Addressability grain, not compression ratio, sets a matched selective-read floor

Type: kb/types/note.md · Tags: document-system, context-engineering

For a known question, compare the material each access path must load to retrieve one discriminating, answer-bearing unit. Measure both units in one declared read-volume currency, such as bytes under one encoding or tokens under one tokenizer. The smaller addressed unit sets the matched selective-read floor. A summary lowers that floor only when its path loads less material than the source path.

The whole-artifact compression ratio alone cannot determine this comparison unless each whole artifact is itself the selected unit. Otherwise, the selected unit can be much smaller than its containing artifact. A whole artifact's total size equals that path's grain only when the whole artifact is its smallest addressable unit.

This claim applies to recoverable content: material that a declared source set can regenerate at the required reliability. It isolates raw read volume only. Answer sufficiency, semantic reconstruction, freshness, retrieval overhead, and maintenance stay in a separate cost ledger.

A matched selective read compares one addressed unit on each side

An artifact's addressability grain is the smallest material a reader can retrieve through the access path in use without loading the rest of the artifact, together with that material's volume in the comparison currency. Here finer means lower measured read volume, not greater topical or structural specificity. Grain depends on the exposed search keys, the reader's question, and the retrieval tool. It is not a fixed property of prose or code.

A matched comparison holds the question, target information, required reliability, and read-volume currency constant. On each side, the candidate answer must be explicit in one unit. A known key must discriminate that unit, and the access path must retrieve it without loading other units. Rule-based context selection needs a pre-existing signal explains the precondition: without such a key, this one-unit comparison does not apply.

These selected units define the raw retrieval floors. A coarser summary can still win overall when it consolidates several source units, disambiguates many hits, supplies unknown vocabulary, or avoids interpretation and verification work. Those wins come from fan-out reduction, discovery help, or cached transformation, not from the two unit volumes alone.

In the Commonplace paths below, source symbols and Markdown metadata or headings provide the keys. Other naming schemes and tools can make either side finer or coarser. The conditional relation stays the same, but the measured units change.

The rule explains two Commonplace paths

Commonplace, the system this note is written inside, provides a candidate helping path and one dated hurting measurement. The first shows the conditions under which structured descriptions lower the floor; it is not a benchmark. The second records the selected units in a historical repository snapshot.

Can help: structured frontmatter descriptions over a note collection. For a candidate whose path and title do not settle a routing question, a scoped search can retrieve its one-line description. When that line is sufficient for the routing decision and smaller than the body material the reader would otherwise need, the description lowers the matched floor. A collection-wide search can fan out across many candidates, multiplying those units rather than turning the whole task into one unit. Commonplace's move from whole generated inventories to scoped path-plus-description selection is recorded in ADR 025.

Measured hurt: a per-module prose reference over a code package. In the 2026-08-23 Commonplace snapshot recorded by commit 3c25fc66, the lib-modules.md section for type_resolver was measured at 3,657 bytes. A symbol search for validate_type_eligibility selected 1,530 bytes of source. For that question, the source path therefore had the lower byte-volume floor, even though the reference compressed the package as a whole. This is a snapshot-bound observation, not a maintained benchmark. It also illustrates the independent sufficiency question: if the approximate reference answer cannot license a reliability-compliant stop, the source remains in the path.

The same one-unit rule can therefore run in opposite directions. A sufficient, smaller description exposes a lower per-candidate routing unit. The historical prose reference exposed a larger symbol-answer unit. Because the paths involve different tasks and other differences, they illustrate the two orderings without estimating a general effect or isolating read volume as the only cause.

Consequences

Specificity helps only until each side reaches its floor. A more precise key can keep shrinking the material selected from the finer side after the coarser side has reached its minimum unit. Once both sides are at their floors, more specificity changes neither cost. Success on vague questions therefore does not show that a coarse layer will stay competitive on precise ones.

Distributed recovery needs an aggregate comparison. A cross-module invariant may be recoverable only by combining several source locations. In that case, the source-side cost is an aggregate reconstruction cost, not one addressed-unit volume, so this test does not decide it.

Recovery failure sets a relative boundary. If a declared source set cannot regenerate a decision, intent, or constraint, that content lies outside a cache comparison relative to that source set. But attempted recovery identifies informational gaps, not provenance or authority. Failed system-only recovery marks a candidate gap; it does not prove global uniqueness, historical derivation, or current governing force.

Operational test. Fix the question, required reliability, and read-volume currency. Ask whether the declared source set can regenerate the candidate answer and whether either path must combine several locations. Then identify one discriminating unit on each path and compare their volumes. Test sufficiency and the rest of the ledger separately. A smaller summary unit is necessary to lower this matched raw read-volume floor, but it is not enough to establish positive net value.

Scope

This claim concerns a matched selective question. Orientation across a whole subject is not selection of one answer-bearing unit. A question that requires synthesis across several units is still selective, but it lies outside the one-unit comparison. Two context boundaries govern collection operations develops the collection-scale regime. Both cases belong in the broader cache-value ledger discussed in whether equivalent recompute is worth avoiding.

A missing or non-discriminating key also blocks the comparison. One-letter names, metaprogramming, generated symbols, repeated prose terms, or unfamiliar vocabulary can force discovery or multi-hit inspection. Named paragraphs, anchors, and precise snippets can make prose finer-grained. Such changes alter or prevent the measurement; they do not refute the conditional relation.

Open Questions

  • What aggregate measure should replace grain when one answer requires several addressed units, or when one query returns several plausible hits?
  • What is the cheapest routing signal for a reader that knows the task vocabulary but not the source's discriminating key?

Relevant Notes: