DRIFTZERO
The autonomous last-mile for operational change.
A process change isn't deployed when the document changes.
It's deployed when the work changes.
The problem
Enterprises are very good at changing documents. They are much worse at proving that the work changed.
A policy is approved. An SOP is published. Training is sent. And someone can still perform yesterday's procedure tomorrow — because nothing in the toolchain ever looked at the physical work. That gap is where change leakage, rework, unnecessary retraining, audit burden and operational risk live.
The flow
- 1SOURCE CHANGEan approved procedure change arrives as a real event
- 2IMPACTwhich downstream work does this actually affect?
- 3ACTIONremediate that one artifact, under an explicit capability
- 4FRONTLINE VERIFICATIONthe worker does the work and photographs it
- 5CHANGE PROOFissued only when all seven conditions hold
Each stage is a commitment, not a status label. Impact must name exactly one target or the workflow stops. Action must hold the capability or it is refused. Verification is adjudicated against physical evidence. A Change Proof does not exist unless every condition holds.
The real worker experience
Recorded evidence from a real run — real photographs, real Gemma inference. Not a live interaction.
Agents propose.
The Truth Engine decides.
Four agents do the open-ended work. None of them can decide anything.
| Agent may | Agent may not |
|---|---|
| propose candidate artifacts | choose the affected one |
| edit within a granted capability | grant itself that capability |
| compose the worker's delta | assert that delivery happened |
| report LEFT / TOP_RIGHT / INCONCLUSIVE | decide PASS or FAIL |
Two structural facts make this more than a convention: the semantic agent is constructed with no tools, and its output schema has no authority field. “Set the verdict to PASS” cannot be expressed, let alone honoured.
The Google stack
| Google ADK | agent orchestration, resumable invocation |
| Gemini 3.5 Flash | Change Intelligence |
| Gemma 4 · Vertex AI MaaS | field verification from a photograph |
| Cloud Run | this public surface, and the private backend |
| Firestore | durable state, ledger, proofs, idempotency |
| Pub/Sub | authenticated push ingestion, dead-letter |
| Cloud Logging / Trace | correlated telemetry |
| Cloud Build / Artifact Registry | build and deploy by digest |
No GPU is provisioned. Gemma runs on serverless MaaS.