CinvanaAI
← All projects

Workbench · Data & provenance

Atlas Kernel

Take an evidence-backed graph proposal through work packets, review and a recorded commit.

Repository: CinvanaAI/atlas-kernel

Recorded synthetic example. Source with a complete offline example.

See the idea in action.

A finding without its evidence edge is rejected. The corrected proposal is committed, and the work packet is completed separately.

Example input
A synthetic benchmark report: 18 of 20 checks passed.
Captured result
{
  "source": {
    "id": "example-source",
    "source_id": "synthetic-benchmark",
    "fragment_index": 0,
    "content": "The synthetic benchmark completed 18 of 20 checks in the recorded run.",
    "content_hash": "769f4082d6de1dff41ff6ea8951bffb2bf9fbd4e900b2a7c3baa5cac1e5d5078",
    "source_type": "benchmark_report",
    "created_at": "2026-01-01T00:00:00Z",
    "metadata": {
      "confidence": 0.9
    }
  },
  "packet": {
    "detected_type": "research_claim",
    "required_checkpoints": [
      "verify_evidence_link.v1"
    ],
    "required_output_schema": "FindingNode"
  },
  "rejected": {
    "notes": "verify_evidence_link.v1: FAIL - No 'derived_from' edge found in proposed_edges \u2014 evidence link required",
    "status": "rejected",
    "graph_writes": 0
  },
  "accepted_proposal": {
    "nodes": [
      {
        "id": "example-finding",
        "type": "finding",
        "name": "Synthetic benchmark pass rate",
        "parent_path": "Research Atlas > Research Notes",
        "summary": "18 / 20 = 90 percent.",
        "confidence": 0.9,
        "domain_pack": "research_notes",
        "created_at": "2026-01-01T00:00:00Z",
        "updated_at": "2026-01-01T00:00:00Z",
        "status": "active"
      }
    ],
    "edges": [
      {
        "id": "example-evidence-edge",
        "source_node_id": "example-finding",
        "target_node_id": "example-source",
        "type": "derived_from",
        "summary": "Evidence is the synthetic benchmark fragment.",
        "created_at": "2026-01-01T00:00:00Z",
        "updated_at": "2026-01-01T00:00:00Z",
        "status": "active",
        "confidence": 1.0,
        "domain_pack": "",
        "label": ""
      }
    ],
    "explanation": "Eighteen divided by twenty is 0.9."
  },
  "verification": {
    "passed": true,
    "notes": "verify_evidence_link.v1: PASS - conditions satisfied"
  },
  "commit": {
    "nodes": [
      "example-finding"
    ],
    "edges": [
      "example-evidence-edge"
    ],
    "candidate_status": "committed"
  },
  "packet_status_after_commit": "in_progress",
  "packet_status_after_explicit_completion": "completed",
  "network_calls": 0
}

Try the example.

From the repository root, follow the dependency requirements in the README. This example uses supplied synthetic material.

python -m pip install -e .
python -m atlas_kernel.walkthrough

A finding without its evidence edge is rejected. The corrected proposal is committed, and the work packet is completed separately.

Complete setup and instructions ↗

The interesting part.

Domain meaning is supplied by JSON packs while packet, commit, storage, and rendering mechanisms remain reusable.

Integrated source snapshot

Where it came from.

Generic kernel of a larger private atlas and memory-system experiment.

Retains graph, work packets, checkpoints, commits and evidence; uses a synthetic domain pack and fixes SQLite cleanup.

The private atlas, memory data and adjacent application layers are outside this snapshot.

Source ↗

Follow the family: Atlas kernel and domain-pack authoring

  • Focused extractionAtlas Domain Pack SDK

    Inspect or reuse this boundary without navigating the whole application. The full kernel retains the validator and stays self-contained.

    Source ↗