This is an audit bundle for the complete v1.5 candidate study, not a qualified buying recipe. Kingy1.5 failed the original frozen conservative supported-answer threshold in all three runs. The current download remains workflow1.4. No fresh 1.5 installation is claimed.

All 240 original attempts and all six model sessions are retained. Taskpack/source/method bytes identify what was tested. The independent synthetic corpus was unseen before measurement and is public/known now. References were used only after measurement; a rerun is reproduction, not unseen qualification.

The original automatic scorer could not ground native chunks spanning PDF pages. A separate same-project unblinded review resolves exact source ranges, credits valid prose and headings, and retains unsupported citations, missing facts and uncertainty. It does not overwrite the frozen grades or grant qualification. Native runtime egress enforcement, novice installation, cloud quality/cost, wall power and hardware upgrades are not established.

Frozen scripts retain their original measurement-workspace paths. These files preserve the audit trail and are not a turnkey native GUI test runner. Model weights and runtime binaries are excluded. Candidate code: MIT; synthetic fixtures: CC0-1.0; Qwen model: Apache-2.0.
