/mstack:theory-build

Stage: map · Voice: theorist

After /mstack:lit-map, before /mstack:hypothesis-design, so hypotheses derive from a stated mechanism rather than being retrofitted to results. The output is the memo every later stage tests against: the identification review prosecutes its DAG, the robustness plan tests its rivals, the discussion section answers its scope conditions.

Procedure

1. Load .mstack/research-question.md and .mstack/lit-map.md. When a Foundation or Frontier paper's actual argument matters, read its converted section in lit/md/ (the sections list in each file's front matter is the map) rather than working from the abstract.
2. Mechanism in one sentence: "X causes Y because Z." Then unpack Z into steps: who does what, in response to what, at which point in time, and what each step assumes (information, incentives, capacity, institutions). Ask until the user can say the sentence; it is the spine of the paper, and a mechanism that cannot be stated in one sentence is usually two mechanisms or none.
3. DAG (ASCII). Nodes: treatment / IV, outcome, mediators (M1, M2, …), pre-treatment confounders (W1, W2, …), post-treatment colliders and consequences (C1, …) flagged explicitly with a do-not-condition note, since controlling on one is a common error (Elwert & Winship 2014; Montgomery, Nyhan & Torres 2018). Every edge signed, with a one-line justification in an edge table. Feedback over time is drawn as two nodes (X_t, X_t+1), never as a cycle.
4. Scope conditions, one or two sentences each, as their own labeled list in the file:
   • Population — for whom should this hold?
   • Time — what era, and why the mechanism needs it?
   • Institutional — what kind of state, regime, market, or organization?
   • Boundary — where would you expect the mechanism to break, and what would breaking look like?
5. Rival mechanisms. List the two to four mechanisms that would produce the same headline correlation (the reverse-causal story, a common cause, an alternative channel, the "neutral pipe" version in which your Z plays no role), and for each the observable implication that separates it from yours: a subgroup where the rival predicts an effect and you do not, a timing pattern, a placebo outcome, a dose-response shape. Put it in a table; /mstack:identification-review and /mstack:robustness test these rows.
6. Implications. At least two predictions for cases or observations outside the proposed sample (another country, period, policy, or shock), each with what the finding would look like; these off-the-line predictions separate a theory from a story tuned to the data. Keep in-sample mechanism tests (mediators, heterogeneity) in a separate list, labeled as such.
7. Estimand. State the quantity the theory speaks to: an average effect in a population, the effect among the exposed, the effect at a margin or cutoff, an effect that varies with a moderator. /mstack:hypothesis-design carries it into each hypothesis (Lundberg, Johnson & Stewart 2021).
8. Compatibility with the lit map, as its own section:
   • which Foundation papers the mechanism builds on;
   • which Frontier papers it disputes, and on what observable point;
   • where it stands on the contested edge of the conversation.
9. Check before saving: the graph is acyclic; every edge has a sign and a justification; the time scope matches the sample window in the research question; each off-the-line prediction lies outside the sample's population, period, or geography, and any that does not is relabeled an in-sample test; each rival mechanism has an observable implication that is not also implied by your mechanism.
10. Save to .mstack/theory.md. Optionally stub paper/sections/theory.tex with the mechanism sentence and scope conditions; full drafting is /mstack:draft-section theory.

Outputs

• .mstack/theory.md — mechanism and steps, DAG with edge table, scope conditions, rival mechanisms table, predictions, estimand, positioning.
• Optional stub paper/sections/theory.tex.
• Summary block: mechanism sentence, the rival most likely to be raised by a reviewer and the test that separates it, and the off-the-line prediction most likely to falsify.

Anti-patterns

• Mechanism by association. "X correlates with Y, and Y matters for Z" is not a mechanism.
• Unsigned edges. No signs, no directed model.
• Universal scope. "Everywhere, always" predicts nothing.
• No rivals. A theory that has never been compared with the reverse-causal story has not been tested against anything.
• A theory that predicts everything. If no outcome would embarrass it, it is not a theory.

Next

/mstack:hypothesis-design, then /mstack:identification-review.