Safety Shift Navigator

Inspiration

What if a new record contradicts the incident handoff that the next shift was about to trust? Most AI summaries can sound decisive even when the original evidence is absent. The UK Health and Safety Executive highlights accurate handovers and independent checking of original information as important human-factors practices (https://www.hse.gov.uk/humanfactors/topics/shift-handover.htm).

We built Safety Shift Navigator, an evidence-to-decision agent that keeps citations, missing records, source contradictions and human authority visible. The question is not only "what does the model say?" but "what changes when the evidence changes?"

Try it now: Interactive no-login demo · MIT-licensed source · 2:24 video. In the demo, choose Network heartbeat → Prepare handoff → Conflicting heartbeat index → Reconcile → Escalate source conflict. Change the selected record and observe how the earlier human choice is invalidated.

What it does

A human reviewer can challenge the entire chain, not just read a one-shot alert.

1. Validate original source observations in a fictional airflow, access-control or network incident; reject malformed data and pin the original source integrity hash.
2. Retrieve the relevant source runbook, with a distinct citation ID; make every cited observation and runbook check directly inspectable in the UI.
3. Reason using NVIDIA Nemotron through the official Nebius Token Factory when a private, scoped, cost-approved key is present. No browser key and no silent substitution of another model. Without a key, the application clearly identifies synthetic offline replay.
4. Critique the model response independently. A strict fail-closed reviewer checks the exact model, source IDs, missing original record, case-specific human question, completion status, token accounting and unsupported certainty.
5. Prepare the handoff for a person, not an equipment controller. No auto-closure, dispatch, machinery actions or real operational changes.
6. Reconcile new evidence. Three fictional follow-up records per incident show partial corroboration, a direct contradiction or an unrelated source. A deterministic asset/signal comparator derives the difference, preserving the original handoff.
7. Record an explicit human-only source-verification choice. The reviewer chooses a permitted next action and receives a local SHA-256 receipt. Changing evidence invalidates that choice. The checksum is not a verified staff signature or real safety approval.

The key idea: the source-specific human decision is part of the agent workflow, not an afterthought following fluent AI text.

How we built it

This is original clean-room work: a small Python standard-library HTTP application and CLI, vanilla HTML/CSS/JavaScript, original synthetic sources and runbooks, a bounded seven-field model contract, a separate source critic, and a human-held decision boundary.

The provider adapter calls Nebius Token Factory at api.tokenfactory.nebius.com with the exact NVIDIA model nvidia/Nemotron-3_5-Lightning. The code independently verifies completion and source consistency; it does not classify every syntactically valid JSON response as true.

The free judge site is a deterministic, clearly labeled synthetic replay (no live Nebius request, paid inference or browser credential). Genuine provider evidence is separate: one authenticated application-to-Nebius model completion returned HTTP 200, the exact model, finish_reason=stop, and 504 input plus 1,959 output = 2,463 tokens. The independent source critic active at the time accepted that known fictional case. The temporary API credential was revoked. See the redacted historical G1 receipt. This is one real integration sample, not a statistical reliability estimate; the stronger critic v2 has not been prospectively tested on unseen live model responses.

Challenges and what we learned

Fluent but incomplete is a failure. Early live Nemotron attempts exhausted their small token budgets in reasoning without producing a usable final answer. A later structured response was superficially valid but missed the exact record a human needed. We retained those failures, made the source critic more specific, and refused to lower its thresholds.

The evidence can change after the first answer. A new heartbeat source may contradict the first. Our comparator reads the actual fictional asset and signal fields rather than accepting an expected outcome label. The system escalates an unresolved source conflict instead of declaring the equipment safe.

The human decision has to be real interaction, not a badge. The second reviewer action requires an explicit selection; no action is preselected. A changed source revokes the previous local choice, and the interface never treats a checksum as an authenticated identity.

Spend is part of reliability. Without a vendor-enforced absolute overage cap, hosting an open-ended public live model endpoint could risk unexpected charges. We deliberately ship a no-login, zero-charge judge replay plus a separate real provider receipt, rather than hiding a billable API key. That choice limits what the public site proves.

Verification and impact

• 97/97 independent Python tests, including malformed inputs, security boundaries and source-specific counterexamples.
• 36/36 local desktop/mobile human-decision browser interactions; 12/12 additional source-inspection checks.
• 18/18 anonymous interactions on the actually published GitHub Pages demo across desktop and 390px mobile; zero vendor or external API requests.
• Nine authored fictional follow-up examples: three each corroborating, conflicting and unrelated, all requiring a human gate. These are deliberately constructed test fixtures, not unseen Nemotron accuracy data.
• Exactly one genuine authenticated application-origin Nebius model completion, plus three separately observed known-case Playground runs after iteration.

Potential impact is a hypothesis: better visibility of missing evidence and contradictory records during a handoff. We have not measured accident reduction, costs saved or staff performance in a field study. The next proper study should compare original-source identification, unsupported factual claims and task completion against a generic AI summary with qualified operators.

What we liked and what Nebius/NVIDIA could improve

What worked: Nebius Token Factory returned the expected NVIDIA Nemotron identity, finish reason and token counts, allowing a straightforward independently checked adapter. Playground access helped us iterate before spending effort on the full UI.

Specific platform feedback: Early tests with 240- and 1,024-token output caps used the reasoning budget without generating a complete final JSON. Clearer separation of reasoning/output budgets, final-answer completeness and retry guidance would improve agent reliability. More verifiable cash-overage controls for promotional credits would make it safer for hackathon developers to offer public live demos.

We would use the Nebius/NVIDIA combination again for bounded evidence-grounded interpretation, with explicit completion checks and cost limits.

Safety and provenance: The demo contains only invented observations, not customer records or real industrial operations. The model is never described as signing off safety, and no real device, payment or staff identity is controlled.

Built With

• ai-agents
• human-in-the-loop
• javascript
• nebius-token-factory
• nvidia-nemotron
• python
• source-provenance