Swift TP1 qualification evidence

This bundle contains redacted derivatives of retained qualification snapshots. Original native artifacts remain authoritative and private; the redaction ledger binds original and derivative hashes. It makes no claim about current deployment, routing, aliases or clients.

The tested BF16/FP8-KV SGLang/MTP3 recipe passed the bounded qualification gates. The fresh ten-cell capacity population (72 requests) is separate from the historical two-arm comparison (432 requests). C1–C8 were matched historically; MTP C16 had an active cap of nine versus sixteen without speculation. Strict formatting failed 4/4, R5 violated the reserve policy without OOM, and the final 32K/C2 decode median missed the soft 15 tokens/s target. SWE resolved 4/5.

• Dated finding
• Manifest and file hashes
• Dated sources
• Measured results and decision
• Evidence summary
• Historical frozen plan
• Configuration snapshot
• Final capacity
• Resource validation
• Historical comparison chart
• Failures and friction
• Limits and withheld records

Snapshot states such as running describe the measurement time only. Operational custody and post-campaign restoration decisions are withheld.