Yang–Baxter GPU Integrability Lab
GPU-resident per-pixel R-matrix computation · independent CPU physics reference · Hilbert-state comparison · TBFT-compatible witness adapter · Merkle receipt
V0.1 · ADDITIVE LOCAL PHYSICS FIXTUREMatched GPU bridge bench + device qualification
Runs the existing KKG-routed Yang–Baxter shader and independent CPU matrix reference repeatedly. GPU timings include dispatch, device execution, projection and readback; neither CPU nor GPU wall timing alone measures shader-only time. Invalid-shear fixtures are rejected; connected TBFT authority remains UNBOUND.
IDLE — local GPU evidence not yet measured on this device.
No benchmark receipt.
GPU pixel-node inspector
Tap a colored node after running the V1 bridge. Inspection uses the GPU readback word and canonical KKG address; unallocated pixels are not KKG nodes.
No bridge GPU state selected.
Device probe: open the eight-gate probe. When served as HTTPS or localhost, the probe can test secure context, storage, user-selected workload files, SHA-256, Canvas2D, CSP denial, WebGPU YBE parity, and reload/reboot persistence. Files are read but not executed by the probe. File-based launches may restrict relative-file opening; ZIP includes both pages.
KKG ⇄ PQS Convergence · V1
Directory → typed GPU allocation → independent YBE calculation → canonical replay commitment → local TBFT-compatible witness. This additive bridge retains the original Yang–Baxter lab unchanged. The default KKG fixture graph is synthetic; real KKG imports must explicitly bind every node to an admitted R-matrix operator.
IDLE — run Bridge compute or qualification; original GPU lab below remains available.
Uses the grid, cycles, and seed controls in the original GPU panel. Admission limits: at most 4,096 active graph nodes for full independent CPU parity. Accepted graph edges are typed provenance/presentation relationships only; unlowered causal edges HALT. GPU assignment and graph identity are bound by SHA-256. Physical panel and external native TBFT authority remain unbound.
Matched workload, identities & resources
Run the bridge to produce evidence.
Bridge tests
Not run.
Local cross-runtime commitments
No local commitment.
Supported graph shape
{"schema":"KKG_PQS_GRAPH_INPUT_V1","nodes":[{"node_id":"CAP::YB::001","operator_id":"R_SWAP_2D","domain":0,"authority":"LOCAL_FIXTURE_ONLY","initial_transforms":0}],"edges":[]}GPU computational surface
Each texel maps to a distinct GPU work item calculating the three-wire braid-form Yang–Baxter residual. Every node retains an 8-byte bounded OmegaNode state across GPU cycles.
No imported evidence verified.
GPU not probed.
Boundary: browser WebGPU computes in GPU storage; it does not convert physical display pixels into independent arithmetic cores. These are GPU-compute nodes indexed by pixel coordinates. The YB fixtures are bounded 4×4 R matrices, not a spectral-parameter integrability theorem. TBFT output is a local schema-compatible adapter, not a call into a connected TBFT runtime.
Qualification tests
| Test ID | Result | Evidence |
|---|
Performance
Measurements separate GPU submission/readback wall time from CPU validation. CPU reference uses the same exact fixtures but a distinct matrix-product implementation. These timings are not a matched-workload GPU speedup ratio.
Mathematical lock
Expected squared residuals from the repository's exact fixtures: swap 0, identity 0, invalid shear 16.
Hilbert probability projection
Selected GPU-computed node
Select a pixel after computation to inspect its GPU-written word, local lineage token and independent Hilbert projection.
TBFT evidence adapter
The existing TBFT_NUMERIC_PHYSICS_STEP_RECORD_V1 contract supplies the record family. This extension emits a separately identified GPU qualification record; remote ledger authority remains unbound.
No evidence receipt.
Evidence provenance and boundaries
Source-locked model inputs: R_MATRIX_NUMERIC_FIXTURE_REGISTRY_V1; R_MATRIX_YB_VALIDATOR_PROFILE_V1; YB_MATRIX_OPERATOR_KERNEL_V1; Hilbert canonical slice.
Original PQS pixel-compute workstation, MC physics solver, Flint IR, Foundation Plate, FIROS, and TBFT repositories are untouched. The source URLs are documentation only; runtime uses zero external resources. Hash commitments authenticate the recorded bytes, not panel photons, attested hardware identity, or a complete physical integrability proof.
Placeholder manifest
PH-YBGPU-L3-FLINT-001 Canonical Flint IR / N16 lowering — UNBOUND PH-YBGPU-L3-SPECTRAL-002 Spectral-parameter Yang–Baxter R(u) and transfer matrix — NOT IMPLEMENTED PH-YBGPU-L3-TBFT-003 Connected native TBFT runtime / remote ledger — UNBOUND PH-YBGPU-L4-PLATE-004 Connected Foundation Plate authorization — UNBOUND PH-YBGPU-L4-SCANOUT-005 Trusted physical display scanout attestation — NOT IMPLEMENTED PH-YBGPU-L4-TIMING-006 Timestamp-query calibrated GPU kernel time — NOT IMPLEMENTED PH-YBGPU-L3-HILBERT-007 Full Hilbert VM adapter including complex vector channels — NOT IMPLEMENTED PH-YBGPU-L4-ARCHIVE-008 Raw all-node snapshot above bounded export quota — EXCLUDED