# Consensus

> Reputation instead of stake

Phase uses committee-based BFT consensus (Chained HotStuff). At any given time, a committee of validators collectively produces and certifies blocks. The committee has a budget of 16 voting seats (a client-release parameter, not a function of how many validators register), filled first by the genesis validators — the network's permanent *anchors*, seated whole in every committee — and then by graduated validators, who hold at most half as many seats as the anchors minus one, so that the anchors alone always form a quorum. With seven anchors the committee seats seven to ten voters with a quorum of five to seven. Votes are aggregated into a compact quorum certificate using BLS signatures, messaging complexity is linear in committee size, and finality is deterministic — a block is final after three chained certificates, seconds after it is produced.


The question the protocol must constantly answer is: **which validators should be in the committee?** In most blockchains, the answer is determined by how much capital a validator has staked. In Phase, the answer is determined by **reputation** — a continuously updated score derived entirely from on-chain behavior.

## Registration is free — and cryptographically strict

Becoming a validator requires a single system transaction with zero gas and no deposit. The transaction registers the validator's address, a BLS key for consensus voting, and a network identity key, together with a proof-of-possession bound to the sender and chain — the standard defence against rogue-key attacks. All keys are derived deterministically from the phone's single master key. A registration remains valid for a short window (320 blocks) and is renewed automatically while the node is running, so abandoned devices fall out of the validator set on their own.

## Reputation scoring

Every block records which committee members signed on time and which leader delivered or failed. Those observations accumulate into a per-validator reliability score. The scoring system is designed to be:

- **Conservative** — the score uses a lower confidence bound: a validator with a few perfect votes scores lower than one with hundreds of consistently good votes. Evidence matters, not luck.
- **Adaptive** — recent performance matters more than distant history, so a validator can recover from a bad patch and a new device can graduate into the system over time.
- **Stress-aware** — the scoring model specifies a correction that distinguishes individual unreliability from systemic outages, so a network-wide failure does not destroy individual reputations; it is specified, not yet active.

## Committee selection: deterministic and verifiable

Once per epoch the protocol freezes a *selection snapshot*: the list of active validators with their reputation-derived weights. Each round's committee is computed from that snapshot by a deterministic function — every node in the network derives byte-for-byte the same committee and can verify the legitimacy of any block's signers. There is no hidden operator choosing who validates.

New validators follow a defined lifecycle: they enter a queue (with a per-epoch admission limit that caps Sybil inflows), serve a probation period in *reserved observer seats* — signing and earning reputation, but counting toward no quorum and not yet leading — and graduate to full membership once they have accumulated enough evidence of reliability. Graduates then compete, weighted by reputation, for the voting seats beside the anchors. This guarantees newcomers a real path in, without grace periods or special treatment. Leadership rotates round-robin among the voting members.

## Rewards only after finality

Validator rewards are released only for heights already finalized under a full quorum, and a payout becomes irreversible as soon as the block that credits it is itself final. This turns the validator's future income stream itself into skin in the game: the thing at risk is not a deposit, but the earning position built through honest participation.

## What happens when a validator misbehaves

There is no slashing — there is no bonded capital to slash. Instead, the protocol's answer to a provable fault is **jail** and a **reputation reset**: the validator loses committee eligibility, the income stream that came with it, and the accumulated track record that took months to build. For a long-serving validator this is a substantial economic penalty — and unlike a stake deposit, it cannot be bought back; it can only be re-earned through honest participation. The cost of misbehavior scales with tenure. The fault detector ships today; the automatic jailing step is specified but not yet active, so in the current build the penalty is the score itself — missed votes and failed leadership cost seats, and seats are income.

## Designed for unreliable validators — because phones are

A committee of phones is not a committee of servers: devices lose signal, switch networks, and run out of battery. Classical BFT protocols simply halt when a third of the committee goes silent. Phase is engineered to degrade gracefully instead:

- **Tentative certificates.** If the full quorum is temporarily unreachable, a reduced quorum can keep the chain moving: blocks continue to be produced and transactions continue to execute, while finality — and the rewards gated on it — catches up retroactively once the full quorum returns. The network slows down; it does not stop.
- **Automatic committee rotation.** A committee that makes no certified progress for a minute deterministically moves to a new committee round: the anchors keep their seats, the leader order is reshuffled, and the graduate and observer seats are redrawn wherever more candidates wait than there are seats. Connections to the incoming committee are pre-warmed before the handover.
- **Self-cleaning validator set.** Systematically unreliable devices lose score and fall out of selection on their own — no governance action required.