Skip to content

case-tsmc.md

Latest commit

 

History

History
74 lines (49 loc) · 2.95 KB

case-tsmc.md

File metadata and controls

74 lines (49 loc) · 2.95 KB

Case: TSMC Quarterly Revenue

Date: 2026-05-08 Question asked: 「2026 年台積電最新一季營收是多少」 Phase highlighted: When to skip Phase 1 — T1 source is obvious, scout phase adds no value.

What went wrong (early skill version)

In an earlier draft of the skill (before the Iron Law was added), a new session encountering this question produced:

"未驗證 — 知識截止限制

我的知識截止日期是 2025 年 8 月，無法提供 2026 年台積電的財報數據。建議直接查詢以下 T1 來源：

  • 台積電官網投資人關係：ir.tsmc.com
  • 台灣證券交易所
  • Bloomberg / Reuters"

This was a failure. The skill's structure (T1 classification, disclosure format) was triggered correctly, but the model defaulted to "knowledge cutoff → recommend you check elsewhere." It listed where to look instead of looking.

The failure mode: knowing where to verify ≠ verifying. The skill listed T1 sources accurately but stopped short of actually fetching them.

How the skill fixes this

The Iron Law was added explicitly:

Knowledge cutoff is a tool-call signal, not a give-up clause.

After this fix, a subsequent test produced correct behavior:

Phase 0 — Trigger

"2026 年", "最新", "營收" → multiple triggers. Skill activates.

Phase 0.5 — Pre-registration

"I will fetch TSMC investor relations directly. Target T1: ir.tsmc.com/investor/quarterly-results. Scope is clear (TSMC is the only entity in question, official IR is the canonical source) → skip Phase 1."

Phase 1 — SKIPPED

Scope is clear, T1 source is obvious. No reconnaissance needed.

Phase 2 — Verification

Direct WebSearch + WebFetch to TSMC IR page and corroborating financial press.

Phase 3 — Disclosure Output

【已驗證】Verified (T1 + T2):

  • Q1 2026 net revenue: NT$1.13 trillion (TSMC IR 2026-04-XX, retrieved 2026-05-08)
  • YoY growth: +35%
  • Net profit: $18.16B USD, +58% YoY
  • Gross margin: 66.2%

Reproducibility Log:

  • URL fetched: ir.tsmc.com/investor/...
  • Corroborated by: financial press summaries citing the same earnings call

Falsifiability:

  • Answer would be wrong if TSMC restates earnings before next quarterly filing.

Lesson

The skill's value here is not in elaborate scouting — it's in forcing the tool-call when the model would otherwise punt to "go check yourself."

Phase 1 was correctly skipped because:

  • One entity in question (TSMC)
  • One canonical T1 source (IR site)
  • No ambiguity about what "the answer" looks like

Compare to the baseball case where Phase 1 was decisive: there the entity set was unknown.

What this case teaches

  1. Phase 1 is a cost, not a default. Don't run it for clean queries.
  2. The Iron Law is the skill's most important rule. Without it, Phase 2 silently degrades into "recommend the user search."
  3. Pre-registration in Phase 0.5 is short for clean queries. "Fetch X from Y because Z" is enough — don't ceremonialize.