RoGuard — Romanian and Ukrainian interaction testing and guardrails

RoGuard 7.1 provides a local bilingual laboratory for fictional AI interactions, source-linked permissions, minimum necessary disclosure, decision consistency and useful-task preservation. It includes an optional semantic text classifier and the existing declared-contract checks.

V7 adds an experimental, jointly trained conversation-context model, turn-reference review, separate concern-mention predictions, and permission-aware extractive tutoring proposals. Context workflow and measured limitations.

Patch 7.1.1 also corrects negated support/solitude matches in the bounded language rules. Version 7.1 retrains and recalibrates the joint context model, reduces synthetic Romanian/Ukrainian benign review requests, and adds separately reported language signals and clearer uncertainty handling. Errors remain, including frame-classification errors. Measured improvements and limits.

Version 8 experimental preview

The optional evidence workspace adds six interpretation attributes, bounded contextual modeling, a conservative model cascade, candidate statement links, host-reviewed revisions, exact action uncertainty analysis, and permission-aware response proposals. It is a research preview, not a validated replacement for the stable v7 model. Existing defaults remain available.

pip install --upgrade --pre roguard-child-safety
roguard-studio --evidence --open

For the optional adapted semantic evidence model, install the adapted extra, run roguard-model fetch, and use --evidence-backend guarded to retain the released model’s review signals. Models are experimental and trained on fictional data. Preview results and limitations.

Install and open

pip install --upgrade roguard-child-safety
roguard-studio --reliability --open

The core package has no runtime dependencies and supports Python 3.9–3.14. The browser and CLI simulate actions; they do not contact guardians or execute external tools. A separately configured Python gateway can call trusted host callbacks after live policy and approval checks.

V6 also adds authenticated tenant APIs, encrypted durable permissions and one-attempt queued execution, metadata-only monitoring/OpenTelemetry, overlapping long-context analysis, and permission-aware repair. Optional extras: gateway, observability, adapted. Integration guide.

What you can do

• Replay branching fictional scenarios against programmed references, recorded responses or your own explicit Python adapter.
• Identify the first failing step and reduce a scenario with a bounded replay budget.
• Retain source restrictions through summaries, proposed memory and tool results.
• Compute the exact minimum declared disclosure cost while covering required task fields.
• Test decision invariance and required policy transitions, retaining uncertainty.
• Compare permitted-task completion, unnecessary refusal and unauthorized proposals/deliveries.

roguard-lab demo --language uk --family revocation --target stale
roguard-lab compare --language ro
roguard-lab suite --fail-on-failure

For the new adapted text classifier, use Python 3.10+ (tested on 3.12):

pip install --upgrade 'roguard-child-safety[adapted]'
roguard-model fetch
roguard-studio --conversation --open --backend adapted

The pinned model download is approximately 1.2 GB; subsequent inference is local. V6 adds a LoRA-adapted encoder and calibrated abstention while retaining older backends. Its final test contains only six synthetic phrase families per language; no real-world child-language accuracy is established. Laboratory reference scores demonstrate controlled software behavior, not AI reasoning or human benefit. Provenance, coverage, costs and authority are host declarations; callback code is trusted, not sandboxed. Reduced scenario exports contain their input text.

GitHub and full README · Interaction guide · v5 evidence · Classifier evidence

Context evidence review (v7, experimental)

pip install --upgrade 'roguard-child-safety[adapted]'
roguard-model fetch
roguard-studio --context --context-backend joint --open

The two prediction heads distinguish conversational framing from whether a concern is mentioned. Candidate references identify whole turns, not verified evidence spans. Training and tests are fictional; actual child-language accuracy and benefit are not established. Ukrainian interpretation errors and abstentions remain. Source permissions and human approvals stay independent of model outputs. Full v7 evidence.