Verity · SpiderMesh

The work runs twice, on two people's machines.

SpiderMesh is a scattered compute network built out of volunteered hardware. Verity calls this execution surface SpiderMesh. The interesting part is not that the work is distributed. It is that a distributed result can be checked without trusting either machine that produced it.

WORK UNITOWNER AOWNER Bnodenodedigest A == digest BVERIFIED
2assignments per work unit
0temperature
7seed
1,024MB storage cap
What a node contributes

A short, honest contract.

Six fields, declared once at registration and checked by the coordinator. Three of them are floors a job can demand; two of them you can decline outright without being penalised.

cpu_coresrequired
Must be greater than zero. Registration is refused without it.
ram_mbrequired
Must be greater than zero. Registration is refused without it.
gpu_countrequired
A required field, but zero is a valid answer. A CPU-only node is a real node.
gpu_vram_mbrequired
Total VRAM across the declared devices. Jobs can demand a floor.
bandwidth_mbpsoptional
Zero means not offered. Nothing about the node is degraded by declining.
storage_mboptional
Encrypted, temporary, and capped at 1,024 MB by the coordinator and again by the daemon, with a per-entry cap on top.

BoundScratch storage is encrypted, temporary and size-capped in two places: the coordinator refuses an over-cap registration, and the daemon enforces the same ceiling locally with a cap per entry on top. It is a scratch volume for a running unit, not a place to keep anything.

One work unit, end to end

Consensus is an equality check, not a vote.

  1. 01

    Register

    A node declares what it has. CPU cores and RAM are mandatory; a GPU, bandwidth and scratch storage are each optional. Over-declaring storage does not get you more work. It gets the registration refused.

  2. 02

    Claim

    The coordinator hands the unit to two nodes and no more, and it will not give both assignments to the same owner. That different-owner rule is the sybil-pair guard, and it is enforced at claim time, not audited afterwards.

  3. 03

    Execute

    Inference runs at temperature 0 with a fixed seed on the node's own model runtime. Determinism is not a nicety here; it is the entire verification strategy.

  4. 04

    Digest

    Each node returns a canonical digest of its output. The coordinator compares digests and never reads the text into a prompt, so a hostile node cannot talk to the orchestrator.

  5. 05

    Verdict

    Equal digests: verified, and both owners earn credit. Unequal: conflicted. Missing: incomplete. There is no averaging, no tie-break, and no benefit of the doubt.

Why determinism

Byte-equal, or it did not happen.

Grading similarity means picking a threshold, and picking a threshold means arguing about it. Most verification schemes for distributed inference end up there.

Fixing temperature at zero and the seed at seven removes the argument. Two honest nodes running the same unit produce the same bytes, so the comparison is string equality on a hash and nothing else.

It also keeps untrusted text out of the loop. The coordinator only ever handles digests, so node output never becomes something the orchestrator reads. That is the rule Verity applies to every other untrusted string.

  • VerifiedBoth digests present and identical. The digest becomes canonical and both owners earn credit.
  • ConflictedDigests differ. Nobody is paid for a result nobody can trust, and the unit does not resolve on a coin toss.
  • IncompleteA digest is missing or empty. The unit simply is not finished yet, and an absence is never read as agreement.

BoundDeterminism holds for nodes running the same model build through the daemon's own inference path. Different builds will disagree, and that disagreement surfaces as a conflicted unit rather than being smoothed over. The network is young: treat published capacity as a description of the protocol, not of a large live fleet.

No comparable product offers user-contributed distributed inference. That finding comes from Verity's own competitor audit, which compares the shipped code against the strongest comparable products and is equally blunt about where Verity is behind.

Where Verity is behind

Lend a machine. Or use the ones that were lent.

Register a node from the workspace, declare what you are willing to give, and watch units resolve to verified or conflicted.