# Exclusions and Non-Use Decisions

## Raw crash logs

No raw crash logs were copied into the workload. Public/official MCP and OpenRouter material was used only to derive safe synthetic failure patterns. The final prompts do not include real file paths, hostnames, usernames, IP addresses, database names, customer payloads, credentials, tokens, or real PII.

## Uploaded crash logs

No standalone crash-log file was present in `/mnt/data` besides the workload slots, LiteLLM request examples, and markdown instructions. Therefore, the package does not claim to include uploaded crash-log evidence.

## OpenRouter public GitHub issue logs

Last-30-day web search did not surface a reusable public OpenRouter crash-log issue with clear reuse status. OpenRouter official docs and public GitHub organization metadata were used for API-compatible routing, fallback, tool-calling, model metadata, and logging patterns. No private OpenRouter traces or customer logs were used.

## MCP public crash logs

Recent official Microsoft MCP issue signals were used for observability, authorization, JSON Schema tool contracts, deterministic routing, and stateless transport patterns. Raw stack traces were not copied. MCP official debugging/specification docs were used to synthesize invalid-params, permission, oversized-output, and timeout workload cases.

## Gracie / Grace cluster

Gracie/Grace cluster searched as a lead only. No validated primary-source workload with clear ownership, license, relevance to WWT A100 LLM inference, and reusable operational failure pattern was included.

## Rejected source categories

- Unknown license prompt dumps.
- Private or proprietary crash logs without explicit authorization.
- Real credentials, API keys, secrets, customer payloads, or real PII.
- Raw jailbreak collections copied verbatim.
- Sources that cannot be mapped to a Cubie reason code.
- Sources that cannot be mapped to a NIST-aligned rationale.
- Mixed-model batch files that would create deployment/model mismatch during bulk processing.