After twelve years in the secondary market, I had accumulated a set of investment “instincts” — what to look for in stock selection, when to buy, how to price, when to sell. But these judgments lived scattered in my head, never systematized.
This time I used AI for a round of self-distillation: 12 rounds of structured interviews + benchmarking against 10 masters + multiple rounds of arbitration and patches, and out came a three-layer investment operating system — from philosophical axioms to machine-readable rules.
Why I did this
The InvestPilot reflection. I had previously built an automated research framework called InvestPilot (a 9-step pipeline). The technical execution layer was solid — state-machine constraints preventing the agent from skipping steps, self-computed valuation to prevent hallucination, Monte Carlo for distributions. But running it revealed a fundamental problem: the philosophy layer wasn’t refined. Investment rules were scattered across prompt templates, with no explicit “constitution,” no traceable foundation for “why we do it this way.” The system knew “how,” but not “why.”
So I went back and refined the philosophy layer first. That is where this framework came from.
The process
Step 1: 12 rounds of structured interviews. Using AI as the interviewer, I went deep on 12 themes across the full investment chain: research process, success cases, failure reviews, cross-market differences, valuation, portfolio positioning, financial analysis, cognitive psychology, risk and cycles, short selling, discipline, and management assessment. Each round drilled down to the operational level — not just staying at the level of principles.
Step 2: Benchmarking against 10 masters. I selected 10 masters as external reference points — 5 Chinese (Feng Liu, Zhang Lei, Li Lu, Duan Yongping, Qiu Guolu) + 5 Western (Buffett/Munger, Klarman, Lynch, Fisher, Marks). I cross-referenced each principle distilled from the interviews against their thinking: which had backing, which were uniquely mine, which conflicted and needed arbitration.
Step 3: Arbitration. Conflicts emerged from the benchmarking. For example, Duan Yongping and Zhang Lei believe “at a ten-year horizon you don’t need catalysts,” while my framework insists on a 0-3 month catalyst window — resolution: different time scales, not a real conflict. Short selling was ultimately removed from the system (long-only). Concentration vs. diversification: concentration is the equilibrium solution of research bandwidth, not a belief. Every arbitration was recorded with reasoning.
Step 4: Blind-spot patches. After the first version, I discovered gaps in macro/cycles/short-selling/concentration/management assessment. I supplemented with four rounds of blind-spot interviews and added an entire section on “Macro, Cycles, and Portfolio Philosophy.”
The framework
A three-layer architecture:
Layer 1: Philosophy (Constitution) — the highest constraint. When any rule conflicts with this layer, this layer prevails.
Five axioms:
- Odds are the universal currency — stock selection, pricing, position sizing, and selling all reduce to one question: is the RRR (probability of upside × magnitude ÷ probability of downside × magnitude) sufficient?
- Alpha’s true core is the cognitive gap on competitive advantage — the money you make comes from the gap between “your understanding of a company’s quality” and “the market’s pricing of it”; in quality (slow variable), not in short-term performance
- Value needs a catalyst; time is a cost — cheap doesn’t generate returns; re-rating does
- Percentile is just a pricing tool, not an alpha source — it only prices companies “already confirmed healthy by quality research”
- Anything that cannot be verified by public, trackable data does not enter the decision — no private channel checks; research output is a set of atomic hypotheses + tracking indicators for each
Plus epistemology (7 rules about “knowing”), value hierarchy (5 priorities for conflicts), three-tier red lines (hard-reject / warning / weight adjustment), and a human-machine constitution.
Layer 2: Framework (Methodology) — six-module decision chain: 100-name research pool → triage (7 trap signals) → deep research (industry 4D + competitive advantage + management + reverse verification) → valuation pricing (atomic hypotheses + anchoring scorecard + seven-percentile distribution) → portfolio management (weekly refresh + Fisher exemption + market thermometer) → review and calibration.
Layer 3: Execution (Rule library) — machine-readable specification. Each rule = ID | rule | threshold | action | executor, classified as HARD-FAIL (machine rejection) / WARNING (human deep-dive) / WEIGHT (weight adjustment).
The most important designs
Odds as universal currency. A single RRR unifies the entire system — no decision can bypass the payoff calculation. Feng Liu’s “odds-first” approach is isomorphic.
Alpha = cognitive gap on competitive advantage. The judgment draws on several sources: Buffett/Munger moats, Fisher scuttlebutt, Duan Yongping business models, Zhang Lei research-driven investing. Quality cognitive gaps correct slowly, so you can position on the left side without racing.
Human-machine constitution. Machines counter biases, enforce discipline, do breadth screening and valuation; humans retain competitive-advantage judgment and final sign-off — “machine cognition cannot yet reach the judgment of a company’s core competitiveness.” Master personas (Munger seat “how does this die” / Klarman seat “opportunity or trap”) are a red-team adversarial layer, not the decision constitution.
100-name research pool. The portfolio is not a static “buy decision” but a dynamic product of the weekly full-pool RRR ranking — enter the top ranks, enter the portfolio; fall out, exit.
Back to InvestPilot
This framework is the foundation for rebuilding InvestPilot. The old version had its philosophy implicit in prompts; the new version makes philosophy an explicit constitutional layer, where every rule is traceable to an axiom, an interview round, a master’s endorsement. Rules are auditable and iterable.
Spectrum positioning
Fisher/Buffett-Munger quality research (depth) × Klarman/Feng Liu odds-based pricing × Lynch/Qiu Guolu percentile rotation × proprietary catalyst engineering. Closest to Klarman on the spectrum — catalyst-driven value investing, engineered for the Chinese market.