dsfb-gpu-debug-demo 0.1.1

CLI binary for dsfb-gpu-debug: generate fixtures, run CPU/GPU pipelines, compare case files.
Documentation

Open In Colab

dsfb-gpu-debug-demo

dsfb-gpu-debug-demo is the command-line replay surface for DSFB-GPU. The binary is named dsfb-gpu-debug. It generates fixtures, runs the CPU reference path, runs the CUDA path when enabled, compares case files, measures selected pipeline paths, and emits audit artifacts.

This crate is intentionally a delivery crate. It does not own the semantic rules and it does not make the GPU authoritative. It wires the reference court in dsfb-gpu-debug-core to the CUDA evidence factory in dsfb-gpu-debug-cuda so a user can reproduce, compare, and inspect the case-file chain from the terminal.

What

The CLI exposes:

  • generate-fixture: produce the canonical synthetic trace fixture.
  • run-cpu: run the deterministic CPU reference pipeline.
  • run-gpu: run the CUDA path when built with --features cuda.
  • compare: compare two case files and report the verdict.
  • bench: measure selected CPU/GPU pipeline paths.
  • bench-gpu-scale: emit the headline money-table benchmark report used by this workspace's local audit trail.
  • s-real-audit: run the sealed S-REAL audit driver over declared datasets.

Where

This crate lives at crates/dsfb-gpu/crates/dsfb-gpu-debug-demo in the DSFB repository. It depends on:

The wider Atlas crates are dsfb-gpu-atlas-corpus and dsfb-gpu-atlas-registry. The public Colab notebook runs the audit-oriented replay flow on a Colab GPU runtime and reports divergence honestly when bytes do not match the committed seal.

Why

Reproducibility needs an operator surface. A library can define deterministic semantics, but a reviewer still needs commands that build, run, compare, and preserve the artifacts. This crate is that surface: a thin CLI over the core and CUDA crates, with exit codes that automation can distinguish.

Mathematical Contract

The CLI does not invent a separate mathematical model. It invokes the same deterministic pipeline:

trace events
  -> window features
  -> residual Q16.16 fields
  -> drift/slew signs
  -> detector motif masks
  -> consensus axes
  -> candidate intervals
  -> bank-admitted episodes
  -> case-file hash chain

For CPU/GPU comparison, equality is byte-level over the declared case file mode and hash chain. A mismatch is not rounded away as numerical tolerance; it is a verdict condition surfaced by compare.

Code

Show commands:

cargo run -p dsfb-gpu-debug-demo --bin dsfb-gpu-debug -- help

Run the CPU path:

cargo run -p dsfb-gpu-debug-demo --bin dsfb-gpu-debug -- generate-fixture --out target/dsfb-gpu/fixture.json
cargo run -p dsfb-gpu-debug-demo --bin dsfb-gpu-debug -- run-cpu --fixture target/dsfb-gpu/fixture.json --out target/dsfb-gpu/cpu.case.json

Build with CUDA dispatch:

cargo run -p dsfb-gpu-debug-demo --features cuda --bin dsfb-gpu-debug -- run-gpu --fixture target/dsfb-gpu/fixture.json --out target/dsfb-gpu/gpu.case.json

Run tests:

cargo test -p dsfb-gpu-debug-demo

Features

  • default: CPU and audit commands build; GPU commands report GpuError::CudaUnavailable when reached.
  • cuda: forwards to dsfb-gpu-debug-cuda/cuda and requires nvcc plus a compatible CUDA environment at build time.

Claim Boundary

This crate is a CLI and reproducibility wrapper. It does not claim new detector mathematics, learned usefulness, probabilistic inference, medical or safety diagnosis, benchmark portability, or independent semantic authority beyond the core and CUDA crates it invokes.

Publish Order

Publish after both dsfb-gpu-debug-core = 0.1.1 and dsfb-gpu-debug-cuda = 0.1.1 are visible on crates.io.

Citation

de Beer, R. (2026). DSFB-GPU: Clear-Box Pure Deterministic Inference CUDA Acceleration for Replayable Trace-Event Verdicts A Prior-Art Architecture for non-probabilistic, non-stochastic, non-weighted, GPU-Accelerated Residual Signs, Detector Motifs, Bank-Governed Fusion, and Byte-Exact Case Files Without Probabilistic Models (1.1). Zenodo. https://doi.org/10.5281/zenodo.20346478

IP Notice

DSFB-GPU Copyright 2026 Invariant Forge LLC This product includes software developed by Invariant Forge LLC. Apache 2.0 (reference implementation). Background IP: Invariant Forge LLC. Commercial deployment requires separate written license. Contact: licensing@invariantforge.net.