Evidence And Claims

PermeantOS claims are evidence-bound. Each public claim should map to a proof report, repeatable command, CI job, and explicit limitation.

Public Evidence Index

The source repository includes:

• docs/evidence-index.md
• docs/evidence-index.json

The index currently covers:

• Qwen2.5 MLX to AWS vLLM long-horizon round trip.
• TinyLlama MLX to AWS vLLM raw-transfer structural E2E.
• Qwen2.5 MLX to llama.cpp canonical KV feed.
• Agent Memory Graph v0 schema and conformance.

Current Claim Boundary

PermeantOS has shown that agents can move on validated paths. That does not mean every model, tokenizer, attention layout, or runtime pair is supported.

The strongest current path is Qwen2.5 from local Apple Silicon MLX to AWS vLLM with production wss:///mTLS transport, Agent Memory Graph binding, reverse runtime import, and origin return-home continuation.

Validation Matrix

Use the repository validation matrix for planned and validated runtime/model profiles:

docs/model-runtime-validation-matrix.md

Evidence Jobs

The repository includes a scheduled/manual Evidence Jobs workflow. The default lane is non-provisioning: it runs adapter conformance checks, regenerates evidence reports, emits validation plans, runs AWS preflight with cloud/source checks explicitly skipped, and uploads a machine-readable report.

The real AWS lane is manual-only and requires explicit cost confirmation, a protected environment, and a self-hosted evidence runner.

Real-Runtime Evidence Reports

• AWS long-horizon Qwen2.5 proof
• AWS TinyLlama proof
• MLX-to-llama.cpp canonical KV proof