Current design, and why each piece
Core; settled
- Writer selection is a lottery.
- Why: it is the only selector no incumbent can gate. The merged rule is a rotation over a committee-held key list:
index = (tip_sequencer + (elapsed // posting_timeframe)) % len(accredited_keys)
assert msg.signer == chan.accredited_keys[current_sequencer_index]
- Anyone absent from
accredited_keys waits forever unless the incumbents vote them in at the configuration threshold.
- The lottery is the base layer’s own, reused rather than invented.
- A note of value
v wins a slot with probability 1 − (1 − f)^(v/D); f is the channel’s rate, D the difficulty.
- The merged spec attributes this to Ouroboros Crypsinous, and states its key property directly: “This matches the slot lottery, which activates a slot with probability f regardless of how many leaders win it.”
- The threshold is checked in the merged order-2 form
t = v(t0 + t1·v), with the constants precomputed outside the proof.
- Why reuse: it is merged, already analysed, and needs no new cryptography.
- A win is worth more than one write.
- Why: one write per win turns the channel’s update latency into a protocol constant equal to
f; at 1/60 that is one inscription per 60 seconds, half the base layer’s own rate. The rotation it replaces caps nobody at all; inside a turn the holder writes as often as it can chain. That regression was ours, and it is now reversed.
- The direction is settled; the shape is not.
- It is a quota of
k writes, or a turn bounded by the freshness window.
- A turn running until the next draw is unavailable under a private draw, because no validator can see a win before it lands.
- Several writes off one win cost either several proofs, or a reused one-time key. The leadership spec forbids the latter: “The key is single-use, as reusing the same one could allow multiple PoLs to be linked to the same identity. An observer could then infer the stake of that identity by observing the frequency at which it emits a PoL.”
- That trade is unresolved.
- A win is exercisable for a window of slots, not one block.
- Validity is
win_slot <= block_slot <= win_slot + Δ.
- Why: a win pinned to a single block dies to ordinary orphaning. A win that is orphaned, or that loses a race for the same parent, re-signs against the new tip inside its window.
- The base layer keeps a set of landed wins so one win lands once; prunable once
block_slot > win_slot + Δ.
- Δ is a per-channel configuration variable, like the rate. Examples for intuition: Δ = 600 slots is 10 minutes, Δ = 3,600 is an hour.
- A channel is created in rotation mode or lottery mode, and can be converted.
- Why convertible: a channel born in lottery mode holds no draw at all until its first declarations play. Converting an existing rotation channel serves that registry lag while the channel is still sequencing normally.
- The honest limit: conversion removes the outage, not the latency. A declaration filed alongside the flip still plays two epochs later.
- Declarations open at the moment the flip is filed, not when it takes effect, so the electorate arrives in the same instant the mode changes and the channel is never dark.
- A flip that would take effect with zero declared stake lapses rather than darkening a funded channel.
- Might want to be one way only, so lottery mode cannot go back to rotation? But we are already assuming an honest and well-behaved committee, so the guard buys little here. Frankly more relevant for the native zone variant, where the committee assumption is what goes away.
- Mode, rate and freshness window are changeable by the channel’s configuration quorum, with a delay, and irrevocably once filed.
- Why changeable at all: a value guessed before a channel has any participants should not be welded shut for life; the merged rotation already lets the same quorum rewrite its own cadence fields with no delay at all.
- Why this matters most on the undeclared variant: there the difficulty is the base layer’s own inferred total stake, so a large movement in base-layer participation moves the channel’s landing rate with nothing on the channel having changed. Being able to retune the rate in response is the only correction that variant has, because it carries no declaration register to read.
- Why delayed: a change taking effect immediately would be a winner-selector. The epoch nonce is public 388,800 slots in advance, so a quorum could try retunes until one produced a winner set it liked. A change effective two epochs out, and irrevocable once filed, removes the re-roll.
- Key rotation stays immediate throughout, because the key set is not an input to the draw.
- This is ours to argue; there is no steer from you behind it.
- Inscription contents are validated zone-side, by sequencers and followers. An invalid inscription is void in full, and is ignored rather than corrected.
- Sequencers treat it as though it does not exist; the next inscription carries no obligation to repair or reference it.
- Why void in full rather than partly applied: partial application would require every follower to agree on which parts of a payload survived, which is exactly the semantics the design keeps out of consensus.
- Why ignored rather than corrected: a correction obligation would make the next writer responsible for a predecessor’s error, an obligation no lottery winner accepted and none can be compelled to discharge.
- This is a convention the design exports to zones, not a rule the base layer enforces. You have already qualified it: “that may not be the case in practice. Maybe some zone users will just trust one of the sequencers through their centralized website for example.” It is stated as a convention for that reason, not as a guarantee.
The fork; should participation be declared?
- The question: is a participant named on the ledger before it can win, or is everyone eligible by default?
- It reduces to one exponent.
- Both sides share
occupancy = 1 − (1 − f)^W, where W is stake actually sequencing divided by difficulty.
- Declared makes the difficulty be the declared set, so
W sits near 1; the rate follows the channel’s own setting, to the extent declarers actually sequence.
- Undeclared leaves
W as a small participating share nobody controls.
- Both sides carry the same rate parameter, set the same way; the fork is not about which side controls its rate. It is about whether the denominator is the same set as the numerator.
- Because occupancy is distribution-independent, the number of participants does not move the rate on the declared side; five declarations or five thousand summing to the same
D behave alike.
Channels have a range of intended uses, and the design accommodates that rather than picking one
- Merged text already tells channels to choose by requirement.
- Asynchronous v Synchronous stated as property, from cross-channel messaging doc: “Channels should choose between synchronous and asynchronous messaging based on their requirements”, then gives two lists. Synchronous when “Atomicity is required”, “Low latency is important”, “Coordinating sequencers have established trust and communication pathways”. Asynchronous when “Simplicity and decentralization are prioritized over latency”, “Sequencer coordination is difficult or undesirable”, “The application can handle eventual consistency.”
- The named use cases differ in kind: zone state updates, liquidity providers (e.g bridged channel funds), atomic swaps between zones, cross-channel funds transfers.
- Broadest purpose: “outsource the hard parts of running a decentralized application to Logos, namely ordering and replicating state updates.”
- And you have said the range extends to how decentralised a channel wants to be at all: “in very niche applications, most of them will have a centralized Sequencer” · “maybe they won’t even want to do this decentralized sequencing”. The design imposes no threshold; a channel can stay in rotation mode forever.
- Not declaring buys the absence of a public roster. Nobody’s participation is disclosed by an act of entry; what leaks, leaks at each landing, through the note that pays the fee.
- But it couples the channel’s cadence to the base layer, because the difficulty is the base layer’s own inferred total stake rather than anything the channel holds.
- The channel’s landing rate is its configured rate multiplied by the share of base-layer stake that happens to sequence it.
- A large inflow of base-layer stake starves the channel: if the inferred total doubles and the channel’s own participants do not, its landing interval doubles, from roughly 99 minutes to 198 at one percent participation, and to about 397 if base-layer stake quadruples.
- A large exit does the reverse and speeds the channel up, roughly doubling its rate within two epochs if half the base-layer stake leaves.
- Either way the channel’s cadence moved and nothing on the channel changed. The configuration quorum retuning the rate afterwards is the correction mechanism, but it acts with a delay (two epochs) and the quorum has to respond.
- The coupling is to the stake estimate, not to base-layer throughput.
- Declaring decouples the channel from all of that, because the difficulty is the channel’s own declared sum and the base layer’s stake estimate never enters it.