See it work

Watch it end to end

Seven minutes, recorded live on an Azure VM against Azure OpenAI with the signing key held in an Azure Key Vault HSM: capture, seal, nova verify, two tamper attempts caught and named, then the same run in the dashboard. The commands and their output are verbatim from that run — re-rendered on screen for legibility, not re-run or edited; the dashboard sections are real screenshots of the same run. The narration is a synthetic voice generated locally with Piper (MIT), not a cloud service.

Interactive demos

One per primitive, each on its own page. They run in your browser on fixtures from the repository and do not talk to a server.

• Asset registry →name@version identity for models, prompts, tools and datasets, with eval-gated promotion.
• Run Capsule →The files one captured run leaves behind, validated in your browser against the shipped schema.
• Replay & diff →What each of the five replay modes does and does not do, and a structural diff of two runs.
• Lineage graph →Provenance, blast radius and replay chain over a small graph of runs, assets and artifacts.
• Evidence Bundle →A signed bundle manifest, verified with real Ed25519 in your browser. Flip a byte and it fails.

NovaFabric wraps a process. Not a framework integration, not a decorator you add to every call site — a wrapper around whatever you already run. A shell script, a notebook cell, an agent, a SLURM training job.

The run exits with its own exit code. If NovaFabric fails, your workload still runs — never blocking the user's work is an architectural invariant, not a setting.

Below is the manifest from that capture, validated against the real JSON Schema 2020-12 file from the repository. Ajv compiles the schema in your browser and reports what it finds.

A green check a website could have hard-coded proves nothing, so: break it. Delete a required field, change a type, put a nonsense value in an enum. If the check stayed green, this demo would be worthless — and for a project selling verifiable evidence, an unfalsifiable demo would be the wrong thing to ship.

{
  "schema_version": "0.1.0",
  "novafabric_version": "0.6.12",
  "run_id": "01KR5SQZPDGTKE3MDP3ZRX8WP1",
  "created_at": "2026-04-15T10:00:00+00:00",
  "finished_at": "2026-04-15T10:00:04+00:00",
  "duration_ms": 4127,
  "command": [
    "python",
    "review.py",
    "--diff",
    "examples/sample.diff"
  ],
  "exit_code": 0,
  "status": "success",
  "capture_mode": "cli-wrapper",
  "working_directory": "~/projects/code-review-agent",
  "host": {
    "arch": "x86_64",
    "os": "linux",
    "python": "3.12.4",
    "cpu_count": 16,
    "memory_bytes": 67108864000,
    "gpu": [],
    "hostname_redacted": true
  },
  "trace_ref": "trace.jsonl",
  "trace_root_span_id": "0123456789abcdef",
  "model_calls_ref": "model-calls.jsonl",
  "tool_calls_ref": "tool-calls.jsonl",
  "assets_ref": "assets.jsonl",
  "environment_ref": "env.lock",
  "replay_policy_ref": "replay.yaml",
  "redaction_proof_ref": "redaction-proof.json",
  "lineage_ref": "lineage.jsonl",
  "model_call_count": 2,
  "tool_call_count": 2,
  "mutating_tool_count": 0,
  "inputs": [
    {
      "name": "diff",
      "path": "inputs/sample.diff",
      "content_hash": "sha256:2a8439e85cae440ff7e28be7009b7056ccff0597d706a093a7e0c44e90f022b7",
      "size_bytes": 1240,
      "media_type": "text/x-diff"
    }
  ],
  "outputs": [
    {
      "name": "review",
      "path": "outputs/review.md",
      "content_hash": "sha256:5d93970a885c44d5babc3cf3dabcbcbd275e1c3df1fdd020f76126bc14bf5819",
      "size_bytes": 612,
      "media_type": "text/markdown"
    }
  ],
  "metadata": {
    "agent": "code-review-agent@0.1.0",
    "prompt": "code-review-prompt@0.1.0"
  }
}

The same check runs locally as nova validate <run-id>.

A replay inspects the capsule, or re-runs it with supported model calls served from the recording, and writes a replay result. The experimental intervention mode also writes a counterfactual capsule you can diff against the original. There are five modes; what each one does is below.

Something changed between Tuesday and Wednesday and the git diff is empty. Structural diff compares two capsules — outputs, environment, recorded files — rather than diffing log text.

A directed provenance graph over runs, assets, and artifacts. It answers the three questions people actually ask:

• Provenance — what produced this artifact?
• Blast radius — if this dataset was wrong, what else is suspect?
• Replay chain — what must be re-run to regenerate this?

SQLite by default. Kuzu, Postgres, AGE, and JanusGraph backends exist for cluster scale and are experimental.

The payoff. A signed bundle an auditor verifies offline, with no NovaFabric installed — only sha256sum and an ed25519 verifier. Evidence only its own tool can check is not evidence; it is a database row. An RFC 3161 timestamp is opt-in (--timestamp).