On-premises, disconnected, and open-weights operations

Prerequisites: distinguish generated output from source evidence; review the AI workflow reference. Public self-study is not official agency training, accreditation or a credential.

What you will be able to do

This lesson develops the ability to design bounded local operations with lifecycle controls. The setting is ORION Service Navigator, a fictional public-benefits information service using public guidance and wholly synthetic constituent questions; it never determines eligibility or takes an action on a person. Every person, organization, system record, metric, and outcome in the case is fictional. Do not replace the case with operational, classified, CUI, export-controlled, proprietary, employer-confidential, or third-party personal information.

By the end of the lesson, you will be able to:

• Demonstrate how to design bounded local operations with lifecycle controls.
• Apply that model to the synthetic ORION Service Navigator environment.
• Produce a reviewable artifact that separates evidence, assumption, uncertainty, and authority.
• Recognize the critical failure and select a safe remediation path.

The performance standard is not confident language. It is a traceable decision, an artifact another qualified reviewer can inspect, a clear boundary around the learner’s authority, and a safe response to uncertainty. Knowledge checks are formative and do not award a certification or endorsement.

This curriculum runs self-study on VetsLearn. Where the text calls for an instructor, peer or formal review, use the stated acceptance criteria for self-review. Reflection questions support reflection; they do not replace independent assessment. A peer review is optional unless your actual organization requires it.

Part 1 Disconnected configuration operations

Model, product, harness and compute are different axes. Open weights do not prove open-source licensing or trustworthiness. APIs can be local; local inference can still use hosted embeddings, tools, telemetry and backups. Containment covers the whole configuration. A disconnected runbook includes approved intake, provenance/hashes, dependencies, administrators, logs, timekeeping, backups, patches, restoration and export. Never install software on work hardware without approval. Capacity includes weights, cache, runtime and concurrency, so measure the workload.

Component and destination record

Component	Version in synthetic case	Destination	Open question
UI	U1	Appliance	Who can administer it?
Serving API	R1, localhost	Same appliance	Bind address and role enforcement?
Weights	W1, quantization Q4	Appliance storage	Provenance/license verified?
Retrieval	Index I1	Local files	Per-role filtering tested?
Embeddings	E1 default	Remote endpoint	Replace or obtain boundary review
Trace collector	L1 default	Hosted dashboard	Disable or establish approved path

Intake and restoration procedure

Create an import manifest recording filename, publisher/source, exact version, license, expected cryptographic hash, dependency/runtime versions, authorized intake method and reviewer. In this simulated case, manifest H1 expects weights hash HASH-W1 and index hash HASH-I1; these are placeholders, not real cryptographic hashes. A mismatch blocks intake. A matching hash checks identity against the manifest; it does not establish publisher trust or authorization.

For the paper restoration exercise: (1) freeze the affected simulated service; (2) identify backup B1 containing W1, runtime R1, corpus D1, index I1 and policy P1; (3) compare manifest versions and hashes; (4) rebuild I1 from D1 if the index is corrupt; (5) run known query Q1 whose expected source is D1:notice-2; (6) run the role-denial fixture; (7) reopen only when outputs and configuration match the recovery criteria. If the backup contains unknown policy P0 or Q1 cites a different source, stop recovery and route the inconsistency. Never substitute an unreviewed model download to make restoration convenient.

Limits: the paper exercise demonstrates reasoning about recovery. Actual containment, throughput and restoration require authorized hardware tests and observation.

Part 2 Synthetic worked example

ORION runs local weights through localhost serving, but retrieval defaults to hosted embeddings and errors go to a cloud collector. Local inference is true; no-data-leaves is unsupported. The synthetic runbook specifies internal replacements and an egress test; real containment requires authorized configuration and observation.

Guided practice: identify the unsupported conclusion, name missing evidence and rewrite with its scope intact. Change one assumption and identify which decision reopens.

Part 3 Feedback and repair

Pass when localhost is distinguished from hosted access, destinations have owners and restoration is demonstrated. Disconnecting inference alone does not settle media, administrator or backup risks.

Keep failed evidence beside the repair. Explain the criterion, correct the artifact and test a different invented case.

Next step: carry your versioned artifact and failures into the next federal lesson; the capstone returns to these records. Check live policy and authorization with current official sources and responsible owners.