Skip to content

Consensus Neural Work

github-actions[bot] edited this page Nov 20, 2026 · 3 revisions

Hyphen Neural Work Consensus Profile

Status: consensus profile hyphen-neural-devnet-v1, 2026-10-19. Research network; not audited, not value-bearing.

This profile replaces the Cunningham-witness block rule with a one-way neural work challenge. It is a separate chain identity, never an in-place upgrade: network_name, network_magic, consensus_features and therefore consensus_params_hash and the genesis hash all differ from devnet v2.

1. The challenge

For a block at height h with parent hash p and miner key m:

seed        = BLAKE3("Hyphen/neural-work-seed/v1" ‖ p ‖ h ‖ m)
weights     = ChaCha20(seed, epoch = h ^ 2^63)      // hyphen-neural::work
x           = the block's witness list, input_dim Q32.32 values
y           = A_weights(x)                           // one fixed-point forward pass
score       = mix64-fold(y)                          // deterministic 64-bit score
accept      <=> score < threshold

x must lie in [-1, 1] per component (ONE = 2^32). threshold is the header's difficulty field. Under an independent uniform-score model, a trial succeeds with probability threshold / 2^64 and the mean trial count is 2^64 / threshold. Consensus checks only the score of the submitted input.

Three properties are load-bearing and each has a regression test:

  1. The seed determines the weights. The reference miner generates candidate inputs and evaluates each with a forward pass. A lower bound on the cost of alternative solvers remains open.
  2. The seed includes the miner key. Without it, a solution found by one miner is valid for every other miner's block at the same height and could be copied directly. Changing the miner key requires evaluating the input under different weights and signing the resulting block.
  3. The header commits to the solution. pow_commitment is BLAKE3("Hyphen/neural-work-commitment/v1" ‖ p ‖ h ‖ threshold ‖ m ‖ x), so neither the threshold nor the solution can change after the search.

Each parent defines a new challenge. Reusing an input requires that it also satisfy the new threshold test.

2. Difficulty

Once 17 parent headers are available, each new block takes its threshold from the latest 16 intervals. Consecutive windows overlap by 15 intervals.

expected = 16 × WORK_TARGET_MS          // 480 000 ms
elapsed  = newest.timestamp - oldest.timestamp   // milliseconds
t'       = clamp(t × elapsed / expected, 1, NEURAL_WORK_MAX_THRESHOLD)

Header timestamps are milliseconds — the same unit as GENESIS_TIMESTAMP_MS and every operator-facing timestamp in the repository. An early build of this profile compared that millisecond elapsed against a second-denominated expected, read every fast window as six times too slow and walked the threshold up to u64::MAX. At u64::MAX the test score < threshold accepts almost every candidate, so the devnet mined at 5 blocks per second with no work at all. WORK_TARGET_MS and the regression test a_window_of_sub_second_blocks_hardens_instead_of_easing pin the unit.

The ratio elapsed / expected is clamped to 1/4 .. 4 before scaling, so one timestamp outlier cannot swing the chain, and the result is capped at NEURAL_WORK_MAX_THRESHOLD = 2^62. The cap is not cosmetic: it is what keeps "hard to find" from degenerating into "free" when hashrate collapses or a retarget loop misbehaves. NeuralWorkParams::validate rejects any profile whose ceiling exceeds it.

Until a full window exists the profile's neural_work_initial_threshold (default 2^56) applies. Both thresholds live in ChainConfig and are hashed into consensus_params_hash whenever FEATURE_NEURAL_WORK is set, so changing either one mints a new chain identity instead of silently redefining an existing chain.

Timestamps MUST exceed the median of the last 11 parent headers (WORK_MEDIAN_WINDOW). This is the deterministic half of the timestamp rule; the "not too far in the future" half compares against the node's NTP-adjusted clock and is applied by the node before consensus validation.

3. Fork choice

The header field is a threshold, so lower is harder. The v1 fork-choice rule uses integer division.

weight(block) = floor((2^64 - 1) / threshold)   // post-genesis; genesis contributes 0
cumulative    = Σ weight

Summing the raw field, as witness-count profiles do, would rank the easiest branch as the heaviest and invert the choice. Blockchain::block_weight and AtomicBlockUpdate::block_weight carry the rule; the state layer never derives it from the header.

4. Activation state

The node validates the profile end to end: genesis, next_difficulty, block acceptance, reorg weight and the API surface all branch on FEATURE_NEURAL_WORK. A local run of the three components mines this profile:

hyphen-node  --network neural-devnet --rpc-bind 127.0.0.1:49333 \
             --template-bind 127.0.0.1:4350 --explorer-bind 127.0.0.1:9080
hyphen-pool-server --network neural-devnet --node 127.0.0.1:4350 \
             --bind 127.0.0.1:4340 --api-bind 127.0.0.1:9181 --payout-mode solo
hyphen-miner --network neural-devnet --pool 127.0.0.1:4340 --wallet-address hy1...

Two protocol properties differ from the witness profile and are enforced by the clients:

  • No credit shares. Partial neural work is not verifiable, so the pool sets the job's share difficulty to the block threshold and every accepted submission is a complete block. PPLNS-style accounting does not apply to this profile; use --payout-mode solo.
  • One solution per job. A miner races its own workers; the first solution claims the job and the rest stop, because a second solution reaches the pool after the template has moved on.

HyphenWallet pins all four identities, including hyphen-neural-devnet-v1, so it can read this chain; it also refuses any node whose network, magic, consensus_params_hash, genesis hash, block version or PoUW protocol version does not match a pinned vector.

Current pinned values for this profile:

network              = hyphen-neural-devnet-v1
network_magic        = 48594e4e                     ("HYNN")
consensus_params_hash= 451ca637dafb07604b8dcabebc86c54d72efe590ac5912bb2a2f35ea6d41b6a9
genesis_hash         = 6d3c028ecd7380ba143a62c8b82fa706155da56fb42ec8cf8b30ec4dc989d058

Because the retarget parameters are hashed, this identity changed when the millisecond fix and the threshold ceiling landed. Every chain mined before that change is a different chain and must be discarded, not migrated.

5. Publishing the client changes

HyphenMiner and HyphenPool pin their shared crates to Git revision 951ea013f0b61115632f463cea22850cd972df83. HyphenWallet/rust uses the same revision. Each repository builds with its own lockfile. From the base-chain checkout, run ./scripts/devnet-smoke.ps1 -Network neural-devnet to compare the first pool settlement with the node's block hash and the independent wallet's decoded amount.

6. What the work is not

The network's output is discarded by the puzzle. This is a hash-gated computation whose per-attempt cost is a forward pass; per unit of security it is less energy-efficient than a BLAKE3 threshold. It is not "useful work" in the scientific sense, and this profile does not claim otherwise. Making the output useful requires a consumer of the result, which is the AetherCompute settlement layer, not the block rule.

Clone this wiki locally