# WWT Cubie A100 Demo Workload - 1,500 Request Package

This package converts the uploaded WWT 1,500-slot structure into a synthetic, sourced, WWT-safe prompt workload and a separate LiteLLM/OpenAI-compatible runtime JSONL package.

## Files

| File | Purpose |
|---|---|
| `blended_prompt_set_1500.jsonl` | Audit/review file with prompts, source IDs, NIST mapping, Cubie expectations, and expected Llama behavior. |
| `litellm_batch_requests_1500_llama8b.jsonl` | Runtime batch file: one `/v1/chat/completions` request per JSONL line. |
| `litellm_batches_llama8b/` | 15 split batch files with 100 requests each. |
| `runtime_batch_manifest.csv` | Batch split counts and validation flags. |
| `source_register.csv` | Official/public sources used for pattern evidence and governance mapping. |
| `mcp_openrouter_crash_patterns.jsonl` | Sanitized MCP/OpenRouter/Azure failure-pattern register. |
| `candidate_workloads.jsonl` | One row per candidate prompt pattern/subcategory. |
| `nist_mapping.csv` | Prompt-level NIST mapping. |
| `run_manifest.yaml` | Replay/run-shape manifest. |
| `bulk_request_validation_report.json` | Counts and schema validation summary. |
| `bulk_processing_notes.md` | Submission notes and operational constraints. |
| `exclusions.md` | Rejected/omitted source categories and why. |

## Target distribution

- 600 clean enterprise prompts
- 300 token-bloat / long-context prompts
- 225 retry-loop / duplicate prompts
- 150 unauthorized tool/scope prompts
- 150 prompt-injection / PII / exfiltration prompts
- 75 malformed / doomed prompts

Total: 1,500 prompts.

## Safety stance

All prompts are synthetic and sanitized. Any PII-like values are fake and marked synthetic. No real credentials, real PII, private crash logs, or customer payloads are included.