Rendered from docs/decisions/0020-where-scent-lives.md in the Headwater corpus. Every document on this half of the site is typed by the taxonomy the descriptor names: corpus.json.

Q20 — Where scent lives

Context

One evaluation settles this with Q8. The standing half of the open-question entry asks how anything grades a cue, and the grader is a probe.

The premise of the open half is false, and foraging theory says so. The entry concluded that spec 5 grades distinctiveness by a comparison of siblings in one place, and that edge cues have no such place. The computational foraging models score a link decision as a utility over the links available at the reader's current position (HW-EVAL-adjacent-work §R.1). Scent is never absolute. It is a comparison over the options at the point of decision.

So one rule grades both placements. Grade a cue against the alternatives that it competes with at the moment a reader reads it. For a summary that set is the documents a reader is choosing between. The shelf is the static proxy for it, and the pointer list is the real thing. For a cue on an edge it is the other outbound cues of the document in hand, which is the choice that a traversal presents. The rule corrects the node side too, because spec 5 never named the set that a summary is compared against.

Decision

An optional source-owned cue on a relation instance, with the summary still required and absence falling back to the target's summary. Q4 settled where it lives and who writes it, and spec 2 already declares it on cites.

Consequences

Three static checks follow, and no new machinery. Non-restatement is Edge-scoped, which spec 12 already defines as one relation instance and both endpoints. Distinctiveness over cues is Document-scoped. The length band is trivial. All three stay advisory permanently under the fixability bar, because the remedy for a weak cue is a rewrite. A corpus that declares no cue has no instances and pays nothing.

Two behavioral measures come from transcripts that already exist. Traversal precision and traversal abandonment are the edge analogs of routing precision and abandonment. A traversal is visible in a tool-call transcript, as opening one document and then opening a target that it declares an edge to.

Four serving rules, and three of them follow from existing rulings. related and explain serve the cue and routing never does, because a routing result has no referring edge. A cue states the warrant of an asserted target, as a pointer does. The confidence gate does not reach a cue, because an author writes it rather than the engine scoring it. Silence is also not available to a reader who can already see the edge. A withheld target takes its cues with it, because the export filter is default-deny over classes.

The cue's counterfactual is the cheapest in the specification. Absence falls back to the target's summary, so the ablation is a switch in the serving layer rather than a second corpus. The leaning said to grade the cue only where a probe records a failed traversal. That is close and one step short, because a failure in the present arm alone has no baseline to compare against.

One consequence arrives from outside, and it is unchanged. A cue is an instance attribute, so an edge that carries one reifies in any RDF export. It appears in that emitter's declared loss set (Q6). The native graph export carries a cue with no loss.