# Runtime validation

This release was gated on successful loading, server startup, and real token
generation. It was not gated on coding-answer quality.

## Structural checks

Run from the release workspace root:

```powershell
$Python = '.\.venv\Scripts\python.exe'
$Validator = '.\tools\validate_rivetcoder_gguf.py'
$Llama = '.\work\llama.cpp-rivetcoder'
$Source = '.\RivetCoder-9B-A4B'
$Package = '.\RivetCoder-9B-A4B-GGUF'

& $Python $Validator "$Package\RivetCoder-9B-A4B-Q8_0.gguf" `
  --llama-cpp-dir $Llama `
  --hf-model $Source `
  --expected-file-type Q8_0 `
  --strict-extra-tensors `
  --report "$Package\provenance\q8-structure.json"

& $Python $Validator "$Package\RivetCoder-9B-A4B-Q4_K_M.gguf" `
  --llama-cpp-dir $Llama `
  --hf-model $Source `
  --expected-file-type Q4_K_M `
  --strict-extra-tensors `
  --report "$Package\provenance\q4km-structure.json"
```

Both reports must contain `"ok": true`, 506 tensors, no missing or extra
tensors, no warnings, and zero routing-control mismatches.

## CLI generation check

After extracting `bundle/rivetcoder-llama-windows-cuda-sm120.zip`:

```powershell
.\llama-cli.exe `
  --model .\RivetCoder-9B-A4B-Q4_K_M.gguf `
  --n-gpu-layers all `
  --ctx-size 512 `
  --batch-size 64 `
  --ubatch-size 64 `
  --predict 16 `
  --prompt 'Reply with OK only.' `
  --temperature 0 `
  --single-turn `
  --simple-io `
  --no-warmup `
  --perf
```

Success requires model load, at least one generated token, and process exit
code `0`. Repeat with the Q8_0 file.

## Server check

```powershell
$server = Start-Process `
  -FilePath '.\llama-server.exe' `
  -ArgumentList @(
    '--model', '.\RivetCoder-9B-A4B-Q4_K_M.gguf',
    '--alias', 'RivetCoder-9B-A4B',
    '--host', '127.0.0.1',
    '--port', '8080',
    '--n-gpu-layers', 'all',
    '--ctx-size', '4096',
    '--parallel', '1',
    '--batch-size', '512',
    '--ubatch-size', '256',
    '--flash-attn', 'on',
    '--reasoning-budget', '0'
  ) `
  -PassThru `
  -WindowStyle Hidden

Invoke-RestMethod http://127.0.0.1:8080/health

$body = @{
  model = 'RivetCoder-9B-A4B'
  messages = @(@{ role = 'user'; content = 'Say OK.' })
  temperature = 0
  max_tokens = 16
} | ConvertTo-Json -Depth 8

Invoke-RestMethod `
  -Method Post `
  -Uri http://127.0.0.1:8080/v1/chat/completions `
  -ContentType application/json `
  -Body $body

Stop-Process -Id $server.Id
```

Success requires `GET /health` to return `{"status":"ok"}`, a non-error
completion response containing generated tokens, and no CUDA/graph/tensor
errors in the server log.

## Recorded results

| Artifact | Load/generate | Structure | SHA-256 |
|---|---|---|---|
| BF16 conversion | pass, exit 0 | pass | `f51dfc1cb82b281156a09360bdbe8fe95734334e01d0d6b6c2df4a27b2619aea` |
| Q8_0 | pass, exit 0 | pass | `6e71dd349df93f734b89fd70f2d0590037ddd3bdc11a4e9533991a0eab47b85e` |
| Q4_K_M | pass, exit 0 | pass | `980adc81f64b52bc71228abd0e2bed02a1b6a00bdda9afbfd6d5e0c124ac5600` |

The BF16 intermediate is intentionally not part of the public quantized
release.