AI Use Disclosure

How tutorials in this cookbook are authored (post-2026-05-23)

Tutorials are drafted in Claude Code sessions by the author (Battam1111) using:

• Claude (Anthropic’s assistant) as a drafting partner
• Sub-agents for parallel research (web search, GitHub repo verification)
• Live web search for up-to-date references against each ZK ecosystem’s official documentation

Each tutorial is then personally reviewed line-by-line by the author, who compiles every code sample in the relevant toolchain before publishing.

Automated linter

Before any tutorial reaches the live site, it must pass an automated AI-tell linter that scans for visible LLM conversational patterns:

• “If you want, I can…” (chatbot offer-to-continue)
• “Let me know if…” (chatbot follow-up offer)
• “Would you like me to…” (chatbot follow-up offer)
• “I do not have live web access…” (model self-disclosure)
• “I cannot honestly claim…” (model self-disclosure)
• “As an AI…” (model identity disclosure)

The linter is open source, see bounty-agent/src/bounty/content_pipeline/ai_tell_linter.py in the private engine repo. The same patterns are blocked at publish time; they cannot reach this cookbook even if a draft slips past review.

What happened on 2026-05-22

Earlier on the 22nd, an automated Frontier API pipeline drafted and published 17 tutorials in three days. Three of those tutorials contained visible LLM conversational tails (“If you want, I can turn this into a research-note version…”). This was a quality-control failure: the pipeline ran without an automated lint and the human reviewer was bulk-processing rather than line-reviewing.

The infrastructure has been corrected as of 2026-05-23:

1. The Frontier-API auto-publish pipeline is disabled by default (config flag BOUNTY_AI_GENERATION_ENABLED=false).
2. Tutorials are now drafted in Claude Code sessions where a human is in the loop in real time, and sub-agents can verify code against live repos and official docs.
3. The AI-tell linter runs automatically before any tutorial publishes.
4. The publish pipeline requires two environment variables and a manual flag, deliberately preventing background jobs from publishing without human intent.

The three problem tutorials had their conversational tails stripped on 2026-05-23; the underlying technical content remains. If you spot any remaining accuracy issue in any tutorial, please open an issue against this repo, corrections are taken seriously.

What Frontier API still does

Internal-only scoring within bounty-agent:

• Bounty go/no-go decisions (numeric scores stored in SQLite)
• Spam classification

These outputs never become public prose, so hallucination risk is bounded.

• Battam1111