Architecture

DeciShift compares the decision transition produced by a versioned structured decision system on the same historical records.

Two execution abstractions coexist.

DecisionPipeline — simple backwards-compatible path

historical records
  -> features
  -> model
  -> calibration
  -> threshold/policy
  -> deterministic rules
  -> binary final decision

DecisionPipeline remains the v0.1/v0.2 public API. Its binary comparison schema, imports, exact/approximate attribution and interaction behavior remain supported.

DecisionFlow — composable advanced path

features
├── risk_model ─┐
└── value_model ├── policy -> rules -> final_action

DecisionFlow is a deterministic row-aligned DAG. It supports branching, merging, multiple models/policies/rules and categorical final actions. It is intentionally not a generic workflow orchestrator.

Nodes execute locally in deterministic topological order. Dependency-aware cache keys include node execution identity and upstream lineage, allowing unchanged branches to be reused while a changed node and its descendants are recomputed.

Baseline and candidate flows may have different topology for independent observed-action comparison. Hybrid node attribution is available only when node names, dependency edges and final node are compatible.

Evidence layers

comparison
  -> attribution + sampling diagnostics
  -> structural impact / transitions
  -> cohorts / optional explicit outcomes / optional fragility
  -> saved evidence bundle
  -> manifest + SHA-256 integrity root
  -> optional Decision Contract gate

DecisionPipeline evidence remains schema 1.0. DecisionFlow introduces explicit schema 2.0 because topology, node identities, action transitions and flow attribution targets materially extend the evidence model. Existing v0.1/v0.2 evidence remains readable and verifiable.

Identity and reproducibility

Both abstractions reuse the same component identity framework. Stable explicit versions/digests and artifact hashes are preferred. Runtime memory addresses may only distinguish unstable objects inside the current process; they are never serialized as reproducible evidence.

Local execution boundary

The package remains CPU-first and file based. There is no required GPU, cloud service, database, Docker runtime, model registry, feature store, telemetry path, LLM/API, scheduler or hosted dashboard.

Trust boundaries

DeciShift can verify its own declared artifact hashes and report supplied/safely derived software provenance. Hash verification does not authenticate a person or organization and does not prove that an external model/data artifact is trustworthy.

Structural impact is graph reachability, not causal impact. Software counterfactual attribution does not establish real-world causality.