Diffing makes change legible without enforcing it.
A local-first, read-only standard for observing drift and decision lineage across AI systems without surveillance, scoring, or control.
Local-first
No outbound calls by default.
Read-only
Observes change without enforcement.
Deterministic
Explainable drift, no scoring.
The problem
Systems change quietly. Output formats shift, decisions drift, and KPI meaning evolves while pipelines keep passing. Most tools show activity, not divergence.
What Diffing makes observable
Evaluation
Record one execution with input and output fingerprints.
Expectation
Derive baseline intent from observed history.
Variation
Detect deterministic divergence from baseline.
Lineage
Reconstruct a neutral timeline on demand.
Proofs and surfaces
DiffGen 1.1 — AI Output Drift
The canonical proof: evaluation capture, deterministic variation, and lineage reconstruction.
Enterprise artifact
AI Change Provenance Record
A neutral disclosure format for audit and governance.
View the artifactInstall in 15 minutes
pnpm install pnpm dev -- init pnpm dev -- hello pnpm dev -- lineage <thread_id>
The Hello World demo produces an evaluation, a deterministic variation, and a lineage narrative. No integrations required.
See the install guideTrust and governance
Diffing is neutral by constraint. It observes and explains without enforcement.
Ready to see drift?
Run the local proof or review the standard.