Decompose a note into premises and counterexample each

Type: kb/types/instruction.md

Experimental, run by hand. This method tests the nodes of a note's argument — the premises its central commitment takes as given — not the edges between them. It assumes every stated inference is valid and asks a different question: is each premise the commitment rests on itself true, or does a concrete counterexample defeat it? This is the "explanation as a falsification surface" move — a commitment is only as strong as the weakest premise it silently assumes. It is not a review-system gate or snapshot-anchored assay: it writes no evidence or freshness-baseline state and is not wired into the review system. Write the report; do not touch the note.

Unlike reconstruct a note's composition friction, which tests whether each inference holds (the edges), this tests whether each premise holds (the nodes). The two are complementary: an argument can have valid inferences resting on a false premise, which the friction check passes and this check catches. Run both when a commitment must be load-bearing.

Run it in a fresh sub-agent, a different runner than wrote the note, so the checker has no sympathy for the note's framing. Separation is what gives the check teeth: the runner that assumed a premise while writing will assume it again while reading.

The caller owns reviewer lifecycle. After the report has been written and verified, close, terminate, or release the checker with the harness lifecycle operation. The checker is a single-use context and must not receive follow-up work.

The hard rule

Do not emit an overall "sound / unsound" or pass/fail verdict, and do not accept the note. The product is routed attention: the premises that a counterexample defeats or dents, and — for each — whether its failure is local or global. Default to DOUBTFUL when uncertain. A premise earns HOLDS only when an active counterexample hunt fails; do not extend it the benefit of the doubt the way fluent reading does.

Step 1 — State the central commitment

State the note's central commitment in one sentence. Match its register:

  • claim — the proposition the note asserts is true;
  • definition — the boundary it draws and asserts is the right cut;
  • procedure — the outcome it asserts following the steps produces.

Step 2 — Decompose into load-bearing premises

Enumerate the premises the commitment rests on and takes as given — including background premises the note treats as obvious and never argues for. What counts as a premise follows the register:

  • claim — a proposition that must be true for the claim to hold;
  • definition — a commitment that makes this the correct boundary (that the distinguished cases differ in the way named, that the distinction is not idle, that nothing important is mis-sorted);
  • procedure — an assumption under which the steps produce the intended outcome (about inputs, the executor, the environment, or the world).

List the premises the argument consumes, not every true sentence in the note. A premise the note never states but the commitment cannot survive without is the most valuable to surface.

Step 3 — Counterexample each premise

For each premise, actively hunt a concrete counterexample or a specific reason it could be false — a case, input, or situation in which the premise does not hold. Record per premise:

  • HOLDS — an active hunt found no defeater;
  • DOUBTFUL — a plausible reason or edge case dents it (state it);
  • DEFEATED — a concrete counterexample makes it false (state it).

Counterexample by register: for a claim, a case where the premise is false; for a definition, a case the boundary mis-sorts (over- or under-includes) or a distinction that changes nothing; for a procedure, a situation where the assumption fails and the steps misfire.

Read each premise at its stated modality (kb/notes/COLLECTION.md, "Claim modality"). A premise stated as a tendency ("usually", "most", "under conditions C") is dented or defeated only by prevalence-shaped evidence — the exceptions being common or the ordinary case going the other way — never by one instance it already concedes. A premise stated as a first-order model with adequacy commitments is attacked through those commitments — an exception its domain treats as ordinary unmarked practice, or the model losing dominance — not by conceded, accounted-for exceptions. A premise with no stated mode reads as universal, and one genuine counterexample defeats it.

For every non-HOLDS premise, also record the counterexample shape: instance (one concrete case), prevalence (evidence the exception is common or ordinary), or priced-exception (the case is marked, fenced, or charged for in its own domain). The shape is routing information for whoever repairs the note — a prevalence-shaped defeat of a universal premise points at a statistical reframe, and a priced-exception defeat points at an ideal-type candidate — not a verdict, and not this method's call to make.

Step 4 — Route each failure local or global

For every DOUBTFUL or DEFEATED premise, classify the scope of its failure:

  • LOCAL — the failure defeats only this premise; the central commitment can still stand on the remaining premises or be rescued by a qualification (narrow the scope, add a condition). The revision changes a premise and keeps the commitment.
  • GLOBAL — the failure propagates to the central commitment, so the commitment itself fails as stated. The revision must weaken, rescope, or retire the commitment, not just patch a premise.

A GLOBAL failure on a load-bearing premise is the strongest signal this method produces; surface it first.

Output

Write to kb/reports/premise-decomposition/<note-name>.premises.md. Mutate nothing else.

# Premise decomposition: <note title>

**Note:** <path>
**Register:** claim | definition | procedure
**Central commitment (one sentence):** <...>

## Premises and counterexamples
1. **<premise, stated as a proposition, at its stated modality>** — <HOLDS | DOUBTFUL | DEFEATED> — <the counterexample or reason> — <LOCAL | GLOBAL if not HOLDS> — <instance | prevalence | priced-exception if not HOLDS>
2. ...

## For the human
<one line: the premise to look at first — the global defeater if any, else the load-bearing doubtful premise most worth the author's attention>