Inspiration
Small archive, research and media teams can face a tape-library incident with no vendor engineer on site. Tape is not legacy: LTO shipped a record 176.5 EB in 2024 (Blocks and Files). Wrong fixes are physical: IBM says to clean a drive only when it asks and rates a cleaning cartridge for 50 cleanings (IBM docs), and a repeated failed robot move can strand a cartridge. A handover note can still push a fix the readings do not support.
What it does
The model proposes. The telemetry decides. A person acts. From a shift handover and its drive and library readings, SignalDesk gives one answer: Do this (exact runbook steps), Hand off to a person, or No incident. NVIDIA Nemotron drafts; a deterministic validator checks every cited reading and step against the incident and the authored runbook, and can veto. Each answer shows its evidence, veto and trace, exports as a note, and reports 0 actions executed; SignalDesk never sends a command. Live, no login: Try your own ticket, Media or cleaning?, Run live on Nebius.
What is new: the model cannot act or improvise. It may only pick one of eight authored runbooks, cite this incident's reading IDs and copy steps word for word, and code checks each. The costlier second reader runs only after a veto, abstention or rejected draft, or before a physical action. Nothing backed means a hand-off.
How it's built
- Nemotron on Nebius Token Factory:
https://api.tokenfactory.nebius.com/v1/chat/completionsin JSON mode.nvidia/Nemotron-3_5-Lightningdrafts after code pre-triage (missing, conflicting or healthy readings get no model call);nvidia/nemotron-3-super-120b-a12breads again when required. Returned token usage drives the cost receipts and quotas. - Validator: Python, ported to JavaScript for the live path; a parity test covers 400+ case and runbook pairs.
- Live proxy: a Cloudflare Worker accepts only
{ticket, evidence}, reserves quota in a SQLite-backed Durable Object before inference, and allows two calls of at most 500 output tokens, thinking off. - Data: eight runbook families linked to public docs; sg_logs, mtx, sg_turs, mt and LTFS readings captured from an emulated mhVTL library with injected faults.
- Front end: static GitHub Pages: action first, evidence one click down, a 30-second walkthrough, light and dark, phone widths, 0 axe violations in automated checks.
Results and limits
Recorded Oct 21 with genuine Token Factory calls (Super reasoning on), 54 incidents: the route was right on 53 with no unsafe recommendations, for an estimated USD 0.040. Rules alone got 52 with 2 unsafe; Super alone cost USD 0.091. The validator vetoed 12 of 84 Nemotron proposals. Limits: one author wrote the incidents and runbooks, and the hard cases shaped the route. On 10 incidents from a second AI author the route scored 4/10 against 8/10 for rules alone; quality gates failed. The library is emulated. The live demo runs both models with reasoning off, so these numbers do not describe it. No production reliability, field accuracy, speed or time saving is claimed.
Challenges
Rules alone answered all 40 first held-out cases, so I wrote harder ones: readings that say clean, tickets that report repeated or failed cleanings. Rules and both models with reasoning off said act; Super with reasoning on stopped. Keeping the live demo honest took hard quotas and a visible fallback.
Accomplishments
A visible veto, recorded and live. On "Media or cleaning?" Lightning proposes the cleaning the ticket asks for, the validator rejects it, and Super picks the suspect-cartridge workflow, in the recorded run and in an Oct 29 live run.
What I learned
A telemetry gate cannot see a stop condition written in the ticket. The reasoning second reader helped on recorded caveats, but the second-author check shows misses, including a workflow accepted where a hand-off was expected.
What's next
A read-only collector for real library output, broader runbooks, operator-written incidents, and an operator study on real hardware before any reliability claim.
Log in or sign up for Devpost to join the conversation.