Decomposition
TL;DR
Decomposition splits one hard question into several answerable ones, answers each on its own evidence, and only then recomposes. As a graph it makes the split explicit — a dedicated node proposes the parts, each part is solved in its own scoped step, and a coverage gate refuses to recompose until every part has an answer and no part of the original question fell between them. Reach for it whenever a question is really several questions wearing one sentence — which, in regulated work, is most of them. It is also the strategy the other strategies run on: every IRG graph is a decomposition that holds still long enough to be inspected.
The question that was four questions
“Should we refund this?” A reviewer answers it in one motion — and the answer is defensible right up until someone asks which part she decided. Was the transaction authorized? Was the dispute timely? Does a liability cap apply? Was provisional credit due in the meantime? Her single yes-or-no silently answered all four, and if any one of them is later shown wrong, nobody can tell whether the error infected the rest.
A careful reviewer works the other way: four questions on four lines, each answered against its own evidence, the final determination assembled from parts that can each be checked. If the timeliness answer falls, the other three stand — and everyone can see exactly what needs redeciding.
An unguided model answers the compound question in one fluent pass. The decomposition graph is required to name the parts first — and cannot conclude until every named part has its own answered, evidenced line.
How the graph works, step by step
Name the parts. A dedicated node proposes the decomposition itself: which separable questions make up the compound one, and what evidence each will need. Making the split an explicit output is the point — a bad decomposition can now be seen and challenged, where in a single pass it is invisible.
Solve each part in its own scope. Every part gets its own step with only the evidence that bears on it. Scoping is what buys independence: the answer to “was it timely?” cannot lean on the vibe of “was it authorized?”, because the step never sees it.
Recompose and evaluate. The parts come back together, and a node evaluates the assembly: do the partial answers actually compose into an answer to the original question, or did the decomposition drop a piece — an interaction between parts, a remainder no part owned?
Arbiter and coverage gate. The arbiter rules on the recomposed whole, and the coverage gate enforces the strategy’s contract: every named part answered, every answer evidenced, nothing in the original question left orphaned. A missing part is not a footnote; it is a failed gate.
The gates, operationally
| Gate | What it checks | Fails when |
|---|---|---|
| Decomposition gate | The proposed parts are separable, individually answerable, and jointly cover the original question | Parts overlap so heavily that one answer decides another, or an aspect of the question belongs to no part |
| Scope gate (per part) | Each part’s answer cites only evidence admitted to that part | A partial answer leans on material outside its scope — the compound question sneaking back in |
| Coverage gate | Every named part has an answer and the recomposition addresses the question as originally posed | Any part is unanswered, or the assembled conclusion answers a narrower question than was asked |
Where it fits — three use cases
1. Element-by-element adjudication
Most regulated determinations are compound by statute: a Reg E error claim decomposes into transaction type, authorization, timeliness, and liability tier. Decomposition mirrors the statute’s own structure, which is why an examiner can read the trace — the graph’s parts are the regulation’s elements.
2. Onboarding and KYC files
“Is this customer acceptable?” is identity, beneficial ownership, sanctions exposure, and expected-activity coherence — four evidence sets, four thresholds. Decomposition keeps a strong identity answer from papering over a thin ownership answer, which is precisely the failure aggregated judgments invite.
3. Exam and audit preparation
A finding like “controls are inadequate” becomes tractable only when split into which control, against which requirement, on what evidence. Decomposing the assertion produces the workpaper structure directly — each part is a row an auditor can verify.
When to reach for it
Reach for decomposition when a question hides several questions with different evidence, different thresholds, or different owners — and especially when a later challenge will land on one part rather than the whole. If a wrong sub-answer should be correctable without redeciding everything, the parts must exist as parts. Skip it only when the question is genuinely atomic; decomposing an atom just manufactures ceremony.
- Decomposition as an output
- The split itself is produced by a node and recorded — so a bad decomposition can be challenged, where a single pass hides it.
- Scoped independence
- Each part sees only its own evidence, so partial answers cannot contaminate each other through shared context.
- Coverage
- The gate condition that no aspect of the original question is left unowned by a part — the failure mode of lazy splits.