Wall Guard — the Evidence/Judgment Wall

The rules (verbatim from CLAUDE.md — load-bearing, never compromise)

• The evidence layer (cited QA, change detection, FIND, COMPARE) is structurally separate from any future judgment layer.
• No verdict fields anywhere in evidence-layer schemas. No buy/sell/hold, no price targets, no attractiveness ranking, no evaluative language in Athena's own voice (attributed source language, cited, is fine — ADR-0007 §3).
• FIND results are ordered by match_strength — a retrieval fact about how well filing TEXT matched the query — never by company judgment. apps/api/research/find.py must never import an answerer module; the zero-answer-model path is its contract (ADR-0011 §1).
• COMPARE synthesizes per-column behind a (filing_id, query) seam with no passage parameter, so cross-company ranking is unrepresentable; no superlatives (most/best/least/worst), no computed ordering, caps enforced as REFUSALS, never silent truncation (ADR-0012).
• A future judgment layer requires its own ADR, must be labeled as judgment, builds only on evidence-layer outputs, and uses read-only market data (FMP/Finnhub/Polygon/FRED — never Trading 212).
• Prefer structural enforcement over test enforcement: make violations impossible by construction (draft/resolved citation split in change detection; import boundaries in FIND; the no-passage-parameter seam in COMPARE), not merely caught by tests.

Pre-commit checklist (run for any commit touching research endpoints)

• No verdict, ranking, score, or attractiveness field added to any evidence-layer schema or response model.
• find.py imports no answerer module (grep -n "answerer\|summarizer" apps/api/research/find.py → only the docstring).
• COMPARE's synthesis seam still takes (filing_id, query) — no passage parameter; entry assembly stays model-free; no model output feeds a model call.
• No most/best/least/worst/stronger/weaker or ordering-by-attractiveness language in outputs or prompts; evaluative words appear only as attributed source language with a citation.
• Caps are min()-clamped ceilings and over-cap requests are REFUSED with a clear message — never silently truncated.
• Every emitted claim cites a source_url stamped from the database (draft labels → resolved provenance), never model-copied URLs.

The test (ADR-0007 §3 / ADR-0011 §5)

Whose judgment does the answer carry? Stating what each filing says, cited, is evidence work. The SYSTEM picking "most exposed" or "best positioned" is judgment — out of scope even when every underlying fact is citable. The in-scope answer is each company's cited language, ordered by nothing.

Rationalization	Reality
"Every underlying fact is cited, so the ranking is grounded"	A superlative requires Athena to judge across companies; that's the wall.
"The user asked which is most exposed"	Answer with each company's cited exposure language and say the verdict is theirs (ADR-0007 §2).
"Silently trimming to the cap is friendlier than refusing"	A truncated comparison the user believes complete is silent wrongness. Refuse, name the cap.
"A prompt instruction not to rank is enough"	Prompt discipline is the rejected shape (ADR-0012 #6b). Hold walls structurally.