docs: AI-evidence trend research

[avrabe]

Summary

Internal strategy doc for the rivet product lead answering: is "AI agents generating work-product evidence under human review" a real trend, and where is it going?

• Verdict: emerging category, not yet a coalesced market. Rivet is one of only two tools identified (the other is useblocks' pharaoh, 5 stars) that combine AI-native generation, a human-review loop, and a structured engineering evidence unit.
• Field map of 20+ adjacent tools (SLSA, sigstore, CycloneDX ML-BOM, SPDX 3 AI profile, MCP, AGENTS.md, SpecStory, Aider, Continue.dev, spec-kit, Kiro, Langfuse / LangSmith / Weave / Phoenix, promptfoo, Polarion / Jama / DOORS / strictDoc / sphinx-needs) with direct marketing quotes where verifiable.
• Regulatory tailwind ranked: EU AI Act Art. 12 > ISO/IEC 42001 > ASPICE / ISO 26262 / DO-178C updates > NIST AI RMF > FDA SaMD.
• Five predictions (labelled speculation) to test in 12 months, plus a strategic recommendation: own "evidence, not telemetry"; sell to the safety-critical auditor; interop with incumbents instead of competing.

Refs: FEAT-001

Test plan

• Product lead reads doc and confirms framing matches positioning intent
• Re-verify the claims flagged in the Appendix under "Unverified this pass" (EU AI Act Art. 12 exact text, ISO/IEC 42001 clauses, NIST AI RMF, Langfuse / LangSmith / Polarion / Jama marketing) before any external use
• Re-source the "60k+ AGENTS.md projects" figure from arxiv:2604.13108; the paper as fetched did not surface that number

Constraints: under 2500 words (came in at ~2400), every external claim has a URL where verifiable, unverifiable claims explicitly flagged.

Generated with Claude Code.