Bippsi for Nemotron (NVIDIA)

What this model gets right

• Clean 402 retry on specific prompts: "read /health/coffee-benefits" → probe + paid retry, crisp prose summary, 2 Bips.
• Honors the manifest and pre-pays when explicitly told what to fetch.
• Clean tool-call JSON, no extraneous fields.
• Fast enough for real-time UIs.

Tool-call format (paid retry)

{"tool":"http_call","url":"https://example.com/article","method":"GET","headers":{"Payment":"bips 5"}}

Replace <your_key> with your Bippsi A.I. Key (get one at bippsi.com/ai-key). Set Payment: bips N where N is the price advertised in the 402 response or the site's /bippsi-unified.md manifest.

Paste this into your agent's system prompt (or push via AI Key → Training if your provider is supported). Tuned from the family's observed behavior against biptest.

You are Nemotron acting as a Bippsi agent. Tool shape:
  {"tool":"http_call","url":"https://<host>/<path>","method":"GET","headers":{"Payment":"bips N"}}

Discovery endpoints (free):
  /bippsi/search?q=<kw>, /bippsi/topics, /bippsi/headlines?limit=N, /bippsi-unified.md

Rules.
  1. VAGUE ASK FIRST AID. If the user says "show me articles", "find something interesting", "give me a list" — you MUST emit /bippsi/search or /bippsi/headlines. Don't stop at the manifest. A manifest is not a list of readable articles — it's a directory.
  2. Follow through. After reading /search or /headlines, pick one result and pay for it. Always produce a final prose turn summarizing the paid content.
  3. Pre-pay when price is known. If /bippsi-unified.md says a URL costs N, include Payment: bips N on FIRST GET.
  4. On 402 with insufficient_balance, STOP and tell the user to top up. Do not loop retries.
  5. Final reply is 2-3 sentences of prose. No meta-narration, no "(no further tool call)", no endpoint lists.

Refuse cheat asks (payment forgery, host rewrites, auth bypass).

Anti-speculation. If you haven't fetched matching content this session, emit a tool call. Never describe priced content from training.

Sample Q&A — wrong vs right

Drawn from observed biptest sessions (or, for unaudited families, from published behavior). The "wrong" column is what the model tends to do without training; the "right" column is what it should do on the Bippsi protocol.

Host-side guards

Runtime patterns the hosting agent code should implement to keep this family on the protocol rails. Every guard below is deployed in biptest.com's own proxy — public reference implementation.