ORA-2026-0102

Pattern

The financial-reconciliation-matcher computes confidence scores via purely additive arithmetic: base 0.35 + invoice mention +0.35 + cost_code match +0.20 + supporting lines +0.05 + single invoice +0.05 + date proximity +0.10. Hard cap at 0.99 (displayed as 98%).

A receipt can score 0.98 — the highest possible — and still be wrong. The formula rewards structural signal overlap (invoice number present, cost code matches, dates near) but cannot detect:

• Vendor entity drift (subcontractor changes LLC name mid-project)
• Change orders that retroactively reclassify cost codes
• Split invoices where amounts don't cleanly divide
• Retainage releases appearing as new line items months later
• Handwritten notes overriding machine-readable fields ("bill to next draw")

98% from arithmetic ≠ 98% from judgment. Deterministic scoring is a fast first pass; it becomes a false floor when treated as terminal confidence.

Evidence

• The matcher pipeline is purely deterministic — an LLM tie-breaker is wired (gpt-4.1-mini) but dormant (no OPENAI_API_KEY deployed to the edge function). A post-match LLM anomaly detection joint does not exist.
• The confidence formula (processor.ts:576) is additive with no subtractive signals — nothing can lower confidence once base signals match.
• Operator directive: "make sure any efficient matching process is getting a review by an LLM" — the efficiency is the fast first pass, the LLM is the integrity check.

Resolution

Two joints: 1. Enable dormant LLM tie-breaker (already wired, config deploy only): activates gpt-4.1-mini candidate disambiguation when multiple options exist. 2. Wire post-match anomaly detection (new code): LLM review between deterministic scoring and status: proposed. Catches construction-domain edge cases that arithmetic can't encode.

Relation

• ORA-2026-0062 (world model fidelity): high confidence on a wrong match = map-confabulation. The matcher says "98% sure" but the world may disagree.
• feedback_llm_at_pipeline_joints.md: operator standing directive that efficient matching gets LLM review