AI Iteration 1 — Comparison vs Golden Truth
2026-04-22 19:16 · 2 KB
RECALL SCORES
| Dimension | Score | Detail |
|---|---|---|
| Key item recall | 80% (12/15) | Found major monopolies. Missed: MGC BT resin, Meiko PCB, Hirose connectors |
| Financial metrics | 62% | PE, ROE, OpMar, D/E, FCF present. Missing: Forward PE sparse, PB sparse, Div Yield sparse |
| Value chain layers | 100% | All 13 layers mapped |
| Comparison table | YES | Present |
| Moat rating | YES | ★ system used |
| Conviction tiers | YES | HIGH/MEDIUM/LOW |
| Japan vs Taiwan | YES | Good comparison |
| Source citations | 0% | ZERO URLs in the entire report — CRITICAL gap |
GAPS (must fix in prompt v2)
1. Source citations: 0% — CRITICAL
Not a single URL in the report. Every claim is unsourced.
Root cause: Prompt says "cite sources with URLs" but doesn't make it a hard requirement per-claim.
2. Financial metrics incomplete
Forward PE, PB, Dividend Yield are sparse.
Root cause: Same as nuclear — metrics not explicitly listed as ALL REQUIRED.
3. Niche companies missed
MGC (BT resin inventor), Meiko (AI server PCB), Hirose (connectors) — all smaller/niche.
Root cause: Agent found the big monopolies but stopped before exploring niche layers.
4. Ticker format inconsistent
Generated report uses different ticker formats, making automated comparison difficult.
Root cause: Prompt doesn't specify "use XXXX.T format for all tickers"
DISCOVERIES
The generated report appears to have good depth on the major monopolies (TEL, Lasertec, Shin-Etsu, etc.) and maps all 13 layers. The quality of analysis is strong — the gap is in completeness of smaller names and sourcing.
PROMPT v2 CHANGES NEEDED (additive to nuclear findings)
6. Add: "EVERY factual claim must have a source URL — no exceptions"
7. Add: "Use XXXX.T format for all Japanese tickers consistently"
8. Add: "For niche/small-cap companies, search specifically for 'Japan {product} manufacturer listed stock'"