How it works

From input to verified outcome.

FactumOS is built for determinism. The same snapshot produces the same outputs—every time—because operators, prompts, and policies are versioned and bound to the graph. Short, auditable stages replace improvisation.

Deterministic pathSame snapshot => reproducible outputs

1. Ingest
Replayable step
Normalize sources into snapshots with Merkle proofs.
- Immutable snapshots pin every byte to a Merkle root.
- Residency + consent flags baked into the ingest contract.
- Schema and PII checks run before any operator can read.
2. Build
Replayable step
Chunk, embed, and project into a governed graph.
- Chunkers and embedders are deterministic and versioned.
- Graph edges carry lineage and cost annotations.
- Vector + graph stores stay aligned to the same snapshot.
3. Reason
Replayable step
LLMs become constrained operators inside the graph.
- Policies select operators; prompts are templates, not chat.
- Evidence grounding enforced through graph adjacency.
- If an LLM drifts, outputs are rejected before they can land.
4. Receipt
Replayable step
Record the run with replayable artifacts.
- Receipts capture inputs, edges, and operator attestations.
- Replay: same snapshot → reproducible outputs by design.
- Conflicts logged with severities so humans can intervene calmly.

Stage receipts (all visible)

Each stage renders its receipt slice by default—no hidden accordion.

IngestReceipt visible

Normalize sources into snapshots with Merkle proofs.

Immutable snapshots pin every byte to a Merkle root.
Residency + consent flags baked into the ingest contract.
Schema and PII checks run before any operator can read.

Receipt anchor: Receipt binds source checksum + residency tags before any downstream operator executes.

BuildReceipt visible

Chunk, embed, and project into a governed graph.

Chunkers and embedders are deterministic and versioned.
Graph edges carry lineage and cost annotations.
Vector + graph stores stay aligned to the same snapshot.

Receipt anchor: Replay uses identical chunk + embed versions; graph edges align to the pinned snapshot ID.

ReasonReceipt visible

LLMs become constrained operators inside the graph.

Policies select operators; prompts are templates, not chat.
Evidence grounding enforced through graph adjacency.
If an LLM drifts, outputs are rejected before they can land.

Receipt anchor: Receipts show which operator versions fired and which outputs were rejected before landing.

ReceiptReceipt visible

Record the run with replayable artifacts.

Receipts capture inputs, edges, and operator attestations.
Replay: same snapshot → reproducible outputs by design.
Conflicts logged with severities so humans can intervene calmly.

Receipt anchor: Receipt includes hash, Merkle root, conflicts, and the replay hash used for evidence export.