A note is an atomic step relative to the check that reads it

Type: kb/types/note.md · Tags: kb-maintenance, evaluation

A knowledge base written and reviewed by agents bounds note size twice, for two reasons that have nothing to do with each other.

The first bound is co-loading. Short composable notes maximize combinatorial discovery because a model's context is finite, so the number of notes that fit in it together sets the surface area for recognizing shared structure. One claim per note, sized to the reader's context.

The second bound is checking. A note is an atomic step when the claim, the material it cites, and the inference from that material to the claim can be checked in one pass by the check that will be applied to it. One checkable inference, sized to the checker's pass.

These are different quantities set by different mechanisms, and a note can satisfy one and fail the other. "Fits one pass" also has no content until the pass is named: a step is atomic relative to a particular check, which is an instance of the general rule that warranted autonomy is bounded by oracle domain — a checker's domain is the range of candidates it can assess with the required confidence, and the boundary of automation is the boundary of verification. The same note may be one step for a check that judges whether quoted passages support a claim and several steps for a check that judges whether an inference across all of them holds. This note states the rule for one check, Commonplace's grounding check, because that is where the unit of cost is clear; the same argument would need its own unit for any other check.

Placing the bound on the artifact rather than the review

The alternative placement is a reading budget on the review: let notes be as large as they are, and calibrate how much the reviewer may open. The two placements differ in their failure mode, not in the constraint they express.

A review-side budget has to be recalibrated whenever the corpus's evidence needs move, and its failure when it lags is silent. A reviewer that stops at its limit and returns a pass is indistinguishable, from the outside, from one that read everything and returned a pass. Disclosure conventions reduce this — a reviewer can name what it left unread — but disclosure reports the budget's boundary, not whether a finding was sitting behind it.

An artifact-side bound does not have that failure mode, because a conforming artifact fits by construction: the bound is held by a validator that fails the note, not by a reviewer that stops reading. The reviewer's limit still exists; on a conforming note it never binds.

Mechanism: for the grounding check, the unit is the unquoted source

The grounding check asks whether the material a note cites supports the claim it is cited for. Its cost is not links, and not bytes: it is the number of artifacts in which the reviewer must find the passage that carries the claim and then judge whether it does. A source that has been quoted verbatim is not one of them.

The reason is that Commonplace has moved that finding step into code. A quotation marked verbatim is resolved by substring match against the file it links, and a mismatch fails the note before any reviewer opens anything (ADR 046); the source-side material is itself held as retained quotes validated against the captured source (ADR 073). This is the general rule that a derived copy of recomputable truth must be checked or absent applied to citation: the quote is a copy of something held elsewhere, and it is admissible because a validator re-derives it and fails on mismatch. A paraphrase has no such derivation rule, and a citation cannot assert more fidelity than its capture preserved — a passage never captured word for word cannot be quoted into this route without returning to the source.

What is left for the reviewer after the quote is a bounded judgment: does this passage support this claim. It costs attention, and nine quoted passages cost more than one, but the cost is set by the author's choice of passage length and requires opening nothing.

Two limits follow, and both are prices already paid elsewhere in the system. Occurrence is not representativeness: the validator establishes that the words appear in the source, not that they mean there what the note uses them for, and a reader who suspects a passage was lifted out of a qualifying context has to open the source after all. And some claims cannot be carried by a bounded passage — claims resting on the structure of a whole argument, on what a source does not say, or on material distributed across a document. Commonplace marks those uses (snapshot required); such a source has no bounded passage and counts fully against the step.

A consequence worth stating, because it answers the obvious objection to bounding the artifact: a claim supported jointly by several sources, none sufficient alone, should be quoted rather than split. Splitting it across notes severs the support, so each fragment fails a check the whole would pass — the hazard the review-side proposal Exceeding a review budget splits the task, now archived, identified when it concluded that any partition would have to be by claim rather than by link. Under an artifact-side bound the author, who knows which passages jointly carry the claim, puts them on the page together and the check reads them together. This is where the two bounds meet: every quoted passage lengthens the note, so quoting spends against the co-loading bound. The bounds are independent as constraints but coupled through this remedy. A claim whose joint support cannot be quoted within a note the reader can still co-load has only the snapshot route or a split left, and the split has to be by claim, so that each fragment states what it establishes. Links to other notes are also not sources in this sense, because a linked note carries its own claim title and its own grounding review; a linked note discharges its own grounding, so a citing note owes it representation, not re-grounding.

The two bounds come apart in the corpus

Across 344 notes in kb/notes/ as of 2026-08-27, eight cite more than five distinct sources, the largest citing nine. All eight are single-claim notes: one proposition, established or bounded by reading six or more sources together. They satisfy the co-loading bound on any reading of it. They are not atomic steps for the grounding check, because checking one means opening six to nine artifacts and finding the claimed support in each. The converse case is available in principle — two independent claims, each grounded in one quoted passage, is one step for the grounding check and two notes under the co-loading bound.

The strongest evidence Commonplace holds about grounding-review reading limits reads differently under this placement. A paired experiment reviewed the same twelve notes under a five-link instruction and an uncapped one; four findings appeared only in the uncapped arm, after the reviewer had opened six, eleven, fourteen and sixteen artifacts. Read as a fact about reviewers, this says five is too few. Read as a fact about notes, it says the reviewer honoured its cap and the notes had put their support beyond it — which is the behaviour an artifact-side bound depends on, and under such a bound the four notes would have had to split or quote before review. Two of the four findings sharpen the reading: one was missing retained support for details attributed to a source, and one was an ingest's analytical prose treated as evidence beyond its retained quotes. Both are absences of a retained passage, and an absence of that kind is a validator's result rather than a reviewer's finding. Had the details been quoted, either the quote would match or the note would have failed validation before review began.

What kind of number a step size is

The bound does not fix its own N, but it does fix what kind of number N is, because three candidates are available and they justify different values.

N is not the point at which reviewer performance degrades. In the same experiment, uncapped reviewers read up to twenty-two artifacts and 322 kilobytes in one pass with no observed loss, and the three notes offering the most material all passed. Whatever a small step protects, on this evidence it is not a reviewer capacity exhausted at current corpus sizes. Nor is N the smallest count at which no miss happened to be observed; that number moves with every experiment and inherits its sample.

N is a convention, chosen so that a second checker — including a person — can redo the step. This is what a small step buys that a well-calibrated reviewer budget does not. A verdict resting on fourteen artifacts costs as much to audit as it cost to produce, so in practice it is trusted rather than audited; a verdict resting on three sources and six quoted passages can be re-checked from the same page. What the page re-checks is the support judgment — that these passages carry this claim; whether a passage was lifted out of a qualifying context still costs a trip to the source, so the audit the bound buys is of the step the validator did not do, not of the one it cannot. The value is not that the reviewer copes, but that the review becomes auditable — and Commonplace's own measurement that language-model verdicts disagree with themselves on unchanged input is the reason auditability is worth paying for.

Sizing a step to its checker rather than to the material is old. Descartes' second precept of method was "to divide each of the difficulties under examination into as many parts as possible, and as might be necessary for its adequate solution" (Discourse on the Method, Part II, verbatim, Veitch's translation). What he claimed for such division was the reliability of long chains: "The long chains of simple and easy reasonings by means of which geometers are accustomed to reach the conclusions of their most difficult demonstrations" (Part II, verbatim) had persuaded him that anything within reach of knowledge could be reached that way. The parts are sized to the solution, and the chain is trusted because each link is simple enough to be evident.

Lamport makes the checker explicit. Asked how much detail a proof step needs, he answers that when the proof is written to convince someone else "the answer depends on the sophistication of the reader" (How to Write a 21st Century Proof, section 3, verbatim). Written for oneself, the test is different: "if the truth of a statement is not completely obvious, or if you suspect that there may be just the slightest possibility that it is not correct, then more detail is needed" (section 3, verbatim). That is step size stated relative to whoever checks the step. His assessment of what the structure delivers is measured: "Structured proofs make it possible, not inevitable" (section 5, verbatim). The same modesty applies here — a bounded check does not catch every error, but an unbounded one cannot be redone. Both precedents concern human checkers reading mathematical arguments, so the transfer to agent-reviewed notes is an analogy about checker-relative sizing, not evidence that any particular N carries over. What the knowledge-base case adds is a validator that turns the commonest kind of step — this passage says this — into a check that costs the reviewer nothing.

Scope

  • Stated for Commonplace's grounding check. The relativity claim is general; the unit — the unquoted source — is derived from that check's cost structure and would have to be re-derived for any other check.
  • Depends on a system where verbatim quotations are mechanically validated against their cited source. Without that validator, quoting moves work around rather than discharging it.
  • Does not decide N, a validator rule that would enforce it, or the fate of the review-side budget currently in force. Those are operational decisions with their own evidence requirements.
  • The corpus counts and the paired experiment are bounded local reports about one knowledge base at one time; they show the two bounds are independent here, not how often they diverge elsewhere.

Open Questions

  • Is the check that a citing note represents a linked note's claim faithfully itself well bounded, or is it where the same problem reappears in smaller form? One of the four findings was a broader mechanism attributed to a link that disclaims it, which no count would catch.
  • Does the artifact-side bound have a cost the review-side one avoids — claims that are genuinely one inference over many sources and cannot be quoted into bounded passages, which would be pushed toward the snapshot route or toward fragmentation?

Relevant Notes: