Docs Protocol
Guarantees
What New Constellation proves, under which model, and the Fault Lab scenario that shows each property on a running cluster.
New Constellation treats Alpenglow almost as a black box and proves four properties on top of it. Each one has a scenario in the Fault Lab that runs the attack on a sandbox cluster and records the protocol's answer.
The model
| Assumption | |
|---|---|
| (M) Model | Byzantine stake is below 20%. After GST, at least 60% of stake is correct and online, messages between correct nodes take at most δglobal, and messages from a correct proposer reach its co-proposers within δlocal. |
| (D) Determinism | Merging, ordering and execution are deterministic. |
| (T) Timing | Δtimeout ≥ 3δglobal and 2δlocal ≤ δglobal. After GST, a correct proposer's block reaches every correct validator through Rotor within 2δglobal of completion, and a correct validator executes a slot within δblock once it holds the slot's blocks, its attestations and the state it builds on. |
| (AS) Alpenglow safety | Correct nodes following Alpenglow's voting rules finalize a single chain, for every block-validity predicate and whenever correct nodes cast their first votes. |
| (AL) Alpenglow liveness | After GST, if every correct validator votes notarize or skip in every slot, every slot gets a notar-fallback or skip certificate; if the same valid block reaches all correct nodes before the timeouts, it is finalized. |
Nodes need no synchronized clocks. Every timeout, the 2δlocal hash exchange included, is started by a local event and measured by the node's own clock; clocks need only run at about the same rate. Safety holds with byzantine stake below 20%, liveness needs 60% correct and online stake, fast finalization 80%.
Theorem 2.1: correctness
New Constellation is safe, and every certified state hash is correct. After GST, every window obtains ParentReady. A cblock is finalized in every consensus slot of every window that starts after GST, has more than k/2 correct proposers while all its other proposers are crashed, and whose parent every correct validator has executed when it emits ParentReady for the window.
Why. Validators run Alpenglow's voting rules on h(c_i), which determines the cblock, its parent, its agreed sets, its attested commitments and its state hash, and every commitment binds its block's ciphertext and key. A correct validator emits at most one Block event per consensus slot, so Alpenglow's safety gives one finalized chain of cblocks, bound to one set of payloads. Each certificate holds 60% of stake and so a correct vote, and a correct vote carries the state a correct validator computed, so every certified state hash is correct.
On Acrux Net. With three correct proposers, every cblock of every window finalizes. With one proposer crashed from the window's start, every cblock still finalizes, on the other two proposers' blocks. Shown by the Crash scenario.
Lemma 2.2: selective-censorship resistance
In a window that starts after GST, in every slot, either the blocks of all correct proposers are part of the slot, or its consensus slot is skipped.
Why. A correct proposer's attestation for the window's first slot reaches every correct validator by the attestation deadline. A validator votes for a cblock of a window only after voting for its first cblock, and it votes for the first only once it holds every proposer's attestation for the first slot or the deadline has passed; either way it holds the correct proposer's attestation, which attests that proposer's own block. So no correct validator notarizes a cblock missing a correct proposer's block, and such a cblock can only be skipped.
In practice. A transaction needs one non-censoring proposer among those it was sent to. Two proposers that drop it cannot exclude the third proposer's block, which carries it. Shown by the Censorship scenario.
Lemma 2.3: proposal hiding
A malicious proposer cannot build a block from any information in a correct proposer's payload and have both blocks in the same cblock.
Why. Until a proposer receives the correct proposer's attestation, it holds at most that block's ciphertext. The key is released only in that attestation, which already names the attacker's commitment as on time or marks it late. In the first case, the attacker's block was fixed before the victim's payload was readable; in the second, the attacker's block is not accepted alongside the victim's. The commitment binds both ciphertext and key, so the attacker cannot open it to a different block later.
In practice. No proposer gets the last look. A proposer that waits to see the others is marked late and its block is dropped. Shown by the Last look scenario. Transaction-level hiding, across the proposers a single transaction is sent to, is in Threshold transactions and the Private front-run scenario.
Lemma 2.4: accountable equivocation
If two correct nodes vote notarize for different cblocks of one consensus slot, there is a proof of equivocation blaming at least one proposer. A correct proposer is never blamed.
Why. Different cblocks need different xblocks for some slot, each supported by unanimous attestations of more than k/2 proposers. Two strict majorities of the same k proposers share at least one member, and that proposer signed two different attestations for the same slot. A correct proposer signs one attestation per slot.
The strict majority is what makes this hold. With k = 2 and one attestation allowed to decide, two honest proposers that each miss the other's commitment would each attest only their own block; validators that see one or the other would vote for different cblocks, and nobody could be blamed.
In practice. Two blocks for one slot, or two attestations, are a slashing proof. The slot is not agreed, and the evidence names the equivocator. Shown by the Equivocation scenario.
What makes a slot skip
The guarantees are paid for with skips, never with a wrong block. A byzantine proposer can make attestations disagree, for example by sending its commitment to only one co-proposer; validators then vote skip and the rest of the window is skipped for everybody. The skip certificate is the evidence. Shown by the Split commitments scenario.
Property and evidence
| Property | Paper | Fault Lab scenario | Evidence recorded |
|---|---|---|---|
| No last look | Lemma 2.3 | Last look | late mark in every attestation, block dropped |
| Censorship resistance | Lemma 2.2 | Censorship | the transaction executed from the third proposer's block |
| Crash tolerance | Theorem 2.1 | Crash | window finalized on two proposers |
| Accountable equivocation | Lemma 2.4 | Equivocation | the two signed messages, the slashing proof |
| Skip on disagreement | Theorem 2.1, Section 4.2 | Split commitments | disagreeing vectors, skip certificate |
| Transaction hiding | Section 3.2 | Private front-run | the unreadable fragment |