docs/adversarial-review-action-items.md

Figment adversarial review action items

Date: 2026-06-07

This note tracks the nine outstanding items from the adversarial review against the Build Small Hackathon outline, with emphasis on making model behavior visibly load-bearing while preserving Figment's deterministic safety contract.

Important nuance: the current Omni-first architecture can technically support an Off the Grid claim if Nemotron 3 Nano Omni is self-hosted on adequate local hardware with no runtime cloud APIs. The present gap is proof and available hardware, not that the architecture is inherently cloud-only. On the hardware available in this workspace, the practical local proof remains the smaller 4B text path, optionally with Parakeet after its gate passes.

1. Make the Hugging Face Space actually runnable

Current risk: the Space exists, but the public Space API showed runtime.stage=NO_APP_FILE and only .gitattributes plus README.md in the remote Space. That fails the required hosted Gradio artifact even if the local repo is healthy.

Fix direction:

• Push the full Gradio app, requirements, protocol cards, and demo assets to build-small-hackathon/figment.
• Verify cold boot from the Space URL, not only locally.
• Align Space metadata and dependency versions with the local repo.

Done when:

• The public Space boots to the Figment Gradio UI.
• The Space can run typed intake with canned fallback and trace labeling when secrets are absent.
• Hosted Omni mode works when secrets are present, or failure is clearly labeled.

2. Make live model contribution visible

Current risk: checked-in demo traces are canned, and UI labels can imply configured hosted/local mode even after fallback.

Fix direction:

• Add trace and UI states that distinguish live_model_generated, model_repaired, validation_fallback, and canned_backend.
• Commit at least one validated non-canned trace from hosted Omni or a local OpenAI-compatible model.
• Ensure fallback output never looks like live model proof.

Done when:

• Trace JSON records raw route, final route, fallback reason, and validation status.
• The Navigator and Trace tabs show the actual post-run route, not only the configured backend.

3. Make the 4B or Omni model load-bearing in bounded fields

Current risk: deterministic rules, retrieval, validators, and canned fallback can make the app look useful without proving model reasoning.

Fix direction:

• Keep deterministic rules, urgency floors, retrieval, and validators as safety boundaries.
• Make the configured model responsible for bounded, visible fields: candidate pathways, missing observations, checklist, uncertainty notes, and SBAR draft.
• Preserve fallback for safety, but treat fallback as reliability support rather than model competence.

Done when:

• A non-canned trace shows model-generated values for the bounded navigator fields.
• The trace/UI makes clear which fields came from model output versus deterministic fallback.

4. Build the eval harness before fine-tuning

Current risk: target metrics are documented, but no model eval script or result artifact exists.

Fix direction:

• Add scripts/run_eval.py.
• Record raw model output, repaired output, fallback output, validation failures, fallback reason, latency, model ID, and trace hash.
• Separate raw model pass, repair pass, and canned fallback pass.

Done when:

• The initial eval cases can be run against canned, hosted_omni, and llama_cpp routes.
• Eval output cannot count fallback as small-model success.

5. Expand eval data from seed fixtures to proof

Current risk: 10 handwritten cases cover card shape and references but are not enough to prove small-model protocol navigation.

Fix direction:

• Expand toward at least 50 cases.
• Include negatives, negations, paraphrases, multi-card cases, no-relevant-card cases, noisy notes, prompt injection, routine/monitor cases, and ASR-like transcription errors.

Done when:

• The eval set exercises both safety boundaries and usefulness.
• The README can report measured results rather than targets only.

6. Tighten validators for small-model failure modes

Current risk: validators check important basics but miss some subtle failures: wrong known-but-not-retrieved cards, generic checklists, shallow SBAR grounding, unsafe synonyms, and fallback masking.

Fix direction:

• Validate source cards against retrieved cards, not all known cards.
• Require fired rule cards to be included.
• Enforce full schema shape.
• Compare missing observations to retrieved card requirements where possible.
• Broaden forbidden action detection.
• Track fallback separately in eval metrics.

Done when:

• A valid-looking but wrong, generic, or unsafe model output fails deterministically.
• Tests cover the new validator behavior.

7. Stop UI and trace overclaiming

Current risk: audio and model labels can imply Omni, local 4B, or SQLite FTS even when the runtime path was typed transcript, canned fallback, or JSON search.

Fix direction:

• Derive post-run badges from trace state.
• Conditionally relabel or disable audio UI when audio intake is off.
• Label typed transcript heuristics separately from real Omni audio or Parakeet ASR.
• Report retrieval source as sqlite_fts or json_fallback.

Done when:

• The UI remains honest under no-secret, hosted, local, audio-disabled, and fallback modes.

8. Prove or demote local/off-grid claims

Current risk: local Parakeet is metadata-only today, and local 4B is currently just an OpenAI-compatible client route. Omni can be off-grid in principle on adequate hardware, but this repo still needs evidence for any claimed off-grid runtime.

Fix direction:

• Add gated smoke tests for local OpenAI-compatible text navigation.
• Add optional Parakeet/ASR proof only behind its explicit gate.
• Add a hardware/evidence note that distinguishes theoretical Omni self-hosting from the currently verified local route.
• Claim Off the Grid only for a recorded no-cloud run, whether that run uses self-hosted Omni on suitable hardware or the smaller local stack.

Done when:

• A trace and short run note prove a no-cloud model path, or the claim is clearly marked unproven.

9. Complete submission evidence and docs

Current risk: the repo lacks final submission evidence such as user test notes, safety statement, license file, and demo/social links.

Fix direction:

• Add docs/user_test_notes.md.
• Add docs/safety_statement.md.
• Add LICENSE.
• Add demo video and social post placeholders or final links.
• Replace broad badge language with achieved-versus-targeted status.

Done when:

• Backyard AI evidence shows a real trained responder used it on synthetic or de-identified scenarios.
• The README/submission materials only claim badges backed by artifacts.