NEB 0.3.1, node display fix

NEB 0.3.1, testnet 2. Same protocol, network, genesis, dataset file and rules at block 8800 as 0.3.0: nodes of both versions work together.

This release fixes the console of neb-node --run. The upgrade is optional. Nodes still on 0.2.1 must move to 0.3.0 or 0.3.1 before block 8800.

• The height on the status line now follows the best chain. Before, it only moved when the node's own work service connected a block, so a node that synced from its peers kept showing the height it had at start. The chain itself was not affected: getblockcount always gave the right height.
• At start the node prints a line before it replays its stored chain, and the time the replay took. The replay checks every block and every share again and can take minutes.

Wallet-only users and miners: no action is needed.

To upgrade, stop the node, replace neb-node, and start it again with the same command and data directory. The dataset file is the same as in 0.3.0.

• Windows and Linux (x86_64): neb-node, neb-wallet, neb-miner-cpu and neb-miner-gpu. HiveOS: nebminer-0.3.1.tar.gz. The Linux binaries need glibc 2.34 or later.
• neb-datasets.bin: unchanged, 259,897,080 bytes, hash 296a174edbf67339fa0af341f063154292bba224b3d2cacf5b18b31c70da80eb (neb-node --dataset-hash). neb-datasets.zip is the same file compressed (26 MB).
• Verify: gpg --import release-key.asc, gpg --verify SHA256SUMS.asc SHA256SUMS, sha256sum -c SHA256SUMS. The fingerprint of the release key is in the README.
• Changes: CHANGELOG.md.

NEB 0.3.0, nebhash v2 from block 8800

NEB 0.3.0, testnet 2. Same network, genesis and chain: the blocks and the coins stay as they are.

Every node and every miner must upgrade before block 8800, with the new neb-datasets.bin. From block 8800 the Kepler model, and from block 8801 the retina model, are small convolutional networks (nebhash v2, spec/nebhash-v2.md), each starting from a new model. The retina model also moves to scans of 64 x 64 and learns normal or abnormal instead of four labels. Block 8800 should come around Friday 13 November, 19:00 UTC; the height decides, not the clock.

In simulation of the chain's training rule: Kepler 90.8 % on stars it never trained on (88.3 % with the current network), retina 81 to 83 % on 500 balanced test scans for normal or abnormal (about 40 % on four classes today).

Up to block 8799 this release accepts the same blocks as 0.2.1, so it can be installed now. A node or a miner still on 0.2.1 stops at block 8800.

Wallet-only users: no action is needed.

One attempt costs more from block 8800: about 2 ms of one CPU core for Kepler and 7 ms for the retina, against 0.4 ms today, and much more on a GPU than a v1 attempt. Hash rates drop by that factor and the difficulty follows within the first hour.

To upgrade, stop the node and the miner, replace the binaries and neb-datasets.bin, and start them again with the same commands and data directory.

• Windows and Linux (x86_64): neb-node, neb-wallet, neb-miner-cpu and neb-miner-gpu. HiveOS: nebminer-0.3.0.tar.gz. The Linux binaries need glibc 2.34 or later. The GPU miner now also covers compute capability 9.0.
• neb-datasets.bin: new, 259,897,080 bytes, hash 296a174edbf67339fa0af341f063154292bba224b3d2cacf5b18b31c70da80eb (neb-node --dataset-hash). neb-datasets.zip is the same file compressed (26 MB).
• Verify: gpg --import release-key.asc, gpg --verify SHA256SUMS.asc SHA256SUMS, sha256sum -c SHA256SUMS. The fingerprint of the release key is in the README.
• Changes: CHANGELOG.md.

Kepler Bonus vetting: season and shape tests (science-kbonus-2)

science-kbonus-2: two more vetting tests for the Kepler Bonus search. No node update.

vet.py now compares the transit depth between Kepler seasons and checks that a planet can make the transit shape around the star. On the survivors of the hunt (revet.py), the two tests leave no inverted-flux survivor above MES 10 (there were 9) and reject the known blend KIC 9966110.

Miners, pools and wallets: nothing to do. The node, the protocol and the binaries stay the same, and v0.2.1 remains the latest release.

KIC 6691030 passes both tests and stays a candidate. Known planets in Kepler Bonus PSF light curves change depth with the season as much as it does (deepest/shallowest ratio up to 1.98 for them, 1.72 for KIC 6691030), and at the 29.4-minute cadence a U-shaped and a V-shaped dip look alike for this signal.

Details and numbers: https://github.com/SilasEigen/neb/tree/science-kbonus-2/science/kbonus

Kepler Bonus transit search (science-kbonus-1)

science-kbonus-1: a transit search in Kepler Bonus with the chain's Kepler model. No node update.

This tag adds science/kbonus, Python scripts that search the Kepler Bonus light curves of background M dwarfs. They vet each signal the way the Kepler pipeline does, run the same search on the inverted flux as a control, and score what survives with the Kepler model the miners trained (slot 0, weights of block 1428).

Miners, pools and wallets: nothing to do. The node, the protocol and the binaries stay the same, and v0.2.1 remains the latest release.

The hunt covered 2,000 background M dwarfs (1,585 usable light curves) and kept one transit candidate, KIC 6691030: P = 4.772 d, depth about 3,800 ppm, MES 30.9, 268 transits. It carries two warning flags. The transit shape gives a lower stellar density than the TIC, and a 74 ppm phase-curve signal at P/2 could mean an eclipsing binary. It needs follow-up by astronomers before anyone can call it a planet. The vetting sheet is attached.

Method, results and how to run: https://github.com/SilasEigen/neb/tree/science-kbonus-1/science/kbonus
Model weights hash (block 1428): d8d471ac7f9ee7a0b28f6f8aa1d5e6e647331a2b10f3929807155f1a14b4334a
Data: Kepler Bonus PSF light curves, Martinez-Palomera, Hedges & Dotson (2023), MAST, doi:10.17909/7jbr-w430, CC BY 4.0.

If you work on Kepler data and want to check the candidate, open an issue on this repository.

NEB 0.2.1, testnet 2

NEB 0.2.1, testnet 2. Same protocol (3), network and genesis as 0.2.0: nodes of both versions work together, and testnet 2 still starts on 2026-11-07 at 19:00 UTC.

Miners should run their node on 0.2.1. When no block arrives for 30 s, the node now sends its miners a fresh job on the same tip. A job holds 2^32 nonces, which a fast GPU tries in about two minutes; with 0.2.0 a miner that had tried them all waited at 0 attempts/s until the next block.

Wallet-only users: no action is needed.

To upgrade, stop the node and the miner, replace the binaries and start them again with the same commands and data directory. neb-datasets.bin is the same file as in 0.2.0.

• Windows and Linux (x86_64): neb-node, neb-wallet, neb-miner-cpu and neb-miner-gpu. HiveOS: nebminer-0.2.1.tar.gz. The Linux binaries need glibc 2.34 or later.
• neb-datasets.bin: unchanged, hash b1008b970f87083179a7b297b8f0fbec4cfc8cc66f605cf2fb2c03bad8b6f37a (neb-node --dataset-hash).
• Verify: gpg --import release-key.asc, gpg --verify SHA256SUMS.asc SHA256SUMS, sha256sum -c SHA256SUMS. The fingerprint of the release key is in the README.
• Changes: CHANGELOG.md.

NEB 0.2.0, testnet 2

NEB 0.2.0, testnet 2. Mining now trains two models on public scientific data, one block each in turn: transit signals of the Kepler mission (NASA Exoplanet Archive), sorted into planet candidates and false positives, and retinal OCT scans (OCTMNIST), sorted into three diseases and normal.

The network starts on 2026-11-07 at 19:00 UTC: no block can be dated earlier.

Testnet 2 starts from a new genesis. Protocol 3. The genesis id is 8271952aba2ac5340c1dfe884caa1f2c6ee765487fa2117ddcd90710ce87248d and the network magic is 1db70ba3; neb-node --genesis prints them. Nodes of protocol 2 do not connect, and the coins of testnet 1 do not carry over. Wallet files and addresses stay the same.

Miners and node operators: download the archive for your system and neb-datasets.bin, stop your node and start the new one with a new data directory (--datadir). The commands do not change; --dataset now takes neb-datasets.bin.

• neb-datasets.bin: the two datasets in one file (38 MB), pinned by hash (b1008b970f87083179a7b297b8f0fbec4cfc8cc66f605cf2fb2c03bad8b6f37a, see neb-node --dataset-hash). tools/make_datasets.py rebuilds it from the public sources.
• Windows and Linux (x86_64): neb-node, neb-wallet, neb-miner-cpu and neb-miner-gpu. HiveOS: nebminer-0.2.0.tar.gz. The Linux binaries need glibc 2.34 or later.
• The miners keep one dataset per model and switch with each block; the GPU miner loads both on the device.
• getmodel [SLOT] gives the weights of a model (0 Kepler, 1 retina); the explorer shows both.
• In a simulation of the protocol, the Kepler model reaches 88.8 % on stars it never trained on (a float network of the same shape: 89.0 %), and the retina model 55 % on four balanced classes after 12,000 blocks.
• Verify: gpg --import release-key.asc, gpg --verify SHA256SUMS.asc SHA256SUMS, sha256sum -c SHA256SUMS. The fingerprint of the release key is in the README.
• Changes: CHANGELOG.md.

NEB 0.1.1

NEB 0.1.1, testnet. Same protocol (2), network and genesis as 0.1.0: nodes of both versions work together, and the chain continues.

Node operators should upgrade. Stop the node, replace the binaries and start it again with the same data directory and options. Miners connected to a node that runs 0.1.1 need no other change.

• Peers: a write to a peer that takes more than 10 s ends its session; a session that has not completed the handshake within 10 s is closed; inbound sessions leave two slots for outbound ones and are limited to two per address; the listener keeps accepting after a transient error. A new block's job goes to the miners before the block is announced to the peers.
• Mempool: a transaction pays at least one base unit per byte (the fee of neb-wallet, 10,000 base units, covers transactions of up to 10,000 bytes); a full pool keeps the transactions of the highest fee rates, and templates take transactions by fee rate.
• Windows and Linux (x86_64): neb-node, neb-wallet, neb-miner-cpu and neb-miner-gpu. HiveOS: nebminer-0.1.1.tar.gz. The Linux binaries need glibc 2.34 or later.
• mnist-q8.bin is the same file as in 0.1.0.
• Verify: gpg --import release-key.asc, gpg --verify SHA256SUMS.asc SHA256SUMS, sha256sum -c SHA256SUMS. The fingerprint of the release key is in the README.
• Changes: CHANGELOG.md.

NEB 0.1.0

NEB 0.1.0, testnet. First release.

The network starts on 2026-11-05 at 19:00 UTC: no block can be dated earlier.

Protocol 2. The genesis id is 7c43ea053c733620043a1aa67d54e66eb41d4ac0c6d5a53f3a29176ace454cd2 and the network magic is 6bcb09e5; neb-node --genesis prints them. A share is bound to the parent of its round, a block lists up to 256 share records of 40 bytes, and nodes relay new records with the shares message (spec/shares.md section 11, spec/node.md).

• Windows and Linux (x86_64): neb-node, neb-wallet, neb-miner-cpu and neb-miner-gpu (NVIDIA GPU of compute capability 7.5 or later, driver 580 or later).
• Optional engines, off by default: neb-miner-cpu --engine avx512 (x86-64 with AVX-512) and neb-miner-gpu --gpu-kernel k2. They are mining optimisations, not consensus; see docs/testnet.md.
• HiveOS: nebminer-0.1.0.tar.gz is the custom miner package (miner name nebminer, see docs/testnet.md, "Mining on HiveOS"). It needs an Ubuntu 22.04 based image.
• The Linux binaries need glibc 2.34 or later (Ubuntu 22.04, Debian 12).
• Mining needs a lock: start the node with --miner-lock (the value printed by neb-wallet lock); without it the node relays blocks but refuses miners. The coins of the genesis block can never be spent.
• neb-wallet new asks for a password and encrypts the wallet file (Argon2id, XChaCha20-Poly1305); neb-wallet encrypt encrypts an existing plain file.
• mnist-q8.bin: the dataset file (47 MB). The node refuses any file that does not have the pinned hash.
• Public relays: a node dials seed1.neuralenginebyte.org:3440 (New York) and seed2.neuralenginebyte.org:3440 (London) by itself. --seed HOST:PORT adds another relay and --no-seeds keeps the node alone on its own chain.
• Measurements: docs/experiments.md describes them, and neb-0.1.0-experiments.tar.gz holds the raw outputs, the bench sources, the scripts and the validation logs. They were taken on one machine, with real nodes joined on the loopback and simulated miners; none comes from a public network.
• The miners show a dashboard in a terminal of at least 120 x 30 (--plain prints text lines).
• Verify: gpg --import release-key.asc, gpg --verify SHA256SUMS.asc SHA256SUMS, sha256sum -c SHA256SUMS. The fingerprint of the release key is in the README.
• Coins on this network have no value.