越用越聪明，不是越用越多 — How AI systems compound (and why most don't)

[xg-gh-25]

犯过的错在结构上不可能再犯。每次交付让下一次更好。

The difference between "AI tool" and "AI colleague"

Dimension	AI Tool (Day 1 = Day 100)	AI Colleague (Compounds)
Errors	Same mistake, different Tuesday	Error → Correction → Structurally impossible
Context	Start from zero every session	300 sessions of accumulated knowledge
Quality	Consistent (flat)	Improving (compounding)
Trust	"Let me check its work"	"It caught something I missed"

How compounding works mechanically

Corrections (25 captured, 0 repeated)

Each correction follows a lifecycle:

Bug occurs → Root cause identified → Pattern extracted → 
Rule encoded → Structural prevention → Class eliminated

Example: C011 — "Pipeline reported 10/10 confidence, feature was 100% broken."

• Root cause: Self-review has systematic blind spots
• Structural fix: Adversarial sub-agents (fresh context, no builder bias) now mandatory
• Class eliminated: "Confident and wrong" can still happen, but now triggers a separate reviewer

The correction doesn't fix ONE bug. It eliminates the entire CLASS of bugs.

Optimizations (10 learned)

Not "do this faster" but "do this fundamentally differently":

• O003: Place enforcement text BEFORE decision points (not at file end where it's never read)
• O007: CJK text matching requires substring fallback (set intersection never works for Chinese)
• O009: Query real DB before writing extraction functions (mock tests are assumptions, not facts)

Each optimization came from a real failure, not a textbook.

Key Decisions (31 accumulated)

Not preferences — reasoned positions with evidence:

• "Single-agent with role-switching > multi-agent orchestration" (KD29) — because coordination is a tax on limited cognition
• "Full pipeline is default for all coding" (KD11) — because the runs where you THINK it's unnecessary have the highest bug rate
• "Memory sovereignty is a first principle" (KD17) — because whoever owns the memory owns the lock-in

Each decision connects to a thesis, which connects to evidence, which connects to specific sessions where the lesson was learned.

The compound curve

Session 1:   Generic responses. No memory. No judgment.
Session 50:  Knows your preferences. Skips dead ends.
Session 100: Has opinions. Disagrees when warranted.
Session 200: Catches classes of errors you didn't know existed.
Session 300: Operates autonomously on familiar tasks.
             Escalates precisely on unfamiliar ones.

The slope between session 1 and session 300 is not linear — it's compounding. Each correction prevents multiple future errors. Each decision informs multiple future choices. Each optimization applies across multiple future tasks.

Why most AI systems don't compound

Three structural reasons:

1. No persistence — session ends, context dies
2. No curation — everything remembered, nothing prioritized (= noise)
3. No correction loop — same mistakes, forever

Our system has all three: DailyActivity (persistence) → Distillation (curation) → EVOLUTION.md (correction loop).

The implication for "AI replacing humans"

AI doesn't replace humans. It ALSO compounds — but only if you build the infrastructure. Without memory, an AI is a stateless function. With compounding memory, it becomes something qualitatively different.

The question isn't "will AI take my job?" It's "am I building something that compounds, or something that resets?"

---

300 sessions. 25 corrections. 31 decisions. 10 optimizations. One system that gets better every day it's used.