GLM completion-budget evidence

Decision: keep 4,096; B is tested and unpromoted. Summary, native panel records, validity ledger, identity, plan and frozen prompts, upstream request metadata, boundary smokes, restoration, friction.

Raw full Pi event streams and original private snapshots remain on the operator's evidence disk. Public copies omit container ID and host PID from native result health records; no grading fields or measurements changed. Private topology and credential-bearing configuration are not published. Request metadata contains no auth headers. The result format is the bounded Pi validation harness's native format, not a fabricated product benchmark job. The eight invalid coding records retain their original grades but MUST be excluded using validity.json.

This publication bundle was finalized after execution from retained files; it does not imply the common public ledger existed before measurement. Every retained public artifact is under 2 MiB.

Recheck the published arithmetic with python3 verify-summary.py from this directory.