sei.lock — drift state

✅ Stable

sei.lock is the freshness anchor file the engine writes to the repository root after every sei run. It holds the digests of pipeline inputs at the time of the last recompute. sei status reads it without running any script to determine which sections have become stale.

---

What sei.lock anchors

The core of the file is the anchored triple (triple), which deterministically identifies the state of the repository at the time of recompute:

triple field	What it measures	How it is computed
code	Code version	SHA-1 of HEAD (via gix, in-process)
model	(reserved)	Empty in GitDvcVersionResolver (model: String::new()); the model version is anchored indirectly via pipeline_lock_digest (dvc.lock/metrics)
dataset	Dataset version	Composed SHA-256 of DVC hashes from data/*.dvc

In addition to the triple, the pipeline_lock_digest field anchors the state of the reproducible pipeline:

• DVC backend (cat. 1): SHA-256 of dvc.lock after dvc repro — anchors deps + params + outs of the recompute.
• MLflow (cat. 2): Canonical SHA-256 of the run metrics (content-addressed).
• Dagster (cat. 3): SHA-256 of the metrics file after asset materialisation.

The pipeline_lock_digest is included in the signed evidence bundle (EvidenceBundle.pipeline_lock_digest) as provenance of the recompute (Annex IV §1).

---

File structure

The schema lives in seigarrena-core/src/lock.rs. The following is an illustrative shape of a real sei.lock (hash values are fictitious):

sei.lock (illustrative shape)

{
  "schema_version": "seigarrena.dev/v1alpha1",
  "program_uid": "550e8400-e29b-41d4-a716-446655440000",
  "triple": {
    "code": "4280c0f1a2b3c4d5e6f7a8b9c0d1e2f3a4b5c6d7",
    "model": "",
    "dataset": "dvc:a1b2c3d4e5f6..."
  },
  "sections": {
    "classification": {
      "status": "valid",
      "input_hash": "sha256:abc123...",
      "edge_type": "scope"
    },
    "measurement_data": {
      "status": "valid",
      "input_hash": "sha256:def456...",
      "edge_type": "data"
    },
    "measurement_model": {
      "status": "valid",
      "input_hash": "sha256:789abc...",
      "edge_type": "impl"
    }
  },
  "signed_bundle_digest": "sha256:deadbeef..."
}

Top-level fields

Field	Type	Description
schema_version	string	Lock schema version (seigarrena.dev/v1alpha1).
program_uid	string	UUID of the risk programme (the risk section of sei.yaml) at run time, derived from the system name (anchors the normative context).
triple	object	Anchored triple (code, model, dataset) — see table above.
sections	map	Typed dependency graph: each section has its input_hash, status, and edge_type.
signed_bundle_digest	string	SHA-256 of the signed evidence bundle emitted by this run (None on the pipeline-less path).

sections entries (LockEntry)

Each entry in the sections map has three fields:

Field	Type	Values	Description
status	string	"valid"	Section status at run time.
input_hash	string	sha256:…	Hash of this section’s inputs. When it changes, the section becomes stale.
edge_type	string	"scope", "impl", "data"	Edge type in the dependency graph (see below).

Edge types (EdgeType)

The edge_type field determines which changes trigger recomputation of each section:

edge_type	Rust value	What triggers recomputation
scope	EdgeType::Scope	Change to system purpose or risk triage (Drift A). Risk classification must be redone.
impl	EdgeType::Impl	Change to the evaluation script or model (Drift B). Re-measure only; no re-triage if purpose is unchanged.
data	EdgeType::Data	Change to the dataset (Drift C, EU AI Act Art. 10). Only data-dependent measurements are recomputed.

This typology maps directly to drift classes A, B and C. See Typed drift for the full model description.

---

How sei status detects drift without recomputing

sei status compares current hashes against those stored in sei.lock without running any script:

1. Resolves the current repository triple (git HEAD + DVC hashes; the model field stays reserved/empty — the model version is reflected in pipeline_lock_digest).
2. Compares each input_hash in sections against the current hash of the corresponding input.
3. Sections whose hash has changed are marked Stale { reason, change_class } with the corresponding drift class.
4. Unchanged sections are marked Reused.

The result is a per-section drift report with its class (A, B or C) and the reason for the change. sei status returns exit ≠ 0 if any section is stale (freshness gate) or if any blocking control in the last bundle fails (risk gate).

# Detect drift without recomputing — useful in CI/CD as a fast gate
sei status

Staleness is selective: if only the dataset changes (edge_type: data), only data-measurement sections are recomputed on the next sei run; system classification and dataset-independent measurements are reused.

---

Canonical serialisation

sei.lock uses an ordered BTreeMap and deterministic serialisation: the same repository state always produces the same bytes. This ensures that signed digests are reproducible across machines.

---

Relationship with the evidence bundle

sei.lock and .sei/bundle.json are complementary:

• sei.lock is the freshness anchor: it tells you when to re-run.
• .sei/bundle.json is the signed evidence: it contains the results of the last run, control verdicts, and the pipeline_lock_digest.

The signed_bundle_digest field in sei.lock links the two artefacts: it is the SHA-256 of the bundle written during the same sei run.

See also: sei CLI Reference, The .sei/* artefacts.