Skip to content

Research evidence map

Start with the public Circle reference collection: actual proof bytes, stage-by-stage timings, raw logs, machine/job records, external manifest pins and a proof-only verification command. Its L1/C1/J1 runs are kept separate from repeated performance studies.

Use the records below to check model and input identities, measurements and failed attempts behind the reported results.

Current scope

The fixed-case LLM integrity MVP proves a short quantized TinyLlama response and composes exact classifier relations over the same request. It does not claim witness hiding, hardware provenance, arbitrary-model support or independent security qualification.

MVP acceptance matrix · Current state · Chronological ledger

Follow a result to its source

Claim Evidence Scope and limits
Generated-constraint CUDA proves the complete fixed case GPU JSON, protocol and measurements One same-H200 SIMD/CUDA pair; CPU witness generation; retained failed telemetry attempt
Earlier CUDA configuration was slower Fallback-baseline JSON Real proofs, but 1,710 CPU constraint fallback calls
Fresh CPU-only LLM workflow reproduces the response Reproduction JSON, LLM guide Fixed model/quantization/token case, not arbitrary prompts
Dedicated classifier proofs bind one request Circle acceptance JSON, results Synthetic task, caller-enrolled mapping
Released weights match the proved models Publication receipt, model releases No new training or quality evaluation; native inference is not proof generation

H200 timing boundaries

The matched pair records 547.75 s SIMD / 53.94 s CUDA summed proving, and 679.49 / 155.89 s complete segment commands. The ratios are 10.15× and 4.36× respectively.

Both use the same host, inputs, parameters and four Rayon threads. All 69 proof/statement/claim files match. CUDA records zero CPU constraint fallback.

The outer cached-build workflow takes 34 min 47 s, because it includes both backends, grid checks, transfers and fresh CPU/Mac verification. That is a different boundary from 53.94 s. The separate original CPU-only workflow used a different 32-vCPU host; do not use its latency as the GPU comparison baseline.

Full timing/resource table and caveats

Classifier evaluation

The contextual CNN has a prespecified synthetic study with three seeds and 100 test rows per seed. The MLP is an arithmetic/marker fixture with no held-out quality result. The three proof-demo cases are engineering controls, not a new test set.

The source grammar/vocabulary is shared across synthetic splits. The earlier CNN screen failed and is retained. Real-world safety performance has not been evaluated; the reported scores cover the synthetic task. Read the model/data notes and evaluation contract.

Open research and qualification

  • Independent backend/security/hiding qualification and setup/enrollment custody.
  • General prompts/models and a reviewed numerical equivalence boundary.
  • Authentic capture, hardware provenance and the broader proposed hybrid system.

The proposed v1 specification is a design target beyond the completed MVP. Check the supported API table before using an interface from that specification.

Reproduction principles

Freeze model/data/evaluator identities before measurement. Keep calibration and test roles separate. Save raw numbers, hashes, environment and failure records before a narrative. A new candidate does not get to change its evaluator. Operations and CONTRIBUTING explain the work loop.