.5% 终止费 + 1.2% 融资成本),在困境价格下可以为负。用名义 pledge 对比会显著低估存储侧的攻击便利性 |
| 2 | "安全性从 ~20% 提升到 ~33%"需要大幅打折 | (a) 只在 binding 时成立;(b) 任何持有 ⅓ 质押的人可以零成本 stall 掉它,把网络打回 20%;(c) 作者自己说"最终性容忍度的增益只在否决权未被行使时成立"。更准确的表述是:这个提案关闭了一条具体攻击路径(深度回滚),并给集成方争取了可见的反应时间 |
| 3 | "必须超过双花价值"的条件没有被验证 | 作者写"那个数额必须超过快速最终性所保护的双花价值",但配套分析没有做这个计算。26.7M FIL(按 $0.77 约 $20M)是否足以覆盖 Filecoin 上不断增长的结算价值?这是一个应该自己补上的关键分析 |
| 4 | 罚没的可执行性完全依赖一个未规格化的子系统 | 故障证明的格式、窗口、escrow 语义、委员会身份模型全部未定。作者自己也说"罚没子系统与最终罚没计划决定了威慑是否成立"。在规格化之前,"可销毁"只是一个承诺,不是一个事实 |
| 5 | 对"攻击者本身可能是 FIL 多头"这一情形的处理不足 | 作者用"carry cost 1.42M FIL/年"论证 stall 很贵。但如果攻击者的动机本来就是长期持有 FIL,机会成本可近似为零。"两类质押"文档恰恰承认了"攻击者已经持有约 21.5M FIL pledge"——这个矛盾没有被正面处理 |
| 6 | uptake 的证据指向悲观,但建模假设指向乐观 | 见 §7.3 第 1 条,这是全文最大的内在张力 |
| 7 | "更深的共识削减会加速扇区终止"这个循环未建模 | 作者明确承认。这意味着再平衡方案的 SP 影响可能被低估(因为 SP 不是被动减收,而是会主动终止扇区、移除算力) |
| 8 | 两处 SP 收入的绝对数字在两份配套文档中不一致 | 页一:30 年 SP 共识奖励 56.05M vs Solstice 72.32M;页二:67.991M vs 84.197M。比率一致(都是 Solstice 的约 70%),差异很可能来自发货/启动日期与情景配置不同,但两页均未解释。引用时要注明是哪一页 |
7.3 明显薄弱 / 未回答的部分
① 最大的问题:uptake 假设与作者自己的证据方向相反
这是我认为整份材料里最重要的一处张力:
| 作者用作建模输入的假设 | 作者自己列出的实测证据 |
|---|---|
| 可触达闲置资本 130M FIL(流通量的 14.6%) | 已接受链上收益的实测基数 约 40M FIL(流通量的 5%) |
| GLIF 约 32M FIL(3,590 个 iFIL 持有者) | |
| 基数 2024 年见顶,之后年流失约 10% | |
| GLIF 正在向非 Filecoin 代币多元化 | |
| stFIL 的失败"烧掉了一批再未回场的人群" | |
| 经验结论:约 10% 是承担信用风险的链上 FIL 产品能吸引到的供应比例上限 |
作者明确标注 130M 和 5M FIL/月"是靠判断力放进去的建模假设,不是实测值"——这是诚实的。但她列出的所有实测证据都指向区间下沿(32M)而不是建模输入(130M)。
而压力测试的结果是:
40M FIL 可触达资本 → 资金池永远无法激活 → 30 年付掉全需求成本的 97%
给你的判断框架:如果实测基数确实只有约 32M,且这个基数在萎缩,那么这个机制最可能的结果是"付了很多年的建议性证书,从未 binding"。这不是灾难(不会更糟),但1 亿 FIL 买了一个可能永远不生效的东西。
要追踪的信号:作者说"第一个经验测试是升级后两年内质押是否达到 80M"。但这个测试来得太晚了——你要的是更早的信号。我建议追踪:升级后 3 个月、6 个月的月均流入(模型假设是 5M FIL/月)。
② 储备金的"机会成本"没有被讨论
推荐方案花掉 91M FIL 储备金(32%)。这笔钱创世时是为了"未来尚未确定的新矿工类型"预留的。如果未来真的出现了那种新矿工类型,这笔钱就没了。 提案把这个成本记为"零",但它其实是一个被放弃的期权。作者没有做这个定价。
③ 没有任何关于 FIL 价格的讨论
整个模型"不考虑价格"。但:
- FIL 价格决定攻击的美元成本(今天 $0.77,很便宜)
- FIL 价格决定"双花价值"的比较基准
- FIL 价格决定质押者是否愿意接受 4-5% 的 FIL 计价收益(如果预期 FIL 贬值,4% 的 FIL 收益毫无吸引力)
这是模型的一个重大简化,而它恰好是持币人最关心的变量。
④ "两类质押"想法与主提案存在张力,但被搁置
配套的 two-class-stake 文档提出:让矿工以 5% 借 pledge(对比 15%),奖励池支出降 12-24%。但它同时把攻击者必须购买的浮存从 26.7M 降到 13M FIL,且攻击者已有的 21.5M FIL pledge 可以自锁进它本该被防御的同一个 ⅓ 权重。作者原话:"这个上限没有被任何人定价。"
这个想法被标记为"探索性,草案不依赖它"——但它描述的问题(质押池的收益率打不过 GLIF 的 15% 借款利率)是真实存在的,而且它可能会在后续讨论中被重新提起。关注这个。
⑤ 提案对"PoS 滑坡"的回应是形式上的
作者说"采纳本提案不是决定迁移到 PoS",但承认建好了 4 个组件,只缺 leader election。这是一个诚实的回答,但它没有回答真正的问题:如果这 4 个组件建好了,下一次有人提"我们只差 leader election 了"时,反对的成本是不是低得多?
这不是提案的缺陷,而是你需要自己判断的战略问题。
⑥ 链上验证证书的技术前提未被确认
草案说"所需的 BLS12-381 运算已在 FIP-0105 下提供,不需要新 syscall"。但:
- FIP-0105 的标题与定位都是"FEVM Precompiles",原文只声明"在 FEVM 中引入预编译函数"
- 而草案 §3a 假设的是在内建 actor(reward-pool actor)中直接验证证书
- FIP-0105 原文并未声明 built-in actor 可以直接调用这些运算
草案自己把"每证书 gas 成本"列为"未测量"的早期交付物,并准备了客户端验证的退路。综合来看,"不需要新 syscall"这个论断在 FIP-0105 文本层面未获确认,是一个应该标注为待验证的技术缺口。
⑦ 最根本的:招募问题被识别为最大风险,但解决方案完全不在协议内
作者自己说得很清楚:
"可能更大的那一摊工作是招募(arguably the larger endeavor)……协议里没有任何东西承诺任何一方去构建那套软件或招募质押。"
"跨越阈值需要说服持有数千万 FIL 的人去质押,这取决于协议之外的触达和分发。"
这意味着:这个提案把最难的问题(如何说服 8000 万 FIL 来质押)明确地推给了协议之外,而它自己在协议内提供的是一个"如果人来,机制就工作"的设计。
考虑到现实基线(现有链上生息产品约 40M FIL,且 2024 年见顶、年流失 10%、最大产品正在多元化),这是一个非常强的假设。而且核心开发者层面已经有人提出同样的疑问(Core Devs 94 上 Jennifer Wang:"约 80M FIL 的目标是否可达?如果永远达不到怎么办?")。
7.4 对作者诚实度的评估
值得单独说:这份草案的自我批判质量远高于行业平均水平。具体表现:
- 主动给反对者递刀子("如果你不接受前提,这个工具就是错的",并列出更便宜的替代方案)
- 公开承认未模型化的循环("更深的共识削减制造终止动态,这里没有建模,在任何更深的削减被提出之前,建模是欠着的")
- 明确标注哪些数字是未验证的快照(
[re-check])、哪些是刻意留空([needs precise spec])、哪些是暂定值([strawman]) - 给出可证伪的条件("两年内到不了 80M FIL 就证伪 uptake 案例")
- 给出压力测试的坏结果(40M 情景下付 97% 成本却永不激活)
- 在 stall 论证结尾明确说"本节不做更多声称"
这提高了我的信任度,但不改变结论——诚实的设计不等于正确的设计。作者把不确定性摊开给你看,恰恰说明这些不确定性是真实的、未解决的。
8. 怎么判断它会成还是会败(可证伪指标清单)
作者给了一个证伪条件("两年内到不了 80M FIL"),但那来得太晚。下表是我补充的、可以更早给出信号的追踪清单。
8.1 制度建设阶段(现在 → 提案是否进入正式 FIP)
| 观察项 | 指向成功 | 指向失败 |
|---|---|---|
| 资金来源是否定案 | 明确采用"纯储备金 91M" | 转向 FIP-0118 切分 → 矿工应该强烈反对 |
| 委员会身份模型是否定案 | 明确"烧谁的",且委托者风险有界 | 持续悬置 → 无法评估委托风险,产品无法成型 |
| 是否新增 inactivity leak | 若加上,stall 否决权被大幅削弱(安全性论证变强) | 不加 → stall 成本维持在 0.08%/周,否决权长期存在 |
N_min/max_share 是否给出具体值 |
且基于实际签名控制权分析(而非地址数) | 基于地址数 → 分散度门槛可被廉价绕过 |
| 是否补上"双花价值 vs 26.7M FIL"的对比 | 补上且结论为正 | 不补 → 安全性论证有一个未闭合的缺口 |
| 每证书 gas 是否测量 | 在可接受范围,架构不变 | 过高 → 需要改用客户端验证,架构变更、依赖 FIP-0107 |
| "两类质押"是否被重新提起 | 若提起且定出锁定权重上限 → SP 重大利好 | 若提起但上限未定价 → 安全性被削弱且无人负责 |
8.2 部署后(阶段 1:质押开放)
先记住现实基线(2026-09,来自独立核实):现有链上生息产品的实测基数约 40M FIL(其中 GLIF 32.9M),2024 年见顶、年流失约 10%、GLIF 正在向非 Filecoin 代币多元化。所以"新流入"几乎必须从零开始创造,而不是从存量里搬。
| 观察项 | 阈值 | 说明 |
|---|---|---|
| 月均质押流入 | 5M FIL/月 = 模型假设 | 16 个月到 80M。这是最重要的早期信号 |
| 3 个月累计质押 | ≥15M FIL → 模型路径上 | <5M FIL → 强烈暗示落在 40M 悲观情景 |
| 6 个月累计质押 | ≥30M FIL → 模型路径上 | 见上 |
| 新池利率 vs GLIF 利率 | 新池利率 ≥ GLIF 的 5.583% → 有竞争力 | 若新池利率显著低于 GLIF 且招募不达标 → 说明"无信用风险"的溢价不足以抵消利率差 |
| GLIF 的 TVL 变化 | 上升 → 资金在进场 | 继续下降或继续多元化 → 最强的不利信号(GLIF 是唯一有规模的现有产品,也是最好的需求代理指标) |
| 是否有大型委托运营方出现 | 出现 2-3 家有规模的运营方 | 迟迟没有 → 招募可能真的做不到 |
| 质押是否以"自持"为主 | 分散的持有者参与 | 集中在少数巨鲸 → 分散度条件可能过不了 |
| 锁定期最终定多长 | 接近 90 天 strawman | 若显著长于 90 天 → 会劝退一部分潜在参与者(Core Devs 94 上已有人反馈"6 个月也被认为过长") |
8.3 阶段 2-3(委员会运行 → 激活)
| 观察项 | 指向成功 | 指向失败 |
|---|---|---|
| 证书是否稳定产出 | 连续产出,无长间隔 | 频繁中断 → 说明激励仍不足,重演 F3 的失败 |
| 每证书 gas 实测值 | 符合预期 | 远超预期 → 架构回退 |
| 是否有故障证明被提交 | 少量且孤立(说明机制在工作) | 大量 → 罚没参数需要调整;完全没有 → 机制的威慑未被检验 |
| 质押深度是否维持 80M 以上连续 3 个月 | 是 → latch 打开 → 证书 binding | 反复跌破 → latch 永不打开 |
| 激活时的一次性标签跳跃 | 集成方正确处理了 ~900 epoch 的跳跃 | 有集成方出事 → 说明披露不足 |
| 分散度统计 | 持续发布,作为被监控信号 | 若决定不发布 → 集中度风险失去可见性 |
8.4 长期(激活后)
| 观察项 | 说明 |
|---|---|
| 质押集中度(Nakamoto 系数) | 今天是 7(2023-01 为 84)。持续下降到 3-4 → 分散度要求形同虚设 |
| 公布利率是否如设计般下降 | 若在深度足够时仍高于 5% → 说明资金池在补贴而非定价 |
| 是否有 stall 事件 | 若发生 → 触发 inactivity leak 的重新审视条件(作者的触发条件是"stall 真被行使"或"溢价跌破 0.25pp") |
| 存储算力是否因为质押而加速下降 | 若是 → 提案的"不损害存储"承诺被证伪 |
| 第 3 年未激活复核的结论 | 若结论是"清盘" → 整件事的失败被正式确认 |
9. 未决项 → 阻塞关系一览
Appendix A 列了 11 个"决定"类开放项。下表按"阻塞谁"重新组织——这比按章节排序有用得多。
| # | 未决项 | 阻塞谁 | 为什么重要 |
|---|---|---|---|
| 1 | 委员会身份模型(worker key / 专用密钥 / 委托运营方密钥) | 所有人 | 下游未决项最多。决定罚没烧谁的质押、权力表是每运营方一条还是每存款人一条。定案前,运营方写不了罚没披露,托管方写不了产品条款,委托者算不出风险 |
| 2 | 每证书 gas 成本(未测量) | 架构 | 过高 → 验证移到客户端 + 系统证明摘要 → 改变架构,依赖 FIP-0107 |
| 3 | N_min / max_share + 度量方式 |
集中度分析(链下) | 决定"分散"的定义;一个成功的运营方在激活前就被设了份额上限 |
| 4 | 故障证明格式与提交窗口 | 解绑延迟的取值 | 窗口变长 → 解绑延迟必须更长 → 影响所有产品条款 |
| 5 | 基础与相关性罚没常量 | 罚没披露 / 风险定价 | 决定实际风险敞口(当前 strawman 1–5%) |
| 6 | α 在公布利率之内还是之上 |
报价基准 | 决定你看到的"净收益率"是什么意思 |
| 7 | 消息容器与批处理方式 | 工程边界 | 受 64 KiB 上限约束 |
| 8 | 公布利率是否追踪链上 pledge 收益率 | 非续约通道的响应性 | 若 pledge 收益率下降,利率曲线不会自动调整 → 可能同时挤压矿工两侧 |
| 9 | stake_activation 最终值 |
激活时点 | 80M 是提议值,最终值开放 |
| 10 | W / W′ 计 epoch 还是实例 |
规范精度 | 实例是自然单位(分布统计按委员会算) |
| 11 | inactivity leak | 明确不包含 | 重新审视的触发条件:stall 真被行使,或所需回报溢价跌破 0.25 个百分点 |
| 12 | 补充数据承诺的编码方式 | 只持证书的验证器 | 这是它们读取权威标记的唯一途径 |
| 13 | 激活后是否继续发布分布统计 | 集中度监控 | FIP 主张继续发布(作为无自动后果的监控信号),但未定 |
一句话总结:最应该盯着的是第 1 项。 它决定了这个提案对**最大的一群潜在参与者(委托质押者)**到底是"低门槛收益"还是"承担了不明风险"。
10. 附录
附录 A:术语表
| 术语 | 含义 |
|---|---|
| F3 / GPBFT | Filecoin Fast Finality,基于 Group-based Practical Byzantine Fault Tolerance 的快速最终性小工具(FIP-0086) |
| QAP | Quality-Adjusted Power,质量调整算力。扇区的原始大小按其质量倍数加权后的算力 |
| 证书(Certificate) | ≥⅔ 委员会权力对某个 GPBFT instance 的签名聚合证明 |
| instance | GPBFT 的一次运行实例,用 CommitteeLookback = 10 个实例的历史权力表 |
| Binding | 证书具备权威性,客户端可上报数十秒级最终性 |
| Advisory | 证书已产出但不具备权威性,网络按 900 epoch 规则上报。仍然给签名者付费 |
| Halted | 委员会无法定人数,证书停止产出。不给钱 |
| Equivocation | 双签——对同一 instance 签署冲突消息。可被罚没 |
| Correlation-scaled slashing | 罚没金额随同一窗口内作恶的质押比例上升,逼近 ⅓ 时到达约 100% 销毁 |
| Sticky latch | 激活开关的不对称设计:连续满足 3 个月才打开,连续不满足 12 个月才关闭 |
| at-will stake | 可随时取回的质押(只有解绑延迟,无固定期限) |
| pledge | 存储扇区的抵押品,绑定到扇区,到期全额返还 |
| unbonding delay | 解绑延迟,strawman 90 天 |
| liquid float | 流动浮增 = 流通量 − 质押 FIL。报告中新增的一行 |
[strawman] |
作者标记的暂定值 |
[needs precise spec] |
作者标记的刻意留空、需要精确规格的部分 |
[re-check] |
作者标记的需在提交前核实的链上快照/引用/建模输入 |
附录 B:三种状态速查
| 状态 | Latch | 法定人数 | 客户端上报最终性 | 奖励 | 触发条件 |
|---|---|---|---|---|---|
| Binding | 打开 | 活跃 | 证书 head,数十秒 | ✅ | 深度 + 分散度连续满足 W(3 个月) |
| Advisory | 关闭 | 活跃 | 900 epoch,~7.5 小时 | ✅ | 激活前,或去激活后 |
| Halted | — | 无 | 900 epoch,~7.5 小时 | ❌ | 委员会失去法定人数 |
附录 C:与既有方案对比
| 方案 | 安全基础 | 成本 | 修复 F3 | 创造代币需求 | 引入 PoS 组件 |
|---|---|---|---|---|---|
| 现状 | QAP(~20% 容忍度) | 0 | ❌ | ❌ | — |
| #1106(区块奖励切分) | QAP(~33% 容忍度) | 约 10% 区块奖励,SP 内部转移 | ✅ | ❌ | 部分 |
| Vote escrow | QAP | 低 | ❌ | ✅(锁定) | ❌ |
| 本提案 | 质押 FIL | 约 100M FIL | ✅ | ✅(原生收益) | 4 个组件 |
附录 D:来源清单
一手材料
| 材料 | 链接 |
|---|---|
| FIP 草案讨论帖 | #1288 |
| 渲染版 + 配套分析入口 | https://hannahhoward.github.io/stake-finality-fip |
| 资金来源选项与推荐 | .../funding-options-and-recommendation.html |
| 奖励支付说明 + FAQ | .../funding-explainer.html |
| 纯存储算力攻击分析 | .../storage-only-attacks.html |
| 两类质押(探索性) | .../two-class-stake.html |
| 链下运营化要求 | .../beyond-the-chain.html |
| Core Devs 94 演示录像 | https://youtu.be/it1ft03pI4A |
本地留存(原始抓取,便于核对)
| 文件 | 内容 |
|---|---|
.context/raw/fip1288.json |
讨论帖 API 原始响应 |
.context/raw/body.md |
FIP 正文全文(64,570 字符) |
.context/raw/comments.json |
6 条评论(5 条作者 + 1 条 @ThatGuyLLC) |
元信息
| 项目 | 值 |
|---|---|
| 作者 | Hannah Howard (@hannahhoward) |
| 类型 | Technical, Core |
| 依赖 | FIP-0086, FIP-0100, FIP-0105, FIP-0115 |
| 创建 | 2026-07-21 |
| 发布讨论 | 2026-09-10 |
| 最后更新 | 2026-09-14 |
| 状态 | Draft,开放讨论 |
| 版权 | CC0 |
附录 E:提案数据的独立核实汇总
核实方法:Filecoin Data Portal 日频数据集(daily_network_metrics.json,2020-08-24 → 2026-09-15,2,214 条记录)、Filfox 实时页面、官方 FIP 仓库原文、Lotus / go-f3 GitHub issues、交易所与第三方快照。
| # | 草案说法 | 核实结果 | 判定 |
|---|---|---|---|
| 1 | "最近 30 天下滑年化 −62%/年" | 基准日 2026-08-23 复算 −61.9% ✅ 准确;但 2026-09-15 已降至 −32.9% | |
| 2 | "QAP 2029–2032 跌破 2.8 EiB" | 当前 QAP 11.996 EiB,2.8 是投影值(需再跌 77%)。按已观测衰减率外推落在 2028–2030,与草案区间相容 | 🟡 逻辑自洽,无第三方来源 |
| 3 | "全网 pledge 65.3M FIL 且在下降" | 2026-08-23 实测 65.28M ✅;2026-09-15 64.03M,仍在跌 | ✅ 准确 |
| 4 | "2027 Q1 pledge 回到 ~70M FIL" | 🟡 需注明条件 | |
| 5 | "FIL 价格 $0.77" | 2026-08-23 日均 $0.7609 ✅;9 月中旬已反弹到 $0.79–0.87 | |
| 6 | "CoinGecko 流通量 795M(2026-07-05)" | 实测隐含聚合方供应 795.32M ✅ | ✅ 准确 |
| 7 | "协议 903.13M vs 聚合方 819.16M(2026-08-13)" | 实测 903.37M vs 819.18M,epoch 日期亦吻合 ✅。缺口稳定在 84–89M FIL | ✅ 准确 |
| 8 | "GLIF 约 32M FIL" | 实测 32.9M FIL ✅ | ✅ 准确 |
| 9 | "3,590 个 iFIL 持有者" | ❌ 无法独立核实(第三方给出的是 6,459–6,493 存款人,口径不同) | |
| 10 | "GLIF 收 15% 借出、付储户 3–4%" | ||
| 11 | "约 40M FIL 已接受链上收益" | ||
| 12 | "F3 停止出证书 / 快速最终性已离线" | ❌ 无官方独立确认(README 仍写 "Live on Mainnet",无事故公告)。但旁证极强:5 次主网 SEV3 停摆(lotus #13094 / #13208 / #13359 / #13473 / #13544 至今 open,2026-03-07 起),根因是参与率长期贴着 66.6% 阈值 | |
| 13 | "网络当前实际最终性 ~900 epoch / 7.5 小时" | ✅ 与 FIP-0086 规定及 Lotus 的 900-epoch fallback 实现一致 | ✅ 准确 |
| 14 | "Nakamoto 系数 ⅓ 阈值处为 7,2023-01 为 84" | ❌ 未查到任何独立来源 | |
| 15 | "前十大所有者获得 38.5% 区块奖励" | ❌ 未查到任何独立来源 | |
| 16 | "FIP-0118 的区块奖励时间表"(15.2/9.5/…) | ❌ FIP-0118 通篇没有年度释放表。那 6 个数字是 #1288 自己用 FIP-0118 权重机制推算的 | 🟡 出处需更正 |
| 17 | "BLS12-381 运算已由 FIP-0105 提供,无需新 syscall" | ||
| 18 | (草案未提)提案成熟度 | 🔴 成熟度需大幅下调 |
数据口径提醒:本报告中"算力"如无特别说明指 QAP(质量调整算力)。社区常把 QAP 叫"总算力"、把 RBP 叫"毛算力",两者相差约 8.5–9 倍(因为 Fil+ 扇区按 10× 计权)。看到"12 EiB"和"1.4 EiB"同时出现时不要困惑。
附录 F:相关 FIP 背景速查
| FIP | 标题 | 状态 | 与本提案的关系 |
|---|---|---|---|
| FIP-0086 | Fast Finality in Filecoin (F3) | Final(主网 2025-04-29 激活,epoch 4920480) | 本提案修改的就是它。定义了 GPBFT 委员会、PowerTableLookback = 10、⅔ 门槛、证书格式、10 TiB 资格门槛。明文决定不给参与奖励——这是 F3 失败的直接原因 |
| FIP-0118 | (Solstice) Deprecate Fil+ and Fund Services via Block-Reward Split | Accepted(非 Final,尚未上线;目标 nv29 ≈ 2026-10-08 当周) | 提案的资金备选来源。废除 Fil+ datacap,所有新扇区自动 10× QAP;区块奖励拆成 w1 consensus / w2 service / w0 burn 三流,w1 从 95% 在 9 个季度内降到 50% 地板 |
| FIP-0081 | Introduce lower bound for sector initial pledge | Final(NV24 上线) | 解释了 pledge 为什么下滑。规定 pledge 下限为穿越前总量的 30%(渐近约为流通供给的 9%)。注意:32 FIL/QAP-TiB 的上限不是它引入的,是 miner actor 既有常量 |
| FIP-0098 | Simplify termination fee to fixed % of initial pledge | Final | 提前终止费 = 成熟扇区 initial pledge 的 8.5%(提案引用正确)。用于论证质押不会抽走存储资金 |
| FIP-0100 | Removing Batch Balancer… | Final | 提案列为 Requires,但正文从未解释为什么 |
| FIP-0105 | BLS12-381 Precompiles for FEVM (EIP-2537) | Final | 提供 7 个 precompile(G1ADD/G1MSM/G2ADD/G2MSM/PAIRING_CHECK/MAP_FP_TO_G1/MAP_FP2_TO_G2)。提案称链上验证证书靠它。 |
| FIP-0107 | Implicit Messages in Block Receipts | Accepted | 只有备用架构(客户端验证)才依赖它。主路径不依赖 |
| FIP-0115 | Premium Percentile Base Fee Target | Final | 提案用它校准 includer 分成 α 的费率 |
| Discussion #1106 | Incentivisation of F3(Kubuxu,2025-01-23) | 仍是 open discussion,从未成为 FIP(被打上 Quiet 标签,最后活动 2025-02-14) |
本提案最直接的竞争对手/替代方案。用最多 10% 区块奖励激励 F3 参与,SP 总奖励不变(纯再分配)。草案的 α ≈ 10% 就是从它借来的。其草案 FIP 正文的 Specification 部分从未写完(全是 "TODO @kuba") |
关于 #1106 的一个重要语境:#1106 的讨论期(2025-01 → 02)里完全没有人提到 quorum 问题——因为 F3 当时还没激活(2025-04-29 才上线)。当时 passive testing 报告的参与率是"77%–79% 的网络算力",看起来很安全。这解释了为什么"不给奖励"在上线时看起来可以接受:核心开发者当时没有预见到无激励会导致 quorum 崩溃。
文档生成于 2026-09-17。所有对提案内容的转述基于 FIP 正文与作者配套分析的原文;标注为"我的分析/判断/建议"的部分是解读,不是原文观点。提案数据已尽最大努力用独立来源核实,核实结果见附录 E,未能核实处已标注。
|
Is there a high-enough network benefit to spend/dilute $70,000,000 to:
If yes, than less money could buy a much tighter finality, ex: Avalanche finality is < 2 SECONDS, so it follows that the benefit is primarily in FIL locking and token-holder participation while PDP grows. |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Draft FIP · Author: Hannah Howard (@hannahhoward) · Type: Technical, Core · Requires: FIP-0086, FIP-0100, FIP-0105, FIP-0115 · Created: 2026-07-21 · Rendered copy and supporting analyses: https://hannahhoward.github.io/stake-finality-fip
How to read this. Four questions, in order.
Supporting analyses are linked in the comments below.
The participation reward requires a funded pool, but this thread leaves the source open to test the mechanism first. Funding the finality reward: options and a recommendation covers the options and is not the subject here. If the mechanism has consensus, funding is the next question; §5 states the requirements, and activation requires an accepted funding source.
Simple Summary
This proposal secures the token. The deeper fixes to Filecoin's storage economics will take years, and FIL needs a durable security and holding incentive in the meantime. The proposal revives Filecoin's fast-finality gadget (F3) by deriving committee voting weight from staked FIL rather than storage power and by paying participants whose signatures appear in finality certificates. The reward comes from a funded pool specified in §5. Leader election and Proof-of-Replication (PoRep) remain unchanged.
Abstract
This proposal secures finality — irreversible chain history — with staked FIL. F3 (FIP-0086) uses a Group-based Practical Byzantine Fault Tolerance (GPBFT) committee weighted by quality-adjusted storage power; that resource also determines leader election through Expected Consensus and PoRep. F3 participation is unrewarded, and the committee has struggled to maintain quorum, leaving the network on the slower 900-epoch Expected-Consensus finality rule.
The mechanism weights the finality committee by staked FIL and rewards signing. Because GPBFT already consumes an abstract power table, it can derive that table from stake instead of storage power. A new staking actor records stake on chain, making GPBFT equivocation slashable against it.
A participant earns only when its signature is included in a certificate. The offered rate declines as stake arrives, letting the market determine equilibrium depth without a rate floor, and depends on the reward pool's released balance. The pool requires about 100M FIL over the program's life; §5 gives the four requirements for any source.
Because a new staking actor starts empty, staking opens at the upgrade epoch, but stake-weighted finality becomes authoritative only after about 80M FIL is staked and sufficiently dispersed, as specified in §6.
This proposal covers finality only. Leader election, block production, and PoRep are unchanged. Stake-weighting leader election — a full consensus pivot that would retire PoRep — is a separate and substantially larger question and is explicitly out of scope.
Change Motivation
No FIL is natively staked today. Filecoin has no native staking mechanism. The FIL that secures the network is pledge collateral bonded to storage sectors, while liquid FIL earns no protocol-native yield. Consensus therefore depends on storage power, and the storage base is declining. This proposal takes that decline as given.
On that premise, the objective is to strengthen FIL's economic security while giving finality a base independent of future storage-power levels. A protocol-native reward gives holders a reason to stake, creating slashable collateral for finality.
If the premise is rejected, this is the wrong instrument. A committee weighted by a stable resource need not be re-based. Discussion #1106 restores the same two-thirds threshold near a storage provider's marginal signing cost using a slice of block rewards and no added funding (Prior art); vote escrow can create locking without changing consensus. If storage remains a sufficient security foundation, those approaches are cheaper.
Under the premise used here, however, a quorum patch would restore certificates on a security base that continues to shrink, backed by collateral that is returned after exit, and it would not address the proposal's token-demand objective. Re-weighting the committee from storage power to stake is what drives the additional funding requirement quantified in the companion analysis.
FIL holders have already allocated capital voluntarily to credit-exposed yield instruments (see the companion analysis), suggesting demand that a protocol-native instrument could serve. The posted rate opens below a 12% ceiling and declines as stake arrives (§3), settling around 4.0–5.0% while the modeled staked pool remains near the target security level (yield comparables: companion analysis).
Filecoin's lack of staking also leaves finality without an independent, slashable security base. FIP-0086 introduced F3 alongside Expected Consensus: participants sign GPBFT messages, and ≥⅔ of committee power produces a certificate that makes a tipset irreversible within tens of seconds, versus the ~900-epoch (~7.5h) Expected-Consensus rule.
FIP-0086 deliberately did not reward participation, reasoning that signing is cheap and participants benefit from faster finality. In practice, the committee lost quorum and F3 stopped producing certificates, leaving the network on Expected-Consensus fallback finality. This proposal gives finality a slashable stake base and fixes the minimum depth required for authority in protocol rather than accepting the market's storage-power level.
Specification
The mechanism has six parts: a staking actor (§1); stake-weighted F3 power table (§2); certificate-gated reward (§3); on-chain certificate submission (§3a); equivocation slashing (§4); and an activation rule (§6). §5 defines the roughly 100M FIL funding requirement and required source properties; the source remains open.
Only the power-table formula (§2) is a trivial change. The certificate-submission path (§3a), fault-proof machinery (§4), and transition (§6) add genuine protocol surface. Everything outside the actor bundle must also be implemented separately in each node client.
1. Staking actor (new)
A new built-in actor is the on-chain source of truth for staked FIL. It records each participant's balance and BLS signing key; supports deposits, delayed withdrawals, and delegation; and maintains the activation state (§6), including the append-only
ActivationRecordthat determines which certificates bind (Appendix B).The unbonding delay has two lower bounds. It must exceed the fault-proof window (§4), so an equivocator cannot withdraw before slashing, and be economically meaningful: exiting stake remains locked long enough that the pool cannot drain immediately and exit is priced against the posted rate rather than treated as an intraday trade (strawman: 90 days). It is a delay, not a fixed term; all figures here and in the companion analysis assume at-will stake.
[needs precise spec — Appendix A]
2. Stake-weighted F3 power table
GPBFT's ⅔ quorum, BLS12-381/BDN signature aggregation, certificate format and signer bitset, and modified Expected-Consensus fork-choice rule are all agnostic to the source of committee power. At the formula level, the change is therefore a single input substitution. Where FIP-0086 computes
this proposal computes
The staking actor supplies the new value at the same
PowerTableLookback = 10lookback (ten GPBFT instances; the implementation names the constantCommitteeLookback). The 16-bit normalization and certificate machinery are unchanged.The current committee's eligibility filters cannot be reused directly: the 10 TiB power floor, zero fee debt, and no unelapsed consensus fault are miner-actor conditions with no equivalent meaning for a staking actor. Stake-side eligibility therefore consists of a minimum stake plus the staking actor's own fault state (§4).
Two encoding limits bound minimum stake. Below 1/65535 of total stake, a participant normalizes to zero weight; CBOR also caps the power table and each certificate delta at 8192 entries, so the minimum must keep the committee within that limit. Rewards accrue outside the power-table balance (§3), avoiding per-epoch compounding that would place every participant in every certificate delta. This weighting becomes authoritative only after activation (§6).
3. Participation reward, gated on signed certificates
Staked FIL earns a reward only when its signature appears in a valid, canonical finality certificate, for the epochs that certificate newly finalizes. Repeating an already-final head earns nothing. Epochs advance every 30 seconds regardless of committee behavior, so faster certificate production cannot increase payout. Stalling, idle stake, and halted F3 earn nothing.
As in Discussion #1106 (Prior art), rewards split into a signer share and a smaller includer share (
α, around 10%) paid to the on-chain submitter (§3a), in practice the block producer, since only the producer collectingαhas reason to include a submission. One signer's reward does not reduce another's, so censoring a peer gains nothing and refusing submission forfeitsα.The offered rate is a published function of two on-chain values: total staked FIL,
S, and the released balance of the reward pool,P:S_offsetkeeps the rate finite at low stake and controls its decline.Pis only the released ledger; funding enters unreleased and moves to released on a fixed schedule (§5).spend_rateandS_offsetare set together with the funding envelope, so their values depend on the adopted source. The table and curve below use the worked configuration from the companion analysis:spend_rate = 0.22andS_offset = 55M. The launch rate is capped at 12% across all options [strawman].The posted rate declines as stake arrives, so the first 80M FIL is recruited along a declining curve rather than at a single fixed rate. The curve uses the same worked configuration.
The schedule has two further properties:
spend_rate × Pat every stake depth. In the worked scenario, year-1 payout is 6.60M FIL on a 30M released pool. There is no catch-up payment for epochs that pay nothing.If the schedule cannot recruit enough stake to reach the activation threshold (§6), the year-3 non-activation review (§5) is triggered if no certificate has ever become binding. Incentive Considerations analyzes that case.
Lifetime payout is bounded by total pool funding, the payment rule above, and the sunset (§5).
For example, suppose a holder stakes 10,000 FIL in the tenth month of signing, when total network stake is 50M FIL. In the worked scenario, the offered rate is 6.29%, or roughly 630 FIL over the following year, accruing only for epochs in which the holder's signature appears in canonical finality certificates.
When total staked FIL reaches 100M — year 6 on that modeled path, with 34.2M FIL released and unpaid — the same 10,000 FIL earns 4.85%, or about 485 FIL per year. Claims are pull-based. The posted rates above are therefore gross of both the includer share
αand the gas paid by a staker to claim.Rewards begin accruing with the committee's first certificate, including during the pre-activation accumulation phase (§6).
[needs precise spec — Appendix A]
3a. Certificate submission and verification (new)
Today F3 certificates remain off chain, in node-local stores and peer-to-peer protocols. §3 requires the chain to receive and verify them, making submission the proposal's largest new protocol component. Discussion #1106 favored message submission over larger block headers; this proposal follows that direction.
A block producer submits each new certificate as a message to the reward-pool actor, which maintains a per-instance watermark. The first valid submission credits signers and pays
α; duplicates fail cheaply. A conflicting certificate for the same instance is accepted as evidence but earns no reward, providing on-chain evidence of correlated equivocation (§4).Submission is permissionless: any account may submit a certificate. Suppressing a valid certificate therefore requires every producer to forgo
αagainst its own economic interest.The actor verifies certificates directly. Aggregate verification requires a coefficient-weighted multi-scalar multiplication over signers plus one pairing; the required BLS12-381 arithmetic already ships under FIP-0105, so no new syscall is required.
Per-certificate gas cost is unmeasured and an early implementation deliverable. If in-actor verification is prohibitively expensive, the fallback is client-side validation feeding the actor a system-attested summary, which would add a block-validation rule to every client and depend on FIP-0107. This proposal assumes in-actor verification.
Resolving a certificate's signer bitset to staker identities requires the ordered committee table for that certificate's instance. Because the staking actor is the power source, it retains per-instance committee snapshots, and the retention window must cover the fault-proof submission window (§4).
Sections §3, §4, and §6 all depend on the same structures: per-instance committee snapshots; the instance-to-epoch mapping supplied by the certificate record; and the number of epochs each certificate newly finalizes.
A chain message is capped at 64 KiB, while the certificate format permits larger power-table deltas. Submitted deltas must therefore be chunked, or the table must read stake snapshots at a bounded cadence. [needs precise spec — Appendix A]
4. Slashing for equivocation
Double-signing (equivocation) in GPBFT — signing conflicting messages for the same instance — is slashable against the offender's staked FIL. This is the key security distinction from storage power: an attacker's staked capital can be burned, while acquired storage power remains recoverable.
Because GPBFT uses explicit signatures, conflicting signed messages attribute a fault to a key. Today F3 nodes drop conflicts and GPBFT messages are not verified on chain, so slashing needs a fault-proof subsystem: nodes retain and gossip conflicting pairs; anyone may submit one; the actor reconstructs the FIP-0086 payloads and verifies both signatures against the registered key.
Correlated equivocation can use a simpler proof. Two valid certificates for the same instance that decide different chains are themselves evidence that every signer appearing in both bitsets double-signed, and those certificates are already on chain (§3a).
The slash scales with correlation. Isolated equivocation receives a small base slash (strawman: 1–5%, plus committee ejection), because an operator accident must be survivable or the penalty will deter participation as well as attacks. Ejection takes effect at the power-table lookback, ten instances after the fault is recorded, because the tables for in-flight instances are already fixed.
The slash then rises with the fraction of stake that equivocates within the same window, reaching a total burn of approximately 100% as the equivocating fraction approaches the ⅓ required to threaten finality. Because the final scale depends on additional proofs that may arrive later, slash amounts remain in escrow until the window closes and then execute at the final scale.
Escrow matters especially in a mass event: equivocation by a third of the committee resembles both an attack and a partition or defective release. A successful finality attack therefore risks roughly the attack stake itself — about one-third of the pool, or tens of millions of FIL at activation. That amount must exceed the double-spend value protected by fast finality, and the activation band should be re-checked as settlement grows.
A valid fault-proof submitter receives a reporter share, capped by the amount burned, mirroring today's consensus-fault report and paying the gas cost of policing the committee. The remainder is burned; returning it to the pool would raise
Pand the offered rate, letting an attack subsidize other stakers.Slashing is active from the committee's first certificate, before activation (§6). The unbonding delay (§1) must strictly exceed the fault-proof submission window, so an equivocator cannot withdraw before a slash can land. The window is defined over instances and anchored to epochs by the certificate record (§3a).
Under delegation, the signing key belongs to the operator. Which stake burns — the operator's bond, delegators' pro-rata stake, or both in some order — depends on the committee identity model (worker key, dedicated key, or delegated operator key). Appendix A records that choice as open.
[needs precise spec — Appendix A]
5. What the funding has to supply
The reward pool requires about 100M FIL of total funding. In the worked model, that supports roughly twelve years at the 80M FIL security level (§6) and remains a strawman. Less funding buys a shallower pool or fewer years at target depth; more raises total consensus expenditure.
Whatever the funding source, the mechanism requires four things from it:
r(S, P)reads only that balance (§3).StateCirculatingSupplyfails.The funding source is open. This proposal does not select one. The companion analysis compares candidate sources against the four requirements above and recommends one; every numerical illustration here uses that analysis's worked scenario. Adopting this mechanism requires an accepted funding source to be adopted with it.
6. Transition and activation
At upgrade, total staked FIL is approximately zero. F3 needs a non-empty initial power table, and a nearly empty committee is worse than none: one participant alone is a ⅔ quorum. Implementation gives the deployment sequence; this section defines its gates.
Activation therefore uses a stake-depth threshold plus distribution requirements. During accumulation, the protocol pays pre-activation rewards for advisory certificates and incurs the blockspace and verification gas required to carry them on chain (§3a; see also Security Considerations).
Staking opens at the upgrade epoch; the committee starts only after it can form. Stake-weighted F3 begins as a fresh instance chain with a new F3 network name, bootstrap epoch, and pinned initial power table. Those values ship in a coordinated client release once registered stake exceeds a formation floor set well below the activation threshold [needs precise spec].
F3's original mainnet rollout similarly staged code deployment and activation and pinned the initial power table in a follow-up release; the formation-floor trigger is new here. A fresh chain also cleanly separates the QAP-era and stake-era certificate streams: verifiers, bridges, and snapshots built against the old stream reject stake-era certificates rather than mistaking them for authoritative.
A handoff within the existing chain is representable — one certificate could replace the entire power table — but it would have to be signed by two-thirds of the old QAP-weighted committee, which is the quorum this proposal is intended to replace. From the first stake-era certificate, stakers earn the participation reward (§3). Rewarding the accumulation phase is necessary to recruit stake toward the activation threshold.
Certificates remain advisory until both depth and distribution conditions are met. The authority condition is
total_staked ≥ stake_activationand the distribution conditions below. Until then, the network continues to use today's 900-epoch Expected-Consensus rule, and stake-weighted certificates are advisory only. This avoids the thin-but-trusted failure mode in which a cheap-to-acquire quorum becomes authoritative.The staking actor maintains total stake, distribution statistics, and the activation latch. Every client reads the same state at finalization and suppresses advisory certificates from finality reporting. The depth latch resembles the power actor's minimum-miner count; distribution statistics differ because they require sorting the operator set. Once all conditions hold, stake-weighted fast finality becomes authoritative.
Activation requires dispersion, not just depth. A deep pool only makes a one-third coalition expensive if stake is sufficiently dispersed. The gate therefore also requires the following strawman conditions [needs precise spec]: the smallest coalition of distinct participants controlling one-third of stake must contain more than
N_minparticipants, and no single participant may control more thanmax_shareof the pool. Both are measured over delegated operators — the entities that actually sign — rather than raw addresses.These tests are necessary but not sufficient because on-chain identities are cheap to multiply. They prevent accidental concentration from being activated and force a deliberate attacker to manufacture at least the appearance of distribution. The underlying economic backstop remains stake depth plus correlation-scaled slashing (§4).
The latch uses asymmetric open and close conditions, making activation deliberately sticky. It opens only when
total_staked ≥ stake_activationand the distribution conditions hold continuously for a sustained windowW(strawman: three months). A single deep epoch therefore cannot activate finality.The latch closes only when
total_staked < stake_deactivationcontinuously for a much longerW′(strawmen: 50M FIL and twelve months). A distribution failure after activation never closes the latch. Otherwise an adversary could create concentration and gain a shutdown switch without one-third of stake or equivocation, cheaper than stalling. Post-activation concentration is handled by correlation-scaled slashing (§4) and, if necessary, a later FIP.The deactivation band is a safety catch, not a live control. It is wide enough that ordinary exit and refill should not trip it. A higher deactivation floor preserves a higher veto price but ends binding finality sooner; Appendix B sizes that tradeoff.
The latch history is the authority rule. The staking actor's
ActivationRecord(§1) is an append-only list of open and close events. A certificate for instanceiis binding exactly when the record contains an activation opened at or beforeithat has not closed byi. Appendix B defines the record layout and the certificate authority mark used to resolve against it.[needs precise spec — Appendix A]
Prior art
GitHub Discussion #1106, "Incentivisation of F3" (Kubuxu, Jan 2025) proposes allocating up to ~10% of block rewards to F3 participation, using evidence in on-chain finality certificates and splitting rewards between signers and the block producers that include them. It remains an open discussion and was never adopted as a FIP. Its text keeps total storage-provider rewards unchanged, so the proposal redistributes existing rewards among providers.
This proposal borrows #1106's inclusion and anti-gaming structure and its preference for message submission (§3a). Change Motivation explains the substantive difference between the two approaches and the additional cost introduced here.
Design Rationale
The security argument depends on what stands behind the consensus threshold, not merely on where the threshold is set. Acquired storage power can transfer without a protocol fee, its pledge is recoverable collateral, and the penalty for manipulating chain selection through provable double-fork mining is a fixed slash of roughly 4 FIL. By contrast, staked FIL that equivocates is burned at a rate that scales with the equivocating fraction (§4).
At the activation threshold, the two security bases are of similar nominal size but differ in what an attacker can lose: roughly 21.5M FIL of recoverable pledge behind one-third of storage power [re-check] versus 26.7M FIL of burnable stake behind one-third of the pool. The required stake depth is enforced by the activation rule rather than left to the current level of storage power.
This comparison addresses safety, not liveness: a committee can stall without equivocating under either design. Security Considerations analyzes that case. The 26.7M FIL figure is computed from the base and correlation-slash assumptions in §4.
Why finality, and only finality. Finality and leader election can be weighted by different resources. Re-weighting finality is comparatively additive because GPBFT already consumes an abstract power table. Re-weighting leader election would require retiring or replacing PoRep, a much larger undertaking.
This proposal therefore captures the security benefit of slashable, stake-backed finality while leaving the storage market's proof machinery intact. If future PoRep research produces a better leader-election proof, changing that proof can remain a separate proposal while finality continues to use stake.
What this does and does not build toward stake-based consensus. Adopting this proposal is not a decision to move Filecoin to proof-of-stake consensus. It does, however, build four components that such a transition would otherwise need: an on-chain stake registry with deposits, delegation, and unbonding; equivocation slashing with fault proofs and correlation scaling; a pool of staked FIL with known depth and measured distribution; and operating experience with a BFT gadget whose committee power comes from stake.
It does not build stake-based leader election. Using stake to decide block production is the largest missing step in any proof-of-stake pivot and introduces defenses against grinding, nothing-at-stake, and long-range attacks. Those problems do not arise here in the same way because leader election remains PoRep/QAP-weighted and finality remains explicit-signature BFT, where the relevant failure is slashable equivocation rather than costless simulation.
If the network later chooses a broader proof-of-stake transition, this proposal would provide some of the substrate, but leader election would remain the largest unresolved component.
Why a stake threshold rather than a QAP→stake ramp. A blended power table, with storage power fading out as stake fades in, could restore storage-secured fast finality immediately. But fast finality is already offline, so there is no urgent live service to preserve during a transition. A blend would also reintroduce QAP into the finality table and require paying storage signers during the ramp. A threshold is simpler: use pure stake-weighting from the start, but make it authoritative only once the pool is deep and dispersed enough.
Why rewards are gated on certificates. Paying all registered stake regardless of participation would recreate F3's free-rider problem. Requiring inclusion in a valid canonical certificate pays for an actual service — a finality signature over the canonical chain — and is self-limiting: no certificates, no payout.
Alternatives considered and set aside
Discussion #1106's block-reward slice is the cheaper alternative, and Change Motivation explains the choice between it and this proposal. Funding-side alternatives — different sources, envelopes, and designs that meet the same requirement — are covered in the companion analysis.
Why the offered rate has no floor. Ethereum's staking-yield curve remains near 1.5% even at high stake levels, so, in the words of EIP-8363, Tapered Issuance Burn (Draft, created 2026-07-14), there is "no point at which the incentive to stake switches off." Correcting that later can require adding a burn against issuance after products have already formed around the base rate.
Here,
r(S, P)is strictly decreasing in staked FIL, has no floor, and approaches zero. The reward pool is depleted by its own payments, and the mechanism sunsets in year 30 (§5). The offer can therefore wind down by design rather than requiring a later governance decision to turn it off.Backwards Compatibility
Test Cases
[to be expanded] Five cross-client test families, each tied to a specific invariant; the funding side has its own tests:
truncate(0xffff * Stake_i / TotalStake)at the lookback, resolves a certificate's signer bitset to the same participants, accepts a certificate signed by ≥⅔ of stake, and rejects one below ⅔.r(S, P)using the current stake depth and released balance for exactly the epochs that certificate newly finalizes, counted once across overlapping certificates. Non-signers, certificates that finalize no new epochs, replays, and halted periods accrue nothing. Annual payout never exceedsspend_rate × P.W. They stop binding only afterstake_deactivationholds throughoutW′, never solely because a distribution condition fails after activation. Historical instances resolve identically as binding or advisory from theActivationRecordfor both syncing nodes and certificate-only verifiers.Security Considerations
Stake-weighted finality adds a large, slashable security budget. In the companion analysis's two demand scenarios, 80–130M FIL is staked from year two through the first decade, depending on the return stakers require. That capital becomes economically slashable collateral for finality, a security layer that does not exist today. At the proposed 80M FIL activation level, a one-third finality coalition must put roughly 27M FIL of slashable capital at risk.
One way to compare the deterrent with the potential prize is to compare one-third of staked FIL with one year of storage-provider consensus rewards. In the worked scenarios, those series cross in the first year of signing, and the gap widens thereafter. The staked series comes from the worked scenario; the consensus-reward series uses the block-reward schedule adopted in FIP-0118.
The crossing holds at both modeled required returns and occurs before the activation gate opens in year 2 on the modeled path (§6). By the time certificates become binding, the slashable budget is therefore already larger than this comparison's modeled prize. A funding choice that reduces storage-provider consensus rewards widens the gap.
Slashing must be enforceable. Stake is a stronger security base than storage power only if equivocation penalties are real and collectible. Attribution can come from either a conflicting signed-message pair or conflicting certificates (§4). The fault-proof subsystem and final slash schedule therefore determine whether the deterrent holds.
While the pool accumulates, finality remains at today's baseline. No committee runs until registered stake passes the formation floor (§6). After formation, certificates remain advisory until the pool is sufficiently deep and dispersed, so a thin committee is never trusted and block-production liveness is unaffected.
During this period, finality continues to use the Expected-Consensus fallback. Published analysis places its adversary tolerance near ~20% of power, compared with ~33% for a live BFT finality gadget (the ~20% figure is from Wang, Azouvi and Vukolić, AFT 2023; ~33% is the standard BFT bound). The network is in that condition today, and no adopted FIP ends it; this proposal does.
The uncertain variable is duration: how long it takes the stake pool to reach the activation threshold. Both inputs to that estimate — reachable idle FIL and uptake rate — are low-confidence today (Incentive Considerations). The design therefore uses a review date rather than relying on a forecast. If no certificate has ever become binding, the year-3 non-activation review (§5) reports the pool's trailing-year behavior and provides an amendment path.
If reachable idle capital is at least the activation threshold, the modeled uptake rate of 5M FIL per month would reach 80M FIL in sixteen months of uninterrupted inflow. If reachable capital is lower, the gate does not open and the network remains on today's finality rule. Pre-activation rewards still pay for advisory certificates within §3's payout bound, and the chain still bears their blockspace and verification-gas cost (§3a).
The first empirical test is therefore whether staked FIL reaches the 80M FIL security level within two years of the upgrade epoch, as the model assumes. A pool that remains far below that depth would falsify the uptake case.
The addressable-pool estimate uses aggregate circulating-supply figures and revealed-preference comparables from existing yield products; it does not use per-holder data.
A signup rush cannot overspend the pool, and stake depth is self-limiting. Additional stake lowers the offered rate, so annual payout remains below the §3 bound regardless of how quickly stake arrives. At a 5% required return, the schedule falls below 5% once about 77M FIL is staked, so a 5% marginal staker stops adding. It falls below 4% around 110M FIL, where the modeled 5% staker begins leaving. In that case stake peaks around 100M FIL, inside the hold level, and settles rather than overshooting indefinitely.
A pool that stalls below activation keeps the posted rate elevated. With only 40M FIL of reachable capital, the rate remains above the early-termination bar (Incentive Considerations) for 46 months rather than four, another reason for the calendared year-3 review (§5).
Two qualifications apply to these comparisons. First, the bars are denominated in FIL and therefore move with the FIL price. Second, the modeling establishes the direction of feedback from consensus rewards into storage-side returns but does not yet establish its magnitude.
This reduces the network's reliance on PoRep-weighted consensus for fast finality; it does not retire PoRep-weighted consensus. Once certificates bind, the security behind fast finality comes from the staked pool, whose minimum depth is enforced by the activation gate, rather than from a storage-collateral series that has declined over time. The fallback remains the 900-epoch Expected-Consensus rule (Appendix B).
Continued decline in the PoRep layer therefore affects the slower fallback path rather than the stake-backed security of a binding certificate. That is the extent of the claim this proposal makes.
A stall stops finality from advancing. It cannot undo finality already reached, and it opens no new attack. A third of registered stake can decline to sign without equivocating: §2 counts registered stake, not participating stake, so §4 never fires, and while the stall runs the chain falls back to the 900-epoch rule it runs on today. So the finality-tolerance gain stated above holds only while the veto goes unexercised — but three things bound what the veto is worth. First, everything already finalized stays finalized; a certificate cannot be recalled by refusing to sign the next one. Second, a reorg still requires storage power, roughly 20% of quality-adjusted power under the fallback rule, exactly what it requires today and no easier, priced regime by regime in the storage-only attack analysis. So the attack that matters needs both: a third of the stake first, and then the same storage-power attack that is available right now without any first step, because F3 is already stalled by apathy at zero cost. This proposal leaves the first lock exactly as it is and adds a second one in front of it. Third, the second lock cannot be picked quietly: certificates stopping is the alarm, visible to every exchange and bridge within hours, which is when they fall back to slow confirmations and take most of the prize off the table. What remains is the chance to reverse whatever moved in that shrinking window, while everything certified stays out of reach; for a profit-seeker the arithmetic does not close, and for an attacker who only wants damage, the carry and visibility below are the deterrent — this section claims no more than that.
Acquiring and carrying the veto is capital-intensive but relatively cheap to maintain. Across the funded years, obtaining one-third of the resulting pool requires roughly 30–43M FIL, more than one-third of the pre-attack pool because the attacker's own stake increases the denominator. A staller earns no participation reward (§3), so carrying the position forgoes about 1.42M FIL per year, or roughly 0.08% of the position per week.
Those figures follow the modeled path of the released pool and assumed staker return. If the market requires a return above what the schedule can offer, the activation gate never opens and there is no stake-based veto to exercise.
Relative to today's stalled F3 committee, some factors make a future stall harder and others make it easier. The capital requirement rises from roughly 21.5M FIL of recoverable pledge to about 36M FIL bought at market, and the current zero-cost route — simple non-participation — disappears. On the other hand, the carrying cost falls roughly fivefold [re-check], and a single pseudonymous buyer could replace the eight-to-ten identifiable incumbents whose coordination difficulty is an important deterrent today.
A stall is also self-limiting. Nobody earns participation rewards while it continues, honest stake can exit, and the pool can fall toward §6's deactivation floor, at which point binding finality returns to advisory rather than remaining hostage to the stalled committee. The veto price therefore declines with pool depth, reaching 9.1M FIL by year 25 in the modeled path; the deactivation band bounds that decline (Appendix B analyzes a higher floor).
An inactivity leak — burning stake absent from the signer bitset that §3 already exposes — remains out of scope. The model estimates that even a 0.25-percentage-point premium in required return costs two years of the funded security window, while a leak also risks burning honest participants during partitions. A later FIP should revisit an inactivity leak if a stall is actually exercised or if that premium falls below 0.25 percentage points.
Binding finality begins with an explicit dispersion requirement. Certificates cannot become authoritative until the pool is both deep and dispersed (§6). Today's committee has no comparable dispersion requirement; its current reading is a Nakamoto coefficient of 7 at the one-third threshold, down from 84 in January 2023, with the top ten owners receiving 38.5% of block rewards [Filecoin Data Portal series behind the public L1-health dashboard, 2026-07-13; re-check].
Concentration can still increase after activation. Delegation aggregates control, and Lido reached roughly 23% of all staked ETH according to its February-2026 tokenholder update through a similar aggregation path. For that reason,
N_minandmax_shareshould be re-checked against who actually controls signing weight, not address counts that are cheap to split. The participation API described in Implementation provides the required reporting; no client exposes it today.The power gained by concentrating stake is limited because this committee controls finality only. An entity with one-third of stake can stall certificates, producing the fallback behavior priced above. Even an entity with two-thirds of stake cannot produce blocks, censor transactions, rewrite state by itself, or mint FIL, because leader election remains storage-power weighted.
The additional capability at two-thirds is to combine captured finality with a storage-power attack and certify a reorg. That compound attack requires roughly twice the veto's stake capital plus the separate storage-power position, all while the stake concentration is visible in an on-chain table. Concentrated stake therefore controls less of the protocol than concentrated storage power does today, because storage power elects every block.
Two important limits remain. First, the dispersion rules constrain participants, not jurisdictions. Delegated stake is likely to concentrate partly in exchanges and custodians, including regulated companies subject to government pressure. This proposal does not measure or solve that risk; it limits the consequence because compelled custodians can at most stall fast finality back to the network's current fallback without also acquiring storage power.
Second, the automatic remedy is intentionally one-way. A concentration finding after activation does not close the latch (§6); correcting it requires a FIP and network upgrade. The alternative — allowing concentration itself to disable finality — would create a cheaper shutdown lever. Dispersion rules therefore make accidental concentration harder to activate, while pool depth and correlation-scaled slashing (§4) provide the economic defense against actual concentrated control.
Incentive Considerations
Uptake is the largest open variable because the schedule buys security by recruiting stake. If the marginal staker requires a higher return than the schedule offers, the pool may recruit and pay participants without ever crossing the activation threshold, so no certificate becomes binding (§6). The year-3 non-activation review is designed for exactly that outcome and therefore keys on activation rather than pool balance (§5). The modeled scenarios assume 130M FIL of reachable idle capital; the companion analysis provides the comparables behind that figure and the demand cases built on it.
The effect on storage providers depends on the funding source, which remains open. The companion analysis allocates each option's costs across constituencies. Under every option, storage providers can benefit in two ways: they may earn the participation reward as FIL holders, and their block producers may earn the includer share for submitting certificates (§3a).
Every option also shares the same review-and-sunset framework: a year-25 review whose stated default is to stop payment, followed by a hard year-30 sunset if payment continues. Schedule constants can be repriced only through a FIP and network upgrade (Governance).
Effect on storage-committed capital. The reward is intended to recruit idle FIL — capital that earns no protocol-native yield today — and the reachable pool is estimated from that idle float. Available evidence suggests that storage-power trends are driven primarily by storage-hardware and deal economics, making participation yield a secondary factor.
The incentive can nevertheless pull capital out of storage through two channels with very different costs:
The declining schedule is the primary mitigation and a constraint that the final constants must continue to satisfy. By the time the stake pool is deep, the marginal staking yield falls to about 4.3–4.9% at the security level, below the modeled return on committed pledge at prices above spot (9.5% at $1.00 and 19.6% at $2.00; see the companion analysis). The high initial rate is also capacity-limited because idle capital can enter more quickly than storage-committed capital can unwind.
One limitation remains:
r(S, P)responds to stake depth, not to the return on storage pledge. If pledge yield falls because of the adopted block-reward schedule or because a funding choice affects it, the staking curve does not automatically adjust. [needs precise spec] whether the posted rate should be linked to on-chain pledge yield or whether depth pricing alone is sufficient, with the rationale stated explicitly.Delegation is not gaming. A GLIF-style operator that stakes on behalf of depositors performs the signing work and shares rewards with those depositors; that is intended behavior. The mechanism prevents two different behaviors: earning on idle, non-signing balances through the certificate gate, and equivocation through slashing.
Curve gaming is bounded. A large staker could hold back deposits to keep the offered rate higher. The cost is the reward forgone on the withheld stake, while the benefit is limited by the slope of the curve; any unused capacity is also available to another participant at the same posted rate. Splitting one economic position across identities does not increase rewards because the rate is uniform.
Waiting is not unambiguously advantageous. Rewards paid to others reduce
P, scheduled releases increase it, and new stake increasesS; in the modeled range the depth term dominates. Stake arriving when the pool is shallow can be offered as much as 12%, while stake arriving around 80M FIL is offered about 4.9%.Product Considerations
Fast confirmation is a feature with a funded lifetime, and integrators can observe its state. An exchange or bridge that treats a binding certificate as settlement can reduce confirmation time from ~7.5 hours to tens of seconds. That guarantee is funded rather than perpetual. In the companion analysis's modeled scenarios, certificates become binding in years 2–4 and the pool remains near the target security level for roughly 12–14 years, with a first review in year 3 and a year-30 sunset.
Outside the binding state, the network uses today's finality rule (Appendix B). Integrators can therefore engineer the fallback they already use today and determine from chain state whether fast finality is authoritative.
Reaching the stake threshold also requires off-chain software and recruitment. Custodians and exchanges need a staked-and-accruing balance concept, pull-based claims, and withdrawal terms that reflect the unbonding delay (Backwards Compatibility). Delegated pools need operator software for BLS key registration and rotation, certificate-inclusion monitoring, reward distribution to depositors, and disclosure of slashing exposure. Wallets need deposit, key-registration, withdrawal, and claim flows for the staking actor.
All of these depend on the participation API described in Implementation, which resolves a certificate's signer bitset into per-participant inclusion.
Nothing in the protocol commits any party to build that software or recruit stake. Likely implementers include operators already running FIL yield products and exchanges or custodians already holding FIL. Recruitment may be the larger challenge: crossing the threshold requires convincing holders of tens of millions of FIL to stake, which depends on outreach and distribution outside the protocol.
If software or recruitment arrives slowly, the consequence is delayed activation rather than a weaker authoritative committee: certificates remain advisory until the threshold is met (§6).
Implementation
Detailed design is out of scope for this discussion. The actor bundle is the only implementation deliverable shared across node clients: it contains the staking actor, reward pool, certificate-submission and verification path, and slashing logic. Power-table derivation, finalization gating, the participation API, supply accounting, staker-facing tooling, and migration are implemented separately in each client, in two languages, and kept consistent through shared conformance vectors. Block-producer software must also submit certificates.
The participation API maps each certificate's signer bitset to per-participant inclusion. Reward verification, delegated-pool operators, and concentration monitoring in Security Considerations all consume that same feed. Only protocol constants depend on the adopted funding source; the implementation path does not.
Deployment sequence. The stages occur in this order and are separately observable:
Each stage depends only on components exercised in an earlier stage. Deployment risk is therefore separated from committee-operation risk, and committee-operation risk from binding-finality risk. The later stages have no fixed date, so this is an ordering rather than a calendar.
This proposal also amends FIP-0086's committee-definition text for eligibility, power source, and signing key. That amendment should also resolve FIP-0086's quality-adjusted-versus-raw-byte divergence.
Governance
Upgradability and parameter governance. The staking actor is built into the system actor bundle and can change only through a network upgrade adopted by node operators through the FIP process. There is no admin key, multisig, or unilateral upgrade path.
The mechanism's parameters — reward-schedule constants, activation and distribution thresholds, slash constants, and any constants fixed with the funding source — are protocol constants and can be amended only through the same process. This proposal deliberately avoids a runtime-governed tuning actor for consensus-critical parameters because hot-tunable finality economics would create an additional attack surface. If the network prefers runtime governance for these values, that should be evaluated separately with its own security review.
Appendix A — the open items, by class
Every
[needs precise spec]marker points to this appendix.[re-check]marks a citation, chain snapshot, or modeling input that must be verified before FIP submission. The table below lists decisions: open items that can change whether the design works, what it costs, or what risks it tolerates. Other unresolved details are treated as refinement bounded by the rules already stated.αis included in the posted rate or paid on topαagainst the FIP-0115 base-fee rulestake_activationN_min,max_share, and measurement over the operator setstake_deactivation,W,W′For activation, the open questions are primarily parameter values rather than the mechanism itself. The hysteresis rule (§6), authority record and readers (Appendix B), and stall behavior (Security Considerations) are specified; the remaining choices are the constants listed above.
Funding-specific open items belong to the separate funding decision, most importantly the source itself.
Appendix B — the activation state
The three states. The network is always in one of the following states. Block production and chain ordering are identical in all three:
All three states retain today's Expected-Consensus fallback, whose published analysis places adversary tolerance near ~20% of power (Security Considerations). That fallback is the baseline contract for integrators. The certificate authority mark below tells an integrator whether a certificate is currently binding or advisory.
The deactivation band, sized. The deactivation floor in §6 trades recovery tolerance against the declining price of a stall. A 60M FIL floor would keep the veto price above roughly 20M FIL but end binding finality around year 17. The 50M FIL strawman permits a wider operating band at the cost of a lower guaranteed veto price.
The floor is intended as a safety catch, not a live control. In the companion model, 50M FIL sits roughly 30M below the depth maintained at the 5.5% required-return case. After an injected exit of 60% from a full pool, the model refills to the new equilibrium within seven to nine months. The pool falls below 50M for one month at a 5% required return and three months at 5.5%. A three-month
W′is therefore close to triggering during a recovery the schedule can complete on its own, whereas a twelve-month window is not. The refill does not restore the pool's former margin above the activation threshold.[needs precise spec] whether
WandW′count epochs or F3 instances. Instances are the natural unit because distribution statistics are computed per committee. Also open is whether the staking actor should continue publishing distribution statistics after activation as a monitored signal with no automatic effect; this proposal assumes that it should.The
ActivationRecord. The staking actor stores an append-only ordered list of(opened_at_epoch, opened_at_instance, closed_at_epoch | null, closed_at_instance | null)entries together with the current latch state. The same consensus rule evaluated by every client at finalization writes the record. The normative binding rule is stated in §6: an instance is binding if an activation entry opened at or before that instance and had not closed by it. This appendix defines the record layout and its readers.The same rule serves three readers. A node tracking chain state reads the record directly; a syncing node uses the same record to reconstruct whether the gate was open for a historical instance. A verifier that holds only certificates instead reads the latch state and active entry from the committee's supplemental-data commitment.
That certificate-only path has an important limitation: the committee is attesting to its own authority. The mark is reliable only after the committee is already authoritative, which is what activation establishes; by itself, it cannot prevent a pre-activation committee from claiming authority. An integrator could pin a published activation epoch, analogous to a network-upgrade height, rather than check the on-chain record for every certificate. Because §6 also permits later deactivation, however, that integrator still needs a trusted way to learn closure events.
[needs precise spec] the commitment encoding and whether a certificate-only verifier must check the mark against the on-chain record once at first use before trusting subsequent marks.
Copyright
Copyright and related rights waived via CC0.
All reactions