ArtemisKit 0.6.2 — Independent workflow outcomes

Independent workflow outcomes

ArtemisKit now checks what an agent actually produced before reporting task success. Native
workflows independently verify files, final state, bounded JSON schemas, successful tool traces
and host policy evidence. Confident prose and normal termination do not replace those checks.

• Deterministic assertions require no additional model calls.
• Declared qualitative criteria use an explicitly configured, bounded independent judge. Target and
  judge identities, usage and budgets remain separate; failed deterministic prerequisites skip judging.
• CLI runs distinguish verified passes (0), valid task failures (8) and invalid/unavailable
  evaluation (9). Existing runtime/policy exits remain distinct; evaluation cancellation returns 130.
• SDK sessions accept semanticJudge. Version 2 records expose assertion coverage and task validity;
  readWorkflowRecord validates V1/V2 metadata without rescoring historical runs.
• New guides document judge configuration, supported JSON Schema limits, denominators and CLI/SDK usage.

Milestone 0.6.2 maps to core 0.6.2 and CLI/SDK 0.6.1, with exact dependency-driven package versions
in docs/releases/0.6.2.json. Read docs/releases/0.6.2.md for qualification and publication status.
These notes are prepared for publication; their presence does not establish an npm or GitHub release.

Durable recovery/fault injection remains 0.6.3; professional workflow reports remain 0.6.4.
Hosted-model quality and prompt-injection immunity are not claimed by deterministic fixture tests.