How to test
The tour as a checklist
Each stop says what to do and what you should see. The guided tour walks the same list with a bar at the bottom of each page.
DoRead what Kuku has said in public and where each Dvar piece fits.
CheckFive rows, each with a linked public source and a page in this demo, then the questions for the meeting.
2. The bench: every model, every task
DoSwitch between the four tasks, then flip the breakdown between Language and Genre.
CheckOne ranking per task by share right and cost per correct outcome. The breakdown shows where a model that looks fine on average is weak.
DoRead the route table, then change scenes a month.
CheckFour ways to run the same five checks. The tiered route runs free code checks first, a small open model on every scene and a frontier model only when unsure. Its monthly bill is a fraction of frontier-on-everything.
4. A Marathi line in a Hindi show
DoRead the scene, then press Run the checks.
CheckSame script as Hindi, so the code check passes it. Only a model sees the language changed. The answer key is shown beside each route's verdict.
5. Search that understands listeners
DoSwitch between the search systems and read the ranked results.
CheckKeyword search matches words; the model step reads the ask as horror, village, ghost. The table below scores every system on the same queries, by query type, with cost and p95 latency.
6. A generator update, blocked
DoRead the red gate card, then open a failing draft.
Checkv4 hits more beats and hooks than live, but drifts into English in Hindi episodes and swears in a 13+ show. Critical failures block the release.
DoSwitch to v4.1 and read the check table.
CheckNo critical failures and more checks passed than live, so the gate says Ship. One long draft is a miss a writer should see, not a blocker.
DoRead accepted dubs per language, then open a dub that needs work.
CheckScript, names, Hindi or English left in and timing are code; meaning and tone are a judge. Each language gets its own number.
9. One gateway, open and frontier
DoFlip a workflow between frontier and routed, then turn on the outage.
CheckEach workflow has a primary, an escalation and a fallback. During the frontier outage, drafts move to the open fallback and search degrades to keyword results instead of failing.
10. Try it live
DoPick a mode, edit the input if you like, and press Run on both models.
CheckBoth models answer now, with what each call cost and how long it took, scored by the same code as the bench. Without keys on the server, each side says which key is missing.