Model Lineage From Forecast to Defensible Output

A forecast without lineage is a forecast you can't defend. Lineage — what model, what data, what version, what decision — is what turns an output into a record a regulator can read.

Lineage is the part of AI work that demos never show and audits always demand. It's the chain from the forecast back to the model, the data, and the version that produced it — and without it, every output is an unexplained number.

Issue-tree what a defensible output needs:

• **Model.** Which model, which version, which configuration.
• **Data.** What the model saw, at what point in time.
• **Decision.** What the model decided, and the confidence.
• **Manifest.** All of the above, attached to the output.

Without the chain, the output is an orphan. With it, the output is a record.

Trust is built on reconstructability. A regulator's question — "why did the model decide this?" — is answered by the lineage, not by the model. The firm that can reconstruct the decision six months later is the firm that can defend it; the firm that can't is the firm that settles.

• The model is updated and the old outputs become unexplainable.
• The data changed and nobody recorded what the model saw at the time.
• The version is gone and the output can't be reproduced.
• The manifest was never attached, so the chain was never there.

Every Deino output carries an audit manifest — model, data, version, decision, lineage — attached to the output, encoded in the Rust workspace. The four invariants refuse to relax; the lineage is load-bearing, not optional. The firm reconstructs and defends from the manifest.

Model lineage is what turns a forecast into a defensible output. Attach the manifest — model, data, version, decision — and the output becomes a record. Skip it, and every output is an unexplained number.