기술 블로그

← 모든 글

· Invest AI

Research Plane vs Decision Plane (ADR-055)

UI는 한국어입니다. 글 본문은 아직 영어 또는 중국어만 있습니다.

The demo that refused to lie

Prep for an Invest walkthrough wanted Map → Micron → 13F holdings. The Micron snapshot had no verifiable 13F citation. The run disclosed the gap and used NVIDIA for holdings instead.

That is the product rule in action: No Citation, No Claim. Evidence for one ticker cannot underwrite another.

The architecture gap

ADR-012 already says the worker computes and the API only reads. It does not name the softer split:

  • Research Plane: what does the evidence support?
  • Decision Plane: what should we do given objectives and constraints?

Wanting a recommendation does not strengthen weak evidence. A cited holding does not by itself mean “this user should buy.”

Evidence Contract (planned)

ADR-055 accepts the pattern and marks it planned. Research should hand Decision a package: claim, source passage, dates, reported/derived/inferred, verification status, unresolved gaps — and preserve missing fields.

What is not claimed: a deployed contract schema, an enforced Research→Decision gate, lineage metrics, or accuracy gains.

Live cousins: public /ai-map citations, ADR-054 hiding unsourced landing cards, Growth Engine facts-only + human gate (ADR-050).

Longer teaching write-up: enterprise design article.

Takeaway

Keep the CQRS line. Add an evidence handoff that can reject. When the source stops, show where — don’t narrate past it.