LOW evidence grade allows zero-effort findings — real-world failure case

[dshanklin-bv]

The Problem

In a real session tonight (aic-director-of-ai-cockpit session #8), an agent used research.md to file 20+ findings across multiple projects. Every single one was graded LOW with no primary source material read. The agent:

• Asked Gemini a question, filed the summary as a finding
• Wrote one-sentence claims with no source verification
• Used research.md as a clipboard, not a research tool
• Never fetched a URL, never read a primary document, never hashed content
• Filed findings in seconds that should have taken hours

The evidence grade system (UNVERIFIED → LOW → REASONED → CONFIRMED) is supposed to gate quality, but LOW requires nothing. The tool accepts 'Gemini said so' as sufficient. There is no enforcement of actual research.

The Impact

• 20+ findings exist that are just AI-summarized opinions
• None of them have been validated against primary sources
• Decisions were almost made based on these hollow findings
• The user caught it and called it 'hot garbage'

Connection to Existing Work

PROPOSAL-evidence-gathering-gates.md already identifies this problem. This issue provides a concrete failure case that proves the proposal is needed.

What Should Change

1. LOW should require at least one fetched URL — not just a claim that a URL exists
2. The tool should refuse findings with no web research — force the researcher to search before recording
3. UNVERIFIED should be the only grade available without source material — and UNVERIFIED findings should be flagged as needs-investigation not filed
4. Consider removing LOW entirely — force researchers to earn at least REASONED (requires content_hash)

The tool should make shallow research HARD and deep research EASY. Right now it is the opposite.

Session Reference

aic-director-of-ai-cockpit session #8, 2026-03-30/31. ASMP research projects 000-006.