streaming-evidence-standards

Streaming Evidence Standards

Purpose: Prove end-to-end streaming chunk delivery from LLM output to HTTP SSE client with timestamped, sequence-level evidence.

Required References

• Bead: BD-iwr
• Roadmap spec: roadmap/2026-02-11-streaming-evidence-reference-standard.md
• PR requirement section: LLM-to-HTTP Chunk Timing Evidence Requirement

Mandatory Evidence (Single Fresh Run)

1. Record identifiers:

• campaign_id
• request_id (or stable equivalent)
• environment/service identifier

2. LLM chunk timeline:

• UTC timestamp for first LLM chunk
• UTC timestamp for every LLM chunk
• sequence index per chunk

3. HTTP SSE timeline:

• UTC timestamp for first HTTP chunk received
• UTC timestamp for every HTTP chunk received
• sequence index per chunk

4. Joined per-sequence table:

• sequence
• llm_ts_utc
• http_ts_utc
• delta_ms = http_ts - llm_ts

5. Summary metrics:

• first-chunk latency summary
• p50/p95/max of delta_ms

6. PASS/FAIL thresholds:

• p95(delta_ms) <= 2000 ms
• max(delta_ms) <= 5000 ms

7. Reproducibility artifacts:

• raw LLM chunk log path
• raw HTTP chunk log path
• joined table/report path

Skeptical Validation Checklist

• Same campaign_id/request_id appears across all artifacts.
• Sequence counts align between LLM and HTTP timelines.
• No skipped/duplicate sequence rows without explanation.
• First chunk is observed in logs (not inferred from request start).
• Reported metrics match the raw joined table.

Failure Conditions

Mark FAIL when any apply:

• Missing per-chunk timestamps on either side
• Missing joined table
• Missing identifiers preventing correlation
• Threshold violations without documented external incident
• Summary claims without raw evidence paths