What is computed, and what is written for the demo

Computed in your browser

• The code checks (word lists, scripts, names against the bible, residue), every route's verdict, catch counts and false alarms against the answer key.
• BM25, spelling and Devanagari folding, the boosts from a model's query reading, nDCG@10, recall, empty pages and the search gate.
• The generator checks (language, rating, cast, residue, beats, length), the release gate and cost per accepted draft.
• Every cost: tokens times list prices (customer-bot/config/models.yaml (Prices: Dvar from portal.dvarlabs.ai/models (INR per 1M tokens, checked 2026-10-21).); frontier from OpenRouter list prices), at ₹84 per US$ (customer-bot/config/models.yaml, fx_inr_per_usd (labelled constant)).
• The localisation checks (script, names, Hindi or English left in, length against the reference) and cost per accepted dub.
• The bench: share right and cost per correct outcome for every model on every task, by language and genre.
• The gateway's routing decisions and the simulated day.

Written for the demo

• Ten invented shows, their bibles and 180 scenes (scripts/generate.mjs, seed 42), with the answer key and which problems only a model can see.
• A 46-title catalog and 74 queries; relevance grades follow from catalog attributes and what each query was written to ask.
• 12 episode briefs and three generator versions' drafts, authored with the problems each version shows. They are not model output.
• Six Hindi scenes with hand-written reference dubs in Tamil, Telugu, Bengali and Marathi (24 dubs): the answer key for length.
• Monthly volumes, token shapes for the gateway workflows, the 400 ms search budget and the 0.7 escalation threshold.
• Model replies: run run-202610230802, measured 2026-10-23.

Measured live, on this server

The Live studio and every Run live button send the same input to an open model on Dvar and a frontier model, now, with the frozen run's instruction and temperature 0:

• check: Dvar · GLM 5.3 and Claude Sonnet 5.5
• search: Dvar · GLM 5.3 and Claude Sonnet 5.5
• draft: Dvar · GLM 5.3 and Claude Sonnet 5.5, judged by Dvar · GLM 5.3
• localise: Dvar · GLM 5.3 and Claude Sonnet 5.5, judged by Claude Opus 5.5

Cost is provider-reported tokens times list price (OpenRouter's reported cost when it returns one). Each visitor has a daily cap; message text is never stored or logged.

Making it a measurement

npm run freeze with DVAR_API_KEY and OPENROUTER_API_KEY set sends the same instruction (version 2026-10-23.2, temperature 0) to every model and stores raw replies, tokens, latency and cost. The pages then parse and score the replies with the code above; no number is typed by hand. Tune instructions on the dev split only.

With Kuku: replace the invented scenes with a few hundred of yours, labelled by your standards team, and the queries with a sample of real searches. Nothing else changes.