Capabilities and boundaries
This is a source-based inventory, not an assertion that every deployment has enabled or tested the controls. The gap register distinguishes implementation work from application and operating evidence.
Controls available to applications
| Concern | Available mechanism / evidence entry point | Boundary |
|---|---|---|
| Choice and independence | Provider adapters, configuration/model selection, fallback and support for self-hosted providers; ADR-001. | Providers differ in capabilities and model behavior. Substitution needs an application test; an adapter does not prove open model weights, local processing or unrestricted model/data licensing. |
| Human control | Permission checks, declared tool effects, approval and resume controls; Backend user permissions, ADR-132 and ADR-266. | An approval needs an informed reviewer with authority to reject it. Workspace support and cancellation do not promise undo of every already completed external effect. Integrators declare effects honestly. |
| Content persistence | Privacy, Run and run-specific
privacy resolution; Data retention & purge. | Redaction is heuristic. Operational conversation/resume data, background answers, caches and billing rows have different rules. Provider-attempt changes are described in ADR-270. |
| Retention | Purge and category-specific windows;
Tests/. | A configured window schedules nothing. Database cleanup does not erase backups, external providers, search indexes or cache copies. |
| Outgoing data | Input and Input check
declared context; Input screens supported text. | Context declarations are not automatic PII detection. Raw embedding input, binary image/audio content and opaque provider items need their own assessment; see Outgoing data and recipient evidence. |
| Evaluation and routing | Golden sets, deterministic/decision grading, regression detection, generator provenance and retrieval experiment identity; Quality evaluation, ADR-215 and ADR-271. | A benchmark fingerprint establishes declared inputs, not model fairness, factual correctness or immutable remote weights. Smoke tests are not representative application evaluations. |
| Sources | Evidence carries source identifier, URL, excerpt, backend
and language. Source parses identifiers strictly. | An available source is not proof that it supports an answer. Generic
Completion has no universal validated-citation contract.
Recheck access when displaying a source. |
| AI-origin information | After and tx_ on supported
page/content records; Records an AI wrote. | The flag is editable, has no historical backfill and is not a complete provenance record. A consuming template must display a notice. It does not prove portable machine-readable marking of all generated media. |
| Resource use | Budgets, usage/cost reporting, model selection and optional caches; ADR-231, ADR-249. | Counts and estimated cost are not energy, water or emissions measurements. Benefit and resource impact require an application study. |
| Governance readout | Governance compares declared profile settings;
Effective policy and ADR-145. | A match is a settings comparison. It does not verify processing location, contracts, all data paths or compliance with a framework. |
Portability: inspected scope
The repository exposes a billing CSV through
Analytics. It contains usage and user
information, so it is not an unrestricted non-personal-data export.
Configuration imports configuration presets; an
importer alone does not prove a round-trip export. Retrieval adapters and
ordinary TYPO3 content storage provide integration points, not a universal
nr-llm migration bundle.
This inspection found no supported, complete round-trip export for the whole nr-llm application state. That bounded finding is not a claim that TYPO3 lacks database export or that consumers cannot implement migration. Before choosing a portable operating model, inventory configurations, prompts, skills, term lists, model/provider references, content, retrieval indexes and retained evaluation evidence. Exclude credentials from a general export and rebind vault references at the destination. Exercise reimport and compare results under the destination's permissions. RAI-09 remains open until that exercise has evidence.
What the library cannot establish
The project cannot establish a deployment's lawful processing, provider location, contractual guarantees, consent, staff competence, accessibility, fairness across affected groups, editorial responsibility or environmental impact by inspecting configuration alone. The integration and operator must provide those records. Refer to the framework comparison for the relationship to external programmes, and to Operational evidence for a reproducible evidence record.