Claim-routed reading may beat reading-first for synthesis notes
Type: kb/types/note.md
Status: conjecture. Motivated by one practitioner account (Learning By Writing); untested in this KB. Treat it as a hypothesis with a named test, not a finding.
The default multistage path reads the sources, then synthesizes: gather everything relevant, then find the claim. Karnofsky's inversion is to write a provisional bottom-line claim first — before being "qualified" — and let that claim decide what to read next: identify the subquestion most likely to change the claim, read narrowly toward it, revise, repeat. Reading is in service of writing, not the reverse.
The conjecture is that, for a note synthesizing several sources, claim-routed reading produces a better-warranted note at lower context cost than read-then-synthesize.
Mechanism: the claim is a selection pressure
A written claim names what would overturn it. That turns an open-ended "read the sources" into a targeted search: read the passage that could move the claim, skip the passages that could not. Warranted reader update is the objective, and a committed provisional claim supplies the selection pressure that an open gathering pass lacks — a bare writing prompt does not determine the contribution, but a provisional claim does, provisionally.
Two benefits are hypothesized. First, context economy: comprehensively reading every source spends bounded context on material that never bears on the claim, while routing reads only what could change it — relevant where context is the scarce resource. Second, warrant: a note built by surviving a sequence of claim-changing challenges retains the reasoning that selected it, rather than a pile of gathered facts whose relevance to the claim was never tested.
Failure mode: anchoring
The same claim that routes reading can bias it. If the provisional claim only ever selects confirming passages, the loop launders a first guess into a fake conclusion. Karnofsky's countermeasure is to route toward the disconfirming subquestion on purpose, and to periodically flip the hypothesis and argue the other side even when unconvinced. Without that, claim-routing is worse than reading-first, not better: it is passive assent dressed as inquiry. So the conjecture holds only for a routing loop that spends its selection budget on what could break the claim.
Scope
The claim is about multi-source synthesis notes — the ingest-and-multistage register where a note's value is the selected relation among sources. It does not apply to single-source capture, faithful transcription, or a note whose contribution is already determined before any reading.
How this could be wrong
The discriminating test is a comparison, not an intuition: build the same synthesis note both ways — claim-routed and read-then-synthesize — and compare the warranted update and the context cost. If routed reading yields the same note, it bought only a feeling of focus. If it yields an anchored note that never seriously read against its first claim, the failure mode dominates and reading-first is safer. The conjecture is that a disconfirmation-routed loop beats reading-first on warrant per unit of context; anything less refutes it.
Open questions
- Can an agent reliably pick "the subquestion most likely to change the claim," or does that judgment need the comprehensive read the routing is trying to avoid?
- Does routing belong as a replacement for the multistage workflow's reading stage, or only as a heuristic inside it?
Relevant Notes:
- Warranted reader update is the objective of substantive writing — grounds: routing is a search strategy for the update, and the provisional claim is what exerts the selection pressure
- context efficiency is the central design concern in agent systems — grounds: reading only claim-relevant passages spares bounded context
- A bare writing prompt does not determine its intended contribution — contrasts: a provisional claim supplies, provisionally, the contribution direction a bare prompt leaves open
- Learning By Writing — abstracted-from: Karnofsky's hypothesis-first investigation loop that motivates this conjecture