Common Foundry v0.1.0-devnet.16 — ProductionV4 Testnet-1

Common Foundry v0.1.0-devnet.16 — ProductionV4 Testnet-1

This binary-only prerelease activates the isolated ProductionV4 transparent
BaseFold proof network. It is incompatible with ProductionV3 Devnet-15 and
uses RPC/P2P ports 22443/22444 and separate chain and wallet storage.

The release contains standalone node, GUI-wallet, and GPU-miner packages for
Windows and Linux. Nodes and wallets download the 6.4 GB model bank. Miners
download 61.2 GB of model and preprocessed prover inputs on first launch.
Launchers fetch parts concurrently, resume interrupted downloads, and require
the pinned byte count and SHA-256 identity for every part and assembled input.
Versioned inputs are served from downloads.commonfoundry.ai on Cloudflare R2,
with the GitHub binary release retained as an automatic fallback. This gives
testers a faster primary path without changing any authenticated input bytes.
Assembly now computes that final identity while writing instead of rereading
the completed file, and up to 16 parts download concurrently. Later launches
authenticate the 6.4 GB model and three Merkle trees once, while the 48 GiB of
row-major proving data is reused as an explicitly untrusted cache. Every row
opened from that cache is checked against a record-pinned tree, and every
finished proof still passes mandatory CPU self-verification before submission.
The Windows miner has a tester-ready START-MINER.bat that prompts for a
receive address. Windows launchers work with the built-in Windows PowerShell
5.1 or PowerShell 7. The Linux packages require curl and Python 3; the GUI
wallet also requires WebKitGTK 4.1.

The release now also includes self-contained reference-pool packages for
Windows/WSL2 and native Linux. Pool operators get durable share, block, credit,
and payout accounting; authenticated automatic Devnet payouts; bounded replay
admission; TLS certificate pinning; a polished loopback dashboard; and the same
resumable authenticated input bootstrap used by the miner. The pool defaults
to a seven-leading-zero-bit share target, one bit easier than the Devnet block
target.

Pool rewards now use exact-work-weighted PPLNS accounting. Each discovered
block freezes its rolling share window and operator fee, waits for the 100-block
coinbase maturity, then distributes every atom remaining after the fee. The
operator fee defaults to 3% and is configurable from both Windows and Linux
launchers; the automatic window represents one expected block of share work.
The dashboard discloses the current policy, per-block split, maturity progress,
and matured operator fees.

Pool packages now include a separate local operator console for Windows and
Linux. It presents live service health and accounting, the latest pool log,
guarded start, graceful-stop, and restart controls, and editable operator-fee
and PPLNS-window settings for future blocks. The console binds only to
127.0.0.1:22448; every state-changing request is same-origin protected, and
stop or restart always uses the pool's authenticated ledger-preserving shutdown
path.

The GUI wallet can mine those ProductionV4 pool jobs directly. It keeps the
payout key inside the wallet, searches server-issued work with the persistent
32-lane CUDA replay worker, and submits only the qualifying nonce. The pool
independently replays each submission before crediting it. Share-result waits
now tolerate normal short socket polling intervals while retaining a bounded,
interruptible five-minute response deadline, so an honest GPU verification no
longer causes the wallet to reconnect or resubmit stale work.

The Windows standalone miner can also mine to a pool using only its payout
address. The official test pool and its certificate pin are prefilled; edit
WALLET_ADDRESS and WORKER_NAME at the top of START-POOL-MINING.bat, then
double-click it. The certificate-pinned pool connection credits accepted work
to that receive address, while the pool still replays and verifies every
submitted nonce independently.

Windows mining now keeps authenticated hot inputs in the WSL-native filesystem
and creates each attempt's temporary trace in native Linux storage. Windows and
Linux launchers keep the replay and proof workers alive across blocks, so model
loading, artifact authentication, memory mapping, and CUDA setup are reused.
Nonce search runs continuously in a 6.021 GiB search-only replay mode, and only
a qualifying nonce is replayed with full traces. Dynamic commitment generation
is fused into the proof worker, which reuses the uploaded traces for relations
and openings. These changes preserve the proof transcript, mandatory CPU
verification, and all authenticated input bytes.

Windows and Linux miner consoles show compact session statistics only. Hashrate
is reported as Forge Work per second (FW/s), where one Forge Work is one
complete ProductionV4 nonce evaluation. Compute is the corresponding TMAC/s
at the fixed consensus geometry of 0.824633720832 trillion multiply-accumulates
per Forge Work. The same line includes accepted and rejected submissions,
average GPU wattage, accepted blocks per kWh, peak GPU temperature, and
last-attempt time. Full prover diagnostics are retained under work/logs and
are printed only when an unexpected failure stops mining.

Nodes now exchange bounded peer lists after the network and consensus
handshake. Newly advertised addresses are connect-back verified before they
are cached or shared, while total, per-IP, retry, and message limits keep
discovery lightweight. The community bootstrap provides first contact, and
compatible nodes can then discover one another automatically.

Wallet transactions now propagate outward to configured and discovered peers
through bounded inventory comparison and the node's normal transaction
validation policy. This lets a standalone miner connected to the community
bootstrap include transactions created by a separately running wallet, while
duplicate, invalid, and over-capacity submissions remain fail-closed.

Wallet and node startup now keep the authenticated fast-start checkpoint
current whenever the canonical chain advances. On the live Windows Devnet
wallet, a warm restart with a 6.25 GB block log reached its ready node listener
in 9.5 seconds, compared with about 8 minutes 40 seconds for the one-time full
historical replay used to establish the checkpoint. The append-only block log
remains authoritative, and deterministic replay remains available whenever a
checkpoint cannot be used.

P2P ingress now includes bounded per-source reputation, temporary bans for
repeated objective protocol failures, and connection-attempt and concurrency
limits. IPv4 addresses and IPv6 /64 prefixes are tracked consistently, while
ordinary outbound outages do not reduce a peer's reputation. These controls
make public nodes more resilient without changing consensus behavior.

ProductionV4 replaces the multi-hour BLS12-381 transition proof with a
KoalaBear-native GPU proof. The complete proof is exactly 12,025,320 bytes
inside the ProductionV4-only 16 MiB block bound; older networks retain their
existing limits.

A live RTX 5090 acceptance run with 32-nonce search batching produced block
236 in 13.8 seconds from snapshot through node acceptance. The batch evaluated
32 complete candidates in 2.421 seconds of GPU time and 2.935 seconds of search
wall time. It reported 10.90 FW/s and 8.99 TMAC/s, up from 4.39 FW/s on the
single-nonce stacked path. Search reached 100% GPU utilization and 570 W during
the batched kernel, while the complete search, replay, proof, CPU verification,
and submission cycle averaged 192.8 W. The winning full replay took 0.132
seconds, fused proof construction took 7.547 seconds, mandatory CPU
self-verification took 0.296 seconds inside that proof phase, and submission
took 0.584 seconds. The finished 12,025,320-byte proof verified and all required
mutation checks were rejected.

Batching changes only search throughput. Candidates share the resident model,
then the first qualifying nonce in canonical nonce order is rebound and replayed
through the unchanged full-trace path. Its final activation must match the
batched result byte for byte before proof construction. An eight-nonce
differential test matched every coefficient stream and every complete final
activation against separate single-nonce replays. Batch size 32 was selected
after testing 1, 2, 4, 8, 16, 32, and 64 candidates; 64 used more memory without
improving throughput. The 32-lane search allocation is 6.445 GiB, preserving
the memory-conserving 16 GB GPU path.

Physical RTX 5070 Ti 16 GB qualification produced and CPU-verified the exact
proof on the prior process path. The new launchers automatically use a
memory-conserving schedule below 20 GB: the replay model is evicted before
proving and restored for the next search. The measured persistent proof and
full-replay allocations total 14.576 GiB at their conservative overlap bound,
keeping the supported 16 GB tier while larger cards retain the model for lower
latency.

An accepted height-2 block was produced through normal P2P admission and was
downloaded, verified, and persisted by an independent second node. This is an
experimental low-difficulty test network. Coinbase rewards remain immature
for 100 blocks before becoming spendable. Coins have no value.