#Reproduction

This protocol separates reproduction into two acceptance levels. Offline evidence verification runs on any machine. A host matching the fixed hardware and software controls can continue with a performance replay that emits a new environment receipt. Both levels use the same model, prompt, ACL, and output identities.

#Acceptance baseline

#1. Acquire the source and freeze experiment identity

Run in Windows PowerShell:

git clone https://github.com/A3S-Lab/Power.git
Set-Location Power

$powerCommit = (git rev-parse HEAD).Trim()
$dirtyFiles = @(git status --porcelain)
if ($dirtyFiles.Count -ne 0) {
  throw 'A clean worktree is required'
}
$powerCommit

The latest peak capture records clean source revision da2c1dd5a2c6a573ef8be7789de4a67fdb2a0eb0; the active quality matrix records 64aef15ddff7232c6261385700c8a912d1ed0963. Replaying from a newer clean revision is valid, but it is a new experiment and must retain its own evidence.

#2. Run the model-free offline verifiers first

The active Q6_K-only verifier does not load the 22.88 GB model and does not require an NVIDIA GPU. It pins the full compact payload and recomputes the 600 request quality matrix:

py -3.13 .\tools\test_qwen38_q6_quality_evidence.py
py -3.13 .\tools\qwen38_q6_quality_evidence.py verify `
  --evidence .\docs\benchmarks\qwen3.8-27b-q6k-rtx4090\quality\pure-q6-rtx4090-3x.evidence.json `
  --json

It recomputes 23.642 token/s for autoregressive Q6_K and 41.035 token/s for the same Q6_K with full-vocabulary MTP. The optional --require-lossless gate intentionally fails because exact output parity is 50/100 and strict scoring has two losses.

#Verify archived peak and calibration evidence

The broader historical verifier checks 23 file SHA-256 values and recomputes the older peak, mixed-artifact quality, and paired-calibration records:

powershell -NoProfile -ExecutionPolicy Bypass `
  -File .\tools\verify-qwen38-q6k-evidence.ps1 -Json

A passing process exits with code 0 and includes:

{
  "status": "passed",
  "verified_file_hashes": 23,
  "quality": {
    "completed_requests": 900,
    "request_wide_tokens_per_second": 83.22814601950864
  },
  "pure_q6": {
    "full_vocabulary_k7_s7_median": 147.020656574707,
    "prefix_fr8192_k7_s6_median": 176.6108685085471
  },
  "deep_optimization": {
    "peak": {
      "median_decode_tokens_per_second": 172.8353133057359,
      "minimum_decode_tokens_per_second": 171.29810355919784
    },
    "general": {
      "target_only_tokens_per_second": 28.71272184998198,
      "mtp_tokens_per_second": 46.92338764288924,
      "speedup_percent": 63.4236833695329,
      "paired_final_answers": 12,
      "fallback_replays": 0
    }
  }
}

Any changed evidence byte or statistic produces a nonzero exit code and names the mismatched field.

#3. Match the acceptance host

These token/s values are boundaries on this host, not cross-hardware promises:

The same protocol can run after changing the GPU, driver, clock, display load, or CPU topology, but label that result as a new platform. Do not copy an inapplicable affinity mask merely to pass the gate.

#4. Build the pinned CUDA profile

cargo fetch
powershell -NoProfile -ExecutionPolicy Bypass `
  -File .\tools\apply-llamacpp-power-patches.ps1

$env:CMAKE_GENERATOR = 'Ninja'
$env:CMAKE_CUDA_ARCHITECTURES = '89'
cargo build --release --bins `
  --target-dir target-native-sm89-ninja `
  --no-default-features `
  --features llamacpp-cuda,llamacpp-mtp-fr

The patch tool must confirm the binding, MTP/FR, and high-priority CUDA-stream patches. The runner rejects any backend other than exclusive llama.cpp.

#5. Verify the model and inputs

Register the untouched Q6_K artifact as qwen3.8-27b-q6-k, then select its Power data directory. These three repository inputs must match exactly:

$powerHome = 'D:\models\a3s-power\qwen38\power-home'
$manifestPath = Join-Path $powerHome `
  'models\manifests\qwen3.8-27b-q6-k.json'
$manifest = Get-Content -Raw -LiteralPath $manifestPath | ConvertFrom-Json
$model = Get-Item -LiteralPath $manifest.path

if ($model.Length -ne 22884408288) { throw 'Unexpected model size' }
if ((Get-FileHash -Algorithm SHA256 -LiteralPath $model.FullName).Hash -ne
    '562FBF760503008F118E5DF38DE5B3E97992D1F693F475815631198547486727') {
  throw 'Unexpected model hash'
}

#6. Replay the prefix-FR8192 peak end to end

Close unnecessary GPU applications and run in a terminal allowed to lock the GPU clock. This command uses a zero performance threshold to capture the real measurement. Only a separate quiet-host service gate should require every sample to reach 175:

$benchmarkRoot = 'D:\models\a3s-power\qwen38\benchmark'
$powerHome = 'D:\models\a3s-power\qwen38\power-home'

.\tools\run-qwen38-q6k-benchmark.ps1 `
  -Label pure-q6-fr8192-k7s6-b11-cudahigh `
  -Config .\docs\benchmarks\qwen3.8-27b-q6k-rtx4090\pure-q6-mtp7-snap6-fr8192-rtx4090-throughput.acl `
  -PromptFile .\docs\benchmarks\qwen3.8-27b-q6k-rtx4090\prompt.txt `
  -BenchmarkRoot $benchmarkRoot `
  -PowerHome $powerHome `
  -ModelHash 562fbf760503008f118e5df38de5b3e97992d1f693f475815631198547486727 `
  -MaxTokens 1024 -NumBatch 11 -WarmupRuns 1 -Samples 9 `
  -MinimumTokensPerSecond 0 -ProcessPriority High `
  -ProcessorAffinityMask 349525 -LockGpuClockMHz 2745 `
  -CudaHighPriority `
  -MaximumIdleGpuUtilizationPercent 8 -IdleGpuSampleCount 3 `
  -TargetDirectory target-native-sm89-ninja `
  -RequireHighPerformancePowerPlan -RequireCleanTree

The runner restores the GPU clock in finally and retains failed reports, so environment contention can be distinguished from a real performance regression.

#7. Accept the new result

Retain these files under $benchmarkRoot:

• pure-q6-fr8192-k7s6-b11-cudahigh.json: nine raw samples, statistics, and output digest;
• pure-q6-fr8192-k7s6-b11-cudahigh.environment.json: Git, binary, model, ACL, prompt, GPU, process, and power identities;
• pure-q6-fr8192-k7s6-b11-cudahigh.preflight.json: startup identity and host-control checks;
• matching stdout and stderr logs: backend initialization and failure diagnosis.

The capture is valid only when all nine requests generate 1,024 tokens, every output SHA-256 is identical, model identity matches exactly, the backend is exclusive, the worktree is clean, and stream priority, affinity, clock, and power controls actually took effect. The caller decides whether it meets a host-specific deployment SLO; the historical 176.61 result is not an automatic failure threshold for the current shared desktop.

#8. Verify the Q6_K-only native DFlash2 capture

Both paired modes use the same 22.88 GB Q6_K target. The 1.14 GB Q4 DFlash2 artifact is an auxiliary proposer, never a target result. Recompute the native five-sample comparison without a model or GPU:

a3s-power-speculative-bench compare `
  .\docs\benchmarks\qwen3.8-27b-q6k-rtx4090\dflash2\native-target-only.json `
  .\docs\benchmarks\qwen3.8-27b-q6k-rtx4090\dflash2\native-dflash2-k7-s6.json

It verifies 33.075 versus 144.453 token/s median decode, a 4.367x speedup, and exact cross-mode output parity. Median end-to-end throughput is 25.744 versus 63.182 token/s. Verify the separate representative quality package:

py -3.13 .\tools\dflash2_evidence.py verify `
  --evidence .\docs\benchmarks\qwen3.8-27b-q6k-rtx4090\dflash2\evidence.json `
  --json

The historical verifier recomputes 29.702 versus 45.143 token/s on the fixed 12-task workload, 12/12 answer parity, and 7/12 complete-output parity. Adding --require-production-default must fail because the representative output gate is not exact; native execution is available. Use the full DFlash2 guide for clean Power commit 72a1ecd, artifact/runtime hashes, A3S_POWER_HOME registration, CUDA build flags, host controls, and exact paired commands.

#9. Reproduce the native DSpark gate

The external-DSpark package is a separate paired experiment. It keeps the same 22,884,408,288-byte Q6_K target and binds the 1,104,594,816-byte DSpark Q4 artifact with SHA-256 12003c7f2642e2e87e979729e16947a913e2213d82136cb5024a36ec4871fef2. Run the four model-free verifiers first:

powershell -NoProfile -ExecutionPolicy Bypass `
  -File .\tools\verify-dspark-evidence.ps1 -Json

py -3.13 .\tools\qwen38_quality_evidence.py verify `
  --evidence .\docs\benchmarks\qwen3.8-27b-q6k-rtx4090\dspark\quality\evidence.json `
  --json

py -3.13 .\tools\dspark_adaptive_evidence.py verify `
  --evidence .\docs\benchmarks\qwen3.8-27b-q6k-rtx4090\dspark\adaptive\evidence.json `
  --json

py -3.13 .\tools\dspark_quality_followup_evidence.py verify `
  --evidence .\docs\benchmarks\qwen3.8-27b-q6k-rtx4090\dspark\quality\followup-evidence.json `
  --json

The accepted context-512, batch-12 capture reports 32.249 token/s target-only and 169.324 token/s with DSpark K10/S6, with a 167.102 token/s minimum. All three 256-token outputs and receipts match exactly. Peak VRAM is 23,847 MiB, so a quiet device and adequate free memory are required even when GPU utilization is low.

The second verifier authenticates the context-1024, batch-12, 600-request quality capture. It recomputes the 22.618 versus 32.678 token/s workload rates, 1.445x speedup, fixed-task scores, replay telemetry, and all 100 paired task vectors. Adding --require-production-default is expected to fail because only 54/100 complete outputs match; the published K10/S6 matrix is diagnostic, not a lossless default.

The third verifier binds the current clean request-local capture. It recomputes a 164.756 token/s peak median and 160.881 minimum, identical peak output and receipt hashes, zero replay, the 22.872 versus 31.052 token/s quality-workload rates, 1.358x speedup, and all 100 paired task vectors. The candidate recorded five lenient gains and three losses and only 55/100 complete-output parity, so --require-production-default must fail here too.

The fourth verifier binds the clean loss-focused follow-up at commit 7bdeb960f5a38ea7515c67a12636a29198fd95f6. It checks three alternating 512-token repetitions and one 1,024-token pair. Every repetition retained 5/5 paired answers with zero gains and zero losses; the 1,024-token pair completed all five tasks with 5/5 answer parity. The 512-token workload reached 30.521 versus 24.967 token/s. Complete output parity remained 0/5, so adding --require-production-default is expected to fail.

Reproduce those two matrices with the hash-locked dspark/quality/divergence-v1.selection.json file and run-qwen38-quality-matrix.ps1: use -MaxTokensOverride 512 -NumCtx 1024 -Repetitions 3 for the performance follow-up, then -MaxTokensOverride 1024 -NumCtx 2048 -Repetitions 1 for the no-truncation quality check. The exact host controls and complete command are in the follow-up protocol.

The exact performance commands in the package attest the requested 2745 MHz GPU clock lock, high-priority CUDA streams, High process priority, 0x55555 affinity, clean worktree, idle GPU, and at least 23,000 MiB free VRAM. Those controls are evidence from this host, not portable defaults for other CPU or GPU topologies.

Use the DSpark reproduction package for the typed registration body, target-only and DSpark ACL files, raw reports, artifact revisions, and exact paired runner commands. DFlash v1 is not part of that result: DFlash v1 and DSpark are alternative artifact contracts, and no compatible DFlash v1 GGUF has completed this gate. DFlash2 uses the separate native package in step 8.

The complete Windows/CUDA guide also covers the paired full-vocabulary control, the 12-task pure-Q6_K calibration, and the previous mixed-artifact gates. The quality-matrix protocol explains how to rerun the existing 100-task × 3-run evaluation.