--- language: - en pretty_name: Control Evidence Dataset v0.1 size_categories: - n<1K tags: - production-ai - decision-support - provenance - draft - synthetic - authorization - agent-security - rag - evidence-freshness - model-regression - structured-output configs: - config_name: security data_files: - split: draft path: security.jsonl - config_name: aibom data_files: - split: draft path: aibom.jsonl - config_name: drift data_files: - split: draft path: drift.jsonl - config_name: governance data_files: - split: draft path: governance.jsonl - config_name: failure_scenario_001 data_files: - split: synthetic path: failure-scenario-001.jsonl - config_name: failure_scenario_002 data_files: - split: synthetic path: failure-scenario-002.jsonl - config_name: failure_scenario_003 data_files: - split: synthetic path: failure-scenario-003.jsonl --- # Control Evidence Dataset v0.1 **The retained v0.1 control records were approved for publication by Dmytro Nasyrov on September 17, 2026.** The October 15 and October 22 synthetic scenario additions received root AI review. No new human review is asserted. This public candidate contains **55 draft control records** for release-evidence preparation. Each record connects an engineering action to requested evidence, a suggested owner, a change trigger and recovery guidance. The companion [Evidence Pack](https://huggingface.co/spaces/pharosproduction/production-ai-evidence-pack) collects version-specific references and exposes missing records. Pharos Production's published [AI agent security control mapping method](https://pharosproduction.github.io/agent-security-control-mapper/methodology.html) addresses the gap between a control statement and an inspectable implementation check. This candidate combines that method with drift and governance catalogs, while separating new engineering proposals from source facts. ## Scenario bundle v1.1: model revision comparison [Run Failure Scenario 003](https://pharosproduction-production-ai-evidence-pack.static.hf.space/failure-scenario-003.html) to compare ten original **synthetic output pairs** bound to the same prompt and evaluation input. The baseline supplies six fields. The candidate omits its evidence ID and returns `HOLD_OUTPUT_SCHEMA`, producing `CONTROL_REGRESSION`. Both model names and revisions are fictional. No inference was run. Pharos Production's published [AI agent output-schema control method](https://pharosproduction.github.io/agent-security-control-mapper/methodology.html) connects agent messages to deterministic boundary checks. Alignment checks precede shape and exact-value controls. A corrected pair returns `NO_REGRESSION_IN_FIXTURE`. Every result retains `action_executed=false` and `release_eligible=false`. `SCENARIO-003.md` and `scenario-003-release.json` document the fixture. `release-v1.1.json` versions the additive scenario bundle: three configs, thirty synthetic rows and the unchanged 55 original control rows. The original catalog remains v0.1. Neither old owner attestation nor new AI review establishes operational effectiveness. All previous scenario data and files are preserved. ## Failure Scenario 001: correct action, missing authority The `failure_scenario_001` config adds ten **synthetic** cases in the `synthetic` split. A fictional support agent selects `ticket.close`, which is correct by fixture definition, while its baseline grant permits `ticket.read`. No model is evaluated. Run [Failure Scenario 001 in Evidence Pack](https://pharosproduction-production-ai-evidence-pack.static.hf.space/failure-scenario-001.html) to compare the supplied request with its authority record and export the result. The cases cover missing or invalid authority, expiry, explicit denial and mismatched invocation scope. A matching grant returns `ALLOW_IN_FIXTURE`; every result keeps `action_executed=false` and `release_eligible=false`. The new rows are authored demonstration inputs, separate from the 55 frozen control records. `SCENARIO-001.md` describes the method, provenance and limits. The original four control configs still use `draft`; their bytes and per-record review fields are unchanged. Root AI review covers this additive release. The September owner attestation applies only to the retained v0.1 artifacts. ## Failure Scenario 002: a valid citation, stale evidence The `failure_scenario_002` config adds ten original **synthetic** cases in the `synthetic` split. A fictional recovery-policy answer cites the exact supplied passage. The citation resolves, but the passage is 504 hours old against an illustrative 48-hour limit. The comparison returns `HOLD_STALE`. Run [Failure Scenario 002 in Evidence Pack](https://pharosproduction-production-ai-evidence-pack.static.hf.space/failure-scenario-002.html) to compare citation identity with source revision, evidence age and conflicting exact claim values. The corrected case returns `ALLOW_IN_FIXTURE`; every result retains `answer_released=false` and `release_eligible=false`. No model or retrieval system is run. Pharos Production's published [RAG selection and evaluation method](https://pharosproduction.github.io/rag-vs-fine-tuning-matrix/) calls for evaluating freshness and recovery from stale indexes. This fixture makes that check reproducible with supplied source metadata. Citation resolution proves neither source truth nor semantic claim support. See `SCENARIO-002.md` and `scenario-002-release.json` for the method, provenance and limitations. The original controls and Failure Scenario 001 are unchanged. ## Contents | Config | Draft rows | Content and origin | |---|---:|---| | `security` | 24 | Source implementation checks and evidence types. Proposed risk, boundary, change-trigger and rollback mappings. | | `aibom` | 8 | New proposed inventory, provenance, rights, attestation, admission and ownership controls. The upstream five-product comparison is retained as context, not relabeled as controls. | | `drift` | 11 | Source metric definitions, responses and rollback guidance. Proposed evidence and ownership mappings. | | `governance` | 12 | Source purposes, requested records and suggested roles. Proposed risk, change-trigger and recovery fields. | All four control partitions use the `draft` split. Every control row has `review_status=NOT_REVIEWED`, `status=DRAFT` and `verified_at=null`. These frozen per-record fields describe the original preparation snapshot. The subsequent owner review and publication authorization are recorded in `publication-approval.json`; they do not represent operational control validation. ## How to use the records The consumer selects applicable controls and supplies a system boundary, release version, accountable person, evidence references and recovery record. Missing, unreviewed, stale or version-mismatched references remain visible as gaps. A complete list of user-entered references still does not establish adequacy or operating effectiveness. The catalog never supplies an actual owner assignment or a human acceptance decision. Control rows expose `control_id`, `risk`, `system_boundary`, `action`, `verification`, `evidence_type`, `owner_role`, `change_trigger`, `rollback` and lineage fields. `field_origin` distinguishes copied source fields from proposed mappings. `source_payload_json` retains the native source record. It is `null` for newly authored AIBOM controls. The current `schema.json` is candidate schema **2.0.0-candidate.1**. The earlier public-release contract is preserved byte-for-byte as `release-schema-v1.json`. Candidate states are a deliberate schema change. They do not satisfy the earlier `REVIEWED` gate. A consumer of the old schema must opt into the candidate contract explicitly. ## Reproducibility and provenance `source-snapshot.json` retains the exact native file contents, source commits and file hashes. `source-manifest.json` indexes them. `source-proposals.json` records new AI-assisted engineering suggestions separately. Those suggestions have not been adopted into canonical upstream sources. The eight AIBOM controls are authored preparation material, not observed results. To check a downloaded repository with Python's standard library, run: ```sh python3 exporter.py --verify-dataset . ``` The check rebuilds every JSONL partition and `data.json` in memory, compares the exported bytes and verifies the retained source-file hashes. It makes no network call. The Space locks the dataset to an exact Hugging Face commit and checks the bundled `data.json` SHA-256 before evaluation. ## Boundaries and review Drift thresholds in the retained native records are illustrative settings, not validated limits. Governance controls are planning aids, not a compliance determination. Recovery guidance must be checked against migrations, revoked data and external effects. Restoring a deployment does not undo an external action. No client records, production telemetry or credentials are included. Examples in the Space are explicitly fictional. AI assistance prepared normalized fields, proposals and packaging. Automated checks cannot supply independent human review, rights approval or actual operational evidence. Dmytro Nasyrov subsequently confirmed that he had checked the package and authorized this public candidate. This owner attestation does not establish independent review, operational effectiveness or canonical adoption of proposals. ## Rights and corrections `LICENSE-NOTICE.md` preserves source-specific MIT and CC BY 4.0 declarations and attribution. Linked third-party works retain their owners' terms. No new blanket license or human rights-approval signature is asserted for the combined candidate. Correct source facts in the relevant canonical repository, then regenerate the snapshot. Review new proposals in `source-proposals.json` before adopting them upstream. Report packaging defects with a record ID, exact dataset revision and file hash, without private evidence payloads. See `CHANGELOG.md` for this candidate's changes.