Layer 0 runs first and determines what type of company this is: Mature Earner, High-Growth Profitable, Narrative/Platform, Pre-Profit Growth, Deep Value/Turnaround, or Dividend/Income.

Different company types need fundamentally different valuation methods. A utility and Tesla cannot be valued the same way.

The classification drives which valuation methods are used and how they're weighted in the final synthesis.

What is EPV? — Bruce Greenwald's model: what is the company worth if it never grows again? Uses normalized operating income divided by cost of capital.

Why it matters: EPV is a floor. If the stock trades below EPV, you're getting future growth for free — the market is pricing the company as if it will shrink. If above EPV, you're paying a premium for expected growth.

Excess cash is added on top (cash minus short-term debt) — that's money shareholders could theoretically receive today.

How it works: Instead of estimating fair value, this flips the question — "what growth rate would justify today's price?"

Uses binary search to find the FCF growth rate that makes the DCF model equal the current market cap. Then compares that implied growth to projected growth from Layer 3.

The gap is the signal:
  Implied < Projected → market underprices the growth (potential upside)
  Implied > Projected → market expects more growth than the data supports (risky)
  Implied ≈ Projected → price fairly reflects expected growth

Two-pass AI investigation. Pass 1: Claude reads all financial data and identifies what's unusual about this specific company — the questions an investor needs answered. Pass 2: Each question is investigated (via web search when available, Claude knowledge otherwise), then everything is synthesised into a 6-dimension analysis.

Full audit trail: Every question generated, every search query, every source hit, and all reasoning is logged with timestamps. Expand the investigation log to see exactly what was asked and why.

If Claude API is unavailable, this layer is gracefully skipped and the synthesis proceeds using quantitative methods only.

For narrative/platform stocks, fair value is meaningless. Instead, this layer asks: what does the market believe, and is that belief reasonable?

Claude evaluates: what each business line needs to generate, historical precedent for the implied growth, what must go right, what could go wrong, and price sensitivity (where it becomes a no-brainer vs clearly overpriced).

Verdict is one of: Reasonable Premium, High Conviction Required, Priced for Perfection, or Disconnected from Fundamentals.

How the composite works: Three valuation methods each produce a fair value. They're weighted: 50% DCF, 25% EPV, 25% Anchored PE.

Signal feedback loop: Strong signals now adjust the fair value itself — not just the verdict threshold. AI-identified tailwinds increase the DCF growth rate, hostile macro raises the discount rate, and competitive moat strength adjusts the PE premium. EPV stays untouched as the growth-agnostic floor.

Adjustment limits: Growth rate ±10pp, discount rate ±2pp, PE premium ±0.15 — prevents runaway estimates while still allowing signals to materially move the fair value.

Method agreement matters: When all three methods point the same direction, the signal is strong. When they disagree, treat the result with caution.

The verdict isn't "buy" or "sell" — it's whether the current price already reflects the company's fundamentals.