privacy/workstreams/esg-reporting-assurance.json

{
  "schema_version": 3,
  "workstream": "esg-reporting-assurance",
  "display_name": "Fascicolo ESG",
  "role": "workflow",
  "governed_paths": [
    ".codex-plugin/plugin.json",
    "skills",
    "scripts",
    "schemas",
    "config",
    "assets",
    "requirements.txt"
  ],
  "governed_shared_paths": [
    "vendor/modules/vera_assurance"
  ],
  "runtime_profiles": [
    "openai-codex",
    "anthropic-cowork"
  ],
  "model_context": {
    "policy": "real_case_data_may_enter_selected_runtime_model_context",
    "classes": [
      {
        "id": "esg-evidence",
        "purpose": "Interpret the selected evidence and proposed observations.",
        "content": "Client and engagement IDs, reporting period, jurisdiction, service, source paths and hashes, CSV cells or text lines, units, supplied numeric observations, missing-data reasons and optional disclosure/metric labels. Local Python reads the entire selected CSV/TXT/Markdown input up to 8 MB. Host-read original documents and helper output may enter model context; no automatic anonymization.",
        "runtime_profiles": [
          "openai-codex",
          "anthropic-cowork"
        ]
      },
      {
        "id": "esg-decisions",
        "purpose": "Explain changes and prepare partial drafts for professional review.",
        "content": "Source titles, URLs, versions and unapproved catalogue status; declared reviewer names, dates, decisions and rationales; exact references, full retained history, stale status and draft text. The script does not call a model. Actual host reads must be recorded by Vera separately; the helper cannot attest to them.",
        "runtime_profiles": [
          "openai-codex",
          "anthropic-cowork"
        ]
      }
    ]
  },
  "external_boundaries": [],
  "security_controls": [
    {
      "id": "archive-isolation",
      "control": "Every operation revalidates a portable v2 Studio Archive context and exact selected input bytes. Mutations require running status; source binding accepts only selected IDs. Cross-client and cross-engagement references fail."
    },
    {
      "id": "version-integrity",
      "control": "State and immutable entity hashes are checked on resume, exact reference versions are resolved and historical inputs remain selected. Changed evidence makes dependent decisions and drafts stale. Expected-state digests reject stale writers; a single-writer lock prevents overlapping mutations."
    },
    {
      "id": "closed-output-paths",
      "control": "Outputs use fixed state filenames and hash-derived draft filenames in the verified run output directory. Traversal and symlinks are rejected and draft bytes are verified before reuse."
    }
  ],
  "review": {
    "reviewed_at": "2026-09-29",
    "reviewed_by": "privacy-surface-review",
    "basis": "external_boundary_review_of_workflow_source",
    "source_fingerprint": "cb6df493287a291a4ecb0a1e8aabc7fb572f73e644c40220a0e40c9238053ec1"
  }
}