FARPY Public Technical Brief

2. Abstract

FARPY’s public model is organized around verifiable execution: published benchmark results connect a controlled workload, measured result, hardware configuration, and verification outcome to inspectable public evidence.

3. Scope and purpose

FARPY exists to make distributed compute easier to trust, compare, and use.

This brief describes the current public model, the evidence presently published, and the boundaries of those claims. It does not describe every internal mechanism.

4. Claims and non-claims

Claims

• Verified benchmark results can be published with inspectable evidence.
• Known physical inventory is separate from verified benchmark configurations and verified results.
• Missing evidence remains missing rather than becoming an estimate.

Non-claims

• Known inventory is not a claim that every listed GPU has a verified benchmark.
• The current evidence does not establish verified RTX 3080 performance.
• The current evidence does not establish verified Octane performance.
• FARPY does not claim that every workload is deterministic, every result is independently reproduced, or arbitrary workloads are supported.
• This brief makes no uptime, capacity, security-certification, compliance, or settlement guarantee.

5. System model

Control

Validates bounded work, coordinates execution, and evaluates returned evidence.

Execution

Runs bounded workloads on eligible GPU hardware.

Truth

Publishes inspectable result, benchmark, and proof artifacts when evidence qualifies for public classification.

6. Evidence lifecycle

1. A bounded workload is submitted.
2. An eligible worker executes that workload.
3. Outputs and execution evidence are returned.
4. Verification classifies the returned evidence as passing, failing, or unavailable.
5. Accepted evidence governs result classification and accounting eligibility.

An accepted receipt records the evidence used to classify completed work and determine accounting eligibility under FARPY’s governing model. It does not, by itself, prove that a payment was completed. Payment completion is a separate event requiring its own evidence. This paper does not claim public payout volume, payment timing, or automatic payment completion.

No receipt, no payout. A payout cannot become eligible without an accepted receipt. This principle does not guarantee payment timing, payment completion, or public payout volume.

7. Verification semantics

• Frozen method. The procedure, software versions, parameters, and evaluation rules are fixed before execution. Freezing a method does not make a workload deterministic.
• Deterministic workload. The workload is expected to produce the same defined output under the same controlled conditions. Determinism does not, by itself, establish reproducibility or prove that independent reproduction occurred.
• Comparable benchmark. Runs use sufficiently controlled methods to support comparison. Comparability does not establish that a result is reproducible.
• Reproducible result. Repeated evidence supports the same result under the stated method and conditions. Reproducibility remains workload- and evidence-dependent.
• Independently reproduced result. A separate execution reproduces the result and is explicitly recorded as such. FARPY does not claim independent reproduction without a separate matching execution record.
• Failures remain visible.
• Incomplete or failed evidence does not graduate into verified truth.

8. Current public evidence

The current public evidence establishes:

• 1 verified benchmark configuration.
• NVIDIA GeForce RTX 3090 with 24 GB.
• Blender using CUDA/OptiX.
• A verified runtime of 4.08 seconds.
• Farpy Score 24203.
• 0 verified Octane results.

The public inventory separately lists 5 known physical GPUs: 4 known NVIDIA GeForce RTX 3090 units and 1 known NVIDIA GeForce RTX 3080 unit. Inventory presence is not benchmark verification.

Result evidence: the specific public result record supports the result identity, NVIDIA GeForce RTX 3090, 4.08-second runtime, and Farpy Score 24203.

Configuration and method evidence: the public verified benchmark data record supports NVIDIA GeForce RTX 3090, 24 GB VRAM, Blender, and CUDA/OptiX.

9. Failure and uncertainty handling

• Missing evidence is reported as unavailable rather than converted into a verified estimate.
• Failed verification remains visible and does not become a passing result.
• The published Octane experiment evidence records a rejected performance benchmark and no verified Octane benchmark result.

10. Current limitations

• Public verified hardware coverage remains narrow.
• Not every known GPU has a verified benchmark.
• Workload support is intentionally bounded.
• Experimental results do not automatically graduate into production truth.
• This document describes current public behavior, not every internal mechanism.

11. Vision

A compute network where completed work is measurable, evidence is inspectable, and economic settlement follows proven execution.