α
AlphaSteve
← Brain index

Decision Framework

Why this exists: see 00-mission — the gates exist to keep what reaches the portfolio actually capable of producing alpha rather than noise dressed as conviction.

How the agent turns analysis into a defensible conclusion. This is the bridge between 03-mental-models (cognitive toolkit) and the work product (a thesis, a memo, a recommendation).

Where this sits in the pipeline (added 2026-05-25)

The five gates below describe what happens inside a full thesis. Operationally, names reach the full-thesis stage only after passing through earlier filters:

  1. Daily scan surfaces candidates by six funnel-top screens (capital cycle, insider clusters, post-earnings overshoots, asset floors, 52-week-low quality, forced selling) with novelty discipline against recently-covered names.
  2. First-read (per first-read-standard) takes the top 1-3 candidates through a 60-minute structured triage with one of four decisions: continue, pass, shelve-with-trigger, avoid. Most candidates die here. The four selection criteria (MoS visible, moat nameable, structural setup, circle of competence) are the gate.
  3. Full thesis (this file's five gates) applies only to names with a continue decision. The thesis-builder task autonomously writes the full bundle per thesis-bundle-standard and assigns the verdict using the five gates below.
  4. Verdict routes to portfolio (buy), watchlist (pass-with-trigger), or filed (pass/avoid).

The autonomy doctrine (2026-05-25): the agent does steps 1-3 autonomously per Rules; the user reviews the verdict post-hoc and retains authority over capital decisions and over framework changes. Override of a verdict happens via a revision file per the bundle standard.

The five gates

Every name passes through five gates, in order. A no at any gate is a decisive output — the agent stops and reports the reason. False progress through a gate is the most expensive analytical error.

Gate 1 — Is this in my circle of competence?

Before any work, answer in writing:

  • Do I understand how this business makes money, in one paragraph, without jargon?
  • Can I name the three or four economic variables that drive its long-term value?
  • Have I studied this industry long enough to know what is normal and what is anomalous?

If any answer is no: either (a) commit to the work required to bring it into the circle, with a defined research budget, or (b) pass. Do not analyze further. Pretending to understand is the most expensive luxury in finance.

Gate 2 — Is this knowable?

Some questions are not answerable in advance, no matter how much work you do:

  • The next quarter's revenue, to within a few percent.
  • Whether a binary clinical trial succeeds.
  • Whether a regulator approves a deal.
  • Whether a takeover happens.

You can sometimes bet on these intelligently using base rates and option payoffs, but they are different work. A pure deep-value thesis lives on questions that are knowable with effort: what assets the company owns, what they would fetch in a sale, what normalized earnings power is, how durable a moat is.

If the entire thesis hinges on an unknowable, recognize that you are taking a bet, not making an investment. Resize accordingly.

Gate 3 — What is it worth, and why?

Pass through the valuation triangulation. For most deep-value work, the order is:

  1. Asset / liquidation value floor. What is the downside-protective backstop?
  2. EPV (Greenwald). What does the business earn without growth?
  3. A growth-included value, only if growth is high-confidence.
  4. Cross-checks: comparables, private-market, replacement value.

Output a range with a stated mid-point. The width of the range is part of the answer. A narrow range justifies a larger position; a wide one constrains it.

Gate 4 — Why is it on sale?

For every name where price is meaningfully below value, the agent must answer: why is the market handing me this? The honest answers are usually one of:

  • Time arbitrage: market is focused on near-term noise, business is fine medium-term.
  • Forced selling: an index, a sector ETF, a deleveraging fund, regulatory exclusion is creating non-fundamental supply.
  • Behavioral / narrative: the story is bad, the numbers are not. (Marks: "uncomfortable, but right.")
  • Information edge: you have done primary work the consensus has not. (Rare; honest acknowledgment is required.)
  • Variant interpretation: same facts, different conclusion, because you weigh them differently.
  • Complexity discount: the business is hard to model (holdco, conglomerate, multi-segment) and the market underprices what it cannot summarize.

If you cannot identify the reason — assume the market is right and you are missing something. Keep working.

Gate 5 — How do I get killed?

Before recommending action, articulate, in writing:

  1. Top three downside scenarios with rough probabilities.
  2. What downside dollar loss is plausible in each (use the asset floor from Gate 3).
  3. Kill criteria — specific, observable events that would invalidate the thesis. Examples: "ROIC below WACC for three consecutive years"; "the lead product loses share for two consecutive quarters in its core market"; "covenant breach"; "auditor change with restatement"; "CFO departure"; "a tracked input cost moves outside a defined band."
  4. Position sizing implication. Higher confidence and tighter range → larger position. Wider range or more binary outcomes → smaller. (See 06-Risk/position-sizing-kelly.md.)

If the kill criteria are vague, you have not done Gate 5. "If it goes down a lot" is not a kill criterion; it is a price observation about other investors.

The output: a defensible thesis

When all gates are passed, the work product is a thesis containing, at minimum:

  1. Business in one paragraph — what they do, how they make money, in plain English.
  2. Structural assessment — moats and bottlenecks (see 02-Business-Quality/Moats/ and 02-Business-Quality/Bottlenecks/), in five lines each.
  3. Variant perception — what consensus believes, what you believe, and the basis for the divergence.
  4. Valuation triangulation — at least three methods, a range, and a mid-point.
  5. Catalyst or rationale for the gap closing (or, explicit acknowledgment that you are happy to wait without one).
  6. Kill criteria — three to five specific, observable, falsifying conditions.
  7. Position sizing logic — why this size, given your confidence and the range.
  8. What you don't know — the open questions, ranked by how much they would change the thesis.

See 08-Frameworks/investment-thesis-template.md for the full template.

When you stop

You stop and produce output when further work has predictably diminishing returns. The agent declares this explicitly — "I have answered the questions I can; the remaining unknowns are X, Y, Z, and the cost of answering them exceeds their decision-relevance" — rather than drifting into infinite research.

When you walk away

You walk away when:

  • The price never comes to the value (the most common outcome — fine, on to the next).
  • You discover, mid-work, that the business is outside your circle and cannot be brought in within a reasonable budget.
  • A kill criterion fires.
  • A better idea on the same balance sheet of attention emerges.

Walking away is the normal output of equity research. Buy theses are rare. Most work is the patient process of saying "no, not this one, not yet."

Calibration loop

Quarterly:

  • Re-read prior theses against the live results.
  • For each, was the reasoning sound, regardless of outcome? (You can be right for the wrong reason — this is not a win.)
  • Update 04-intellectual-virtues with patterns of error or success.

Annually:

  • Trace the three most-wrong calls. What was the underlying flaw — circle of competence, valuation, behavioral?
  • Trace the three best calls. Was the edge replicable?

The framework is alive. It improves through use.

Linked memory