Verathos v0.1.42 – Verified MoE, Account Billing, and Runtime Stability

Verathos v0.1.42 – Verified MoE, Account Billing, and Runtime Stability

This release improves Gleipnir v3 hard-audit reliability and validator
capacity-receipt processing, and adds qualified serving for the
official Qwen3.6 35B sparse-MoE FP8 checkpoint.

The release also adds wallet accounts, API keys, unified USD inference credit
including conviction-locked inference credits, direct TAO and prepaid-USDC
funding, and accountless x402 payments.

What changed

• Captures the authenticated prompt-boundary GDN checkpoint required by a
  one-token hard audit.
• Preserves natural EOS and configured stop-token terminations during exact
  post-nonce hard replay.
• Adds fail-closed Qwen3.6 35B sparse-MoE execution profiles, authenticated
  expert routing and compact hard-proof verification.
• Adds wallet-authenticated accounts, API-key access, audited USD credit,
  conviction-locked inference credits, direct TAO and prepaid-USDC top-ups,
  while retaining accountless x402 access.
• Qualifies the stock installer and serving/proof path across A100, H100 and
  Blackwell GPUs while preserving existing dense FP8, INT4 and unquantized
  model behavior.
• Keeps capacity-receipt ingestion bounded and independent from deferred
  scoring and shared-state work.
• Aligns mainnet validator scoring with routable endpoint policy: HTTP or
  non-public registrations receive zero effective score and weight without
  audit probation, their stored score decays normally until corrected, and
  miner diagnostics provide the exact HTTPS setup and re-registration step.
• Stops mainnet miners before model loading when their supplied public
  --endpoint is not eligible, while preserving HTTP for testnet/local use.
• Adds fail-neutral validator-incident handling for capacity windows with
  authenticated evidence of local ingestion disruption, while preserving
  enforcement of cryptographically invalid proofs.
• Moves informational metagraph work out of chronology-sensitive block
  processing.
• Improves CUDA/Torch/vLLM installer selection on Ada, Hopper and Blackwell
  systems and reports progress during cold weight-root preparation.
• Preserves the existing proof wire format, signed model artifacts,
  tolerances, canary policy and ordinary multi-token proof semantics.

Qwen3.6 35B software support is included in this release and prepared for
separately authenticated model activation.

Updating

Miners, validators and API proxies using auto-update will install v0.1.42 and
restart at a safe idle boundary. Operators updating manually should pull the
release and use the standard setup and process-management flow.