evidence · 3.0 KiB

README.md

Committed at docs/evidence/mainline-init-exec-return/physical-2026-10-05/README.md. This is what grounds the requirements that cite it.

Open raw source

Selected init exec return: physical capture

ONE physical UART capture at revision 69d7cbb3, 2026-10-05 18:30:54–18:34:18 UTC. Total invocation204.576 seconds includes protected preflight/loading; candidate readiness bound180 seconds. Controller exit1: ordinary init login readiness unverified.

Reviewed restored artifact qualification and actual guarded staging passed. This begin ran another complete normal preflight, with exact normal identities/eight boot hashes/three active services/registration absence and staged new metadata/closure. Root exclusively reserved board/UART and ran:

python3 tools/mainline-drm-system-trial.py begin \
  --wait-initramfs-in-initcall --without-boot-markers --init-exec-return \
  --bundle /nix/store/5f2j4y8hy9cclbqjx4nshhxi4jiwz7iq-k230-mainline-drm-trial-boot-files \
  --manifest PRIVATE_MANIFEST --normal-report PRIVATE_NORMAL \
  --state NEW_PRIVATE_STATE --log NEW_PRIVATE_LOG --result NEW_PRIVATE_RESULT

Fresh Linux, exact received323-byte arguments, and the /init announcement preceded ONE valid K230_INIT_EXEC_RETURN_V1 ret=0 record. The parser observed no malformed/duplicate/phase errors. No qualified login or primary prompt appeared within180 seconds. Fixed observations and safe result retain the record and bindings; raw UART/runtime identities remain protected.

The selected kernel exec setup returned successfully. This does not prove the transition to userspace, ELF loader/constructors/systemd main, or that the record’s own printk call returned. Missing later output does not locate a failed instruction or syscall. Adding this optional record may perturb behavior; ordinary root/panel/glass and task5b.5 remain UNVERIFIED.

Complete independent physical review PASS: all five load/CRC pairs, exact arguments, artifact/time/hash bindings and full 73-byte chunked parser replay were checked. The replay consumed all candidate bytes with exactly one valid return record, no readiness, and zero subsequent input. The recovery checker source also passed review and ran only after the subsequent NEW reset confirmation. Bootm is the final command annotation; reviewed source sends no further candidate input on unknown readiness. That is source/transcript evidence, not independent electrical TX proof. The port is released. The subsequent NEW operator reset was consumed once: python3 PRIVATE_RESET_NORMAL_CHECK returned0 after a fresh prompt and full protected normal postflight. A distinct boot, exact normal system/profile/kernel/ init/uname, eight protected boot hashes, three active services and registration absence passed. Recovery receipt binds the private raw/result/ checker/command hashes. Independent complete recovery review PASS. This confirmation check does not claim a passively observed fresh SPL or automatic return. No camera or real-glass acceptance was obtained; ordinary mainline acceptance remains UNVERIFIED.