Human Coder Launch Bundles

scripts/human_reliability/generate_launch_bundle.py creates a coder-facing handoff directory for one approved human reliability cohort. It wraps an already generated blind packet with task overview, training copy, calibration instructions, declarations, allowed-reference rules, return instructions, and a bundle manifest.

Run:

python3 scripts/human_reliability/generate_launch_bundle.py \
  --case <case_id> \
  --cohort cases/<case_id>/quality/human-reliability/cohorts/<cohort>.json

The output is written beneath:

cases/<case_id>/quality/human-reliability/launch-bundles/<cohort-id>-<version>/

Bundle Contents

Each generated bundle contains:

  • README.md: task overview, time estimate, included files, rights/storage note, and packet hash;
  • packet/: blind packet payload, blank CSV/JSON response templates, and packet manifest;
  • training/human-coder-training-guide.md: required training guide copy;
  • references/: coder-facing method and submission-contract references;
  • calibration-instructions.md: source-language calibration requirements;
  • coder-declarations.md: conflict-of-interest, independence, AI-assistance, and source-language competence declarations;
  • allowed-references.md: allowed references and prohibited materials; and
  • return-instructions.md: return, contact, and escalation procedure.

The bundle manifest records hashes for generated and copied files so the distributed handoff can be audited.

Safety Rules

Generation requires:

  • an approved cohort manifest with a current approval hash;
  • a current packet manifest and payload hashes;
  • matching cohort and packet identity;
  • storage_policy: repository_allowed; and
  • ai_assistance_allowed: false.

The generator refuses path members that point at internal answer keys, adjudication, model-reliability, normalized submissions, or accepted-artifact subtrees. It copies packet payloads from the packet manifest and verifies their hashes before writing the launch bundle.