A confident guess
meets a hard stop.
A messy tape handover becomes a sourced incident note. Watch Lightning draft, the validator veto and Super take a second look. Recorded Nebius calls. mhVTL readings.
Start with a convincing ticket, a wrong proposal and a visible veto. Explore the cases, take the tour, or bring your own handover.
Try your own ticket
Start with a messy shift handover. Each example also fills the numeric readings below; edit either before running.
Run sends this ticket and the numeric readings to Nebius through the demo proxy. Public or synthetic text only. No equipment actions are executed.
Response playback. Live call timing includes validation; separate validator timing is unavailable.
Readings and local validator
Validator only · no model call · 0 actions executed
Enter readings to check the same numeric gates locally. Leave a field blank to omit it. Decimal and scientific notation are accepted; malformed or non-finite values remain invalid readings.
Choose a runbook to see one authored reference step.
A backed result means an evidence and wording match, not action approval. This validator cannot read the incident ticket or apply runbook caveats. The original route still needs the reasoner and operator to review context before any physical action.
From ticket to checked note
Animation is playback. Model latency and tokens are observed; separate validator timing was not recorded. The recorded route combines separately saved calls.
Why: the readings and the validator's checks
| Reading | Metric | Value | Gate | Verdict |
|---|
A reading over its gate supports a hypothesis; it does not prove root cause. The recorded validator is Python; the live proxy uses the same numeric and exact-citation checks in JavaScript. Neither asks a model to judge a model.
| Model | Draft | Validator | Latency | Tokens in / out | Cost |
|---|
Authored tape runbook · each step cites public man pages, T10 tables or open-source project docs · MIT
What’s next
Validate with an independent operator cohort and real hardware, then measure how well the notes help a shift handover. Extend context checks before considering equipment integration.
Re-evaluate the bounded live configuration: it caps outputs and disables hidden thinking for both models. The recorded Super reasoning-on score does not transfer to your live tickets. Provider availability and current billing remain unverified.
Evidence and current limits
This incident, selected model: · ·
The hero readings come from mhVTL, an open-source virtual tape library, with faults injected on purpose; it is not real drive hardware or a real site. The other incidents are hand-written by the same author as the runbooks. No operator study or time saving is measured. A cache record is local evidence, not cryptographic attestation by the provider. The validator checks that a claim is backed by the readings; it cannot see context the readings do not encode, which is why a stronger reader is required before any physical action.
Second-author synthetic check: Lightning, Super and the route each scored 4/10; rules-only scored 8/10. Quality gates failed.