PiPER contact profiling v0.1 — experimental evidence

Experimental PiPER source patch and complete synthetic profiling evidence. The contact scan uses a per-call scratch force buffer and precomputed local membership sets; every native force query and 1 ms check remains active.

• Four ABBA runs: pooled mean environment-step wall compute 15.70% lower; process CPU 12.60% lower. Two pairs, one preserved synthetic command tape, shared host; policy inference excluded.
• Ten separate diagnostics: 200,010 snapshots exactly matched across five baseline/candidate pairs, including native state, contacts and independently recomputed outcome state.
• The original reader refused three expanded JSON pairs at 64 MiB. Its failed report and the successful supplemental streaming audit of the same files are both retained. No simulation or timing retries.
• Prior held-out policy accuracy remains 5/12. The jaw pulse is equivalence coverage, not recovery validation. No new wheel or production promotion, real-time guarantee, hardware test or arbitrary-video learning.

The evidence ZIP uses content-addressed blobs and includes a guarded verifier/reconstructor, source snapshots, all attempts and licenses. Verify SHA256SUMS before extracting or running the reviewed tools.

Research note · Hugging Face evidence