Skip to content

failure-log.md

Latest commit

 

History

History
89 lines (75 loc) · 4.25 KB

failure-log.md

File metadata and controls

89 lines (75 loc) · 4.25 KB

Failure Log — Kenji Watanabe Personal Life System

Purpose: Record errors and near-misses in my AI-assisted personal planning.
Owner: Kenji Watanabe
Note: In personal planning, "failures" are often softer than in professional work — but they still cost time, money, or wellbeing if not caught.

Failure Log

F-001 — AI gave a tax estimate without my real numbers

  • Date caught: 2025-08-20
  • Category: F3 — Numerical Error
  • What happened: I asked AI to help me think about whether to increase RRSP contributions. AI said "based on a typical salary around $90K you'd save about $1,200–1,500 in taxes." This was not based on my actual numbers. AI invented a plausible range.
  • Impact: I almost made a contribution decision based on the estimate. When I checked with my accountant, my marginal rate and situation were different enough that the figure was meaningfully wrong.
  • Root cause: I hadn't set up canonical-numbers.md yet. No constraint on AI making estimates.
  • Fix: Got real numbers from accountant. Added rule: AI does not produce any financial estimate unless based on canonical numbers.
  • System change: Created canonical-numbers.md. Added rule to running document.
  • Status: ✅ Resolved

F-002 — Pillar bleed: relocation logistics ate a career-focused session

  • Date caught: 2026-01-15
  • Category: F1 — Context Drift / Pillar Interference
  • What happened: Session was supposed to be focused on career: preparing for Fenwick interview process. AI was helpful at the start, then I mentioned "oh and I'm also dealing with the move" — AI pivoted to logistics. The last 30 minutes of a 60-minute session was about Toronto neighbourhood research.
  • Impact: Career prep was incomplete. Had to do another session the next day.
  • Root cause: I opened the door by mentioning the move. AI responded helpfully to what I said, not to the original session goal. No explicit pillar focus rule was active.
  • Fix: None for that session. For next career session, focused it successfully.
  • System change: Added to running document rules: "One pillar per session unless I explicitly open cross-pillar. If I mention another pillar in passing, AI acknowledges and returns to focus."
  • Status: ✅ Resolved

F-003 — AI recommended a course of action on career decision

  • Date caught: 2026-02-28
  • Category: F4 — Boundary Violation (AI overstepping)
  • What happened: When discussing my 90-day plan at Fenwick, AI said "you should push for a promotion path conversation early — managers respect ambition shown in the first 90 days." This is career advice framed as a directive. My career boundary says AI offers options, not directives.
  • Impact: No concrete harm — I noticed and flagged it. But if I'd been less aware, I might have taken advice that wasn't right for my specific manager and culture.
  • Root cause: No explicit prompt constraint in that session. AI defaulted to helpful advice-giving mode.
  • Fix: AI used ARGUMENT FLAG correctly when prompted to ("I am suggesting X — this is a recommendation, do you want to accept it?"). I declined.
  • System change: Added explicit constraint to session prompts: "You may offer options but not recommend what I should do on personal career, health, or financial decisions."
  • Status: ✅ Resolved

Near-Misses

Date What was caught Who caught it Action
2025-09-10 AI suggested a specific investment fund by name Kenji flagged immediately Replaced with category discussion; reminded AI of investment boundary
2026-01-30 AI described knee pain management advice as if providing medical guidance Kenji flagged Redirected to "here's what physiotherapists typically suggest" framing
2026-02-15 AI included a "compare yourself to peers your age" savings benchmark without being asked Kenji flagged Reminded AI: track against own baseline only (Principle 4)

Summary

Category Count Resolved
F3 — Numerical errors 1 1
F1 — Context/pillar drift 1 1
F4 — Boundary violations 1 1
Near-misses 3 3
Total 6 6