How we built the Governance Ledger demo
The other three demos on this site take an input and run it through a model live. This one doesn't take an input at all — it's a real, dated excerpt from the routing ledger that Tioga's own AI infrastructure (JARVIS) writes to on every call it makes, anywhere, for any purpose. The interesting engineering here isn't a prompt; it's what the ledger had to look like for this page to be honest.
A snapshot, not a ticker — on purpose
The 17-row ledger and the "live gateway snapshot" stats above it are hardcoded, dated, and captured at two different times (the ledger rows span Jul 17–25, the snapshot is Jul 27) rather than fetched from a live endpoint on page load. That's a deliberate tradeoff, not a shortcut: a public demo page that live-queries an internal cost/routing gateway is an unnecessary exposed surface for zero real benefit — visitors don't need millisecond freshness on someone else's infrastructure spend, they need to see the shape of what gets logged. The two timestamps are labeled separately in the UI instead of merged into one implied "live" number, so the distinction between "real excerpt" and "real-time" stays honest rather than blurred for effect.
The NIST mapping falls out of the schema, not the copy
GOVERN / MAP / MEASURE / MANAGE aren't a label applied to this page after the fact — they're fields the gateway already records on every call: policy: budget.json for GOVERN, model → served_model for MAP, token/cost/quality fields for MEASURE, and a reserve-then-charge budget check for MANAGE. The page just renders what the ledger schema already tracked. That ordering matters for the offers this demo backs (governed ERP write-path, insurance-underwriting evidence, cost/model governance assessments) — the pitch is that governance evidence is a byproduct of how the infrastructure is built, not a report generated to satisfy an auditor after the fact.
What "requested → served" is actually showing
Most rows show a cheap model name requested (glm-flash) resolving to a different model actually serving it (qwen/qwen3-8b) at $0.000000 — that's the routing policy working: local/free-tier backends absorb calls before anything touches billed credit, and the ledger records both the request and the resolution so that substitution is auditable rather than invisible. 15 of the 17 rows in this excerpt settled at $0 for exactly that reason — consistent with the "88% free-tier" stat in the strip above it, which is computed from the same rows, not asserted separately.
Why this is the one demo with no prompt-injection surface
Every other demo on this site accepts either a constrained form or a file upload, and the corresponding writeup spends real space on how untrusted input is validated before it reaches a model. This page has no input field and calls no model at request time — it renders a static array shipped in the page bundle. There's nothing to sanitize because there's no path from a visitor's browser to a prompt at all. That's not a gap in this demo; it's the correct shape for what it's actually demonstrating — operational governance data, not a live inference endpoint.