Knowledge graphs · Agentic retrieval

Inbound Edges Are a Different Question, Not More Links

Outbound edges answer where you might go from here. Inbound edges answer what later or larger thing found this page relevant — consequence, use and later interpretation, not more of the same graph.

Scott Farrell · LeverageAI · July 2026

Your graph already has plenty of links. Someone proposes reverse traversal — walk not only where this page points, but who points here — and the room splits. Half the table hears “more edges,” which sounds like padding. Half hears “PageRank, but for our wiki,” which sounds like a ranking project nobody asked for. Both miss the point. The reverse direction is not a volume upgrade on the same relation. It is a different question, asked of the same graph, and it returns a different class of answer.

That is the whole thesis of this piece. Outbound edges answer: given this page, where might I go? Inbound edges answer: what larger or more peculiar thing considers this page relevant? Outbound supplies antecedent and association. Inbound supplies consequence, use and later interpretation. Once you hear the two questions as distinct, reverse stops being “more of the same plumbing” and becomes a design axis of its own.

Direction is not symmetry. The two walks answer different questions, so they carry different information.

What the parent piece measured — and what it left open

This article is an extender, not a rematch. Next door, in The Novelty-Preserving Carve-Out, a dated replay of stored walk traffic made a hard economic argument: most “enrichment” in a bloom payload was already visible on the outbound edge lines the model could already read. In that replay, about 99.7% of convergence nominations were already present in the visible neighbourhood; the protected residual that actually bought novelty was small — and a disproportionate share of that residual was the uncommon inbound-only relationship. The resulting redesign cut convergence entries from 6,472 to 987 and payload from 4.75 MB to 0.33 MB while improving measured precision and cost per eventual take. Those numbers belong to that piece. They are cited here once so you know the residual class was real under a context-budget audit — not so this article can re-own the measurement argument.

What #184 proved is that a payload should spend context on novelty, not on mechanically repeating salience. What it did not prove — and what this piece owns — is what the reverse direction means. Economics without semantics do not generalise. You can protect inbound-only entries as “novel residual” and still treat them as an accident of encoding. The design advice only travels when you can say, cleanly, which question reverse answers that forward cannot.

One more boundary before the work. Bloom originally allocated attention according to graph structure: convergence counts, hotspot resolution, richer packages where the neighbourhood already implicated a node. That architecture still stands as a way of spending resolution where structure warrants it. Telemetry then refined which structure deserves extra payload. This article continues that refinement on the meaning axis: not denser bloom, not another Top-K, but the orientation of the edge itself.

Two questions, not two volumes

Imagine you are standing on a concept page mid-walk. The page is competent. It names a claim, lists a few typed neighbours, points at a source chapter. From here the agent can ask the natural forward question:

Given this page, where might I go?

That is outbound. It returns antecedents the page already knows: the framework it develops, the source it compresses, the related concept it contrasts with, the project it originated in. The walk extends along associations the author could write because those targets existed — or were later edited in — when the page was shaped.

Now invert the question without inventing new relation kinds:

What larger or more peculiar thing considers this page relevant?

That is inbound. The answer is not “the same list, reversed for fun.” It is a different retrieval job. Scientific literature already treats the distinction as ordinary: references are what a paper points at; cited by is who used the paper later. No serious researcher confuses the two panels of a paper’s landing page. Classic link analysis, for all its sins as a ranking cult, at least encoded the same asymmetry: importance flows from what points at you, not only from what you point at.¹ You do not need PageRank scores to borrow that orientation. You only need to stop treating directed edges as undirected for design purposes.

The failure mode is collapsing those rows into one line: “links.” Once you do that, every reverse feature looks like a volume knob. Turn it up and you get denser payloads. Turn it down and you get thrift. Neither knob asks whether you just spent budget restating what the model could already see from the outbound neighbourhood — or whether you just omitted the only orientation that could correct a walk that landed on yesterday’s page as if it were still the genba.

Why reverse is not free inference

A common objection: if the graph stores directed edges at all, reverse is a free view. Why build anything? Why protect reverse-only residual in a payload?

Because storage reverse and payload reverse are different, and because authorship-time is asymmetric even when the store is bidirectional.

First, payload. A walk response does not hand the model the entire graph. It hands a neighbourhood — search hits, one-hop gists, convergence, bloom packages, whatever your stack ships. The model is excellent at noticing that several visible edges point at the same hub. It is not entitled to invent a true reverse-only neighbour that was never returned. “Invisible reverse relationship from outside the returned neighbourhood” is exactly the class of thing a novelty-preserving carve-out is trying not to discard. That is the parent piece’s residual class, restated only as design context.

Second, authorship-time. A page can write edges to what exists when it is written, and to what a later edit deliberately adds. It cannot spontaneously know, at write time, which distant project will operationalise it next year, which framework will elevate a minor concept into a load-bearing dependency, which later synthesis will absorb it, or which capstone will supersede it. Those facts often appear as edges from newer nodes to the older page. From the older page’s local view, they are reverse. From the graph’s global view, they are simply later authorship. The orientation matters for the walker who is currently standing on the old page and does not yet know the future exists.

Edges in a compiled layer are already not a cache of what sources said; they are added structure synthesised across sources — a compile, not a re-derivation. Direction is part of that added structure. Treating it as cosmetic is how you smuggle an undirected mental model into a directed graph.

Five things a reverse edge can reveal

Here is the core substance of the piece. Each item is a different job reverse can do for a walker. None is “another related link.” Develop the habit of classifying reverse hits into these jobs; if a reverse edge does not land in one of them, ask whether it is true reverse information or just a bidirectional twin of an outbound you already returned.

1. A later synthesis that absorbed the current idea

What it looks like. You are on an early essay that introduced a sharp observation — say, that a compiled claims-and-edges layer is a transformation of sources rather than a cache of them. The essay is still true. It is no longer the densest place the idea lives. Months later a capstone chapter absorbed the observation into a wider maintenance doctrine: chronological stacking, janitor jobs, lint, contradiction-as-edge, the numbers-out rule. The capstone points back at the early essay as origin or develops-from. The early essay, unless someone edited it, does not point forward at the capstone that now carries the load-bearing package.

What reverse surfaces. Standing on the early essay, outbound neighbours look like the world that existed when the idea was young. Inbound surfaces the later synthesis: “this idea was absorbed here.” The walker that only follows outbound may treat the origin page as current doctrine. The walker that can see reverse gets offered the denser basin and can still return to the origin for history.

Why this is not “more links.” The relation kind might be ordinary — develops, compresses, originated-in. The direction from the old page is the only way the absorption is visible without already knowing the later title. If your payload only ever shows outbound, absorption is a secret kept by people who already know the canon map.

A practical check: if removing reverse would make “has this idea been synthesised more tightly elsewhere?” unanswerable from the current node, you are looking at reveal type 1, not padding.

2. A framework that applies a seemingly minor concept

What it looks like. A small concept page exists because someone needed a name for a sharp constraint — “access without citizenship,” “numbers out,” “relational arity boundary.” On its own page the concept looks modest: a definition, a couple of edges, a warning about a failure mode. Elsewhere, a framework page treats that concept as a load-bearing dependency. The framework is where the concept becomes operational design: reverse indexes for quote exhibits without granting them graph citizenship; maintenance rules that refuse to store raw figures in the graph; pair-space constraints that force a typed edge instead of a unary tag.

What reverse surfaces. From the minor concept, outbound may only show sibling concepts and a source pointer. Inbound shows the frameworks that apply it — the places where the small idea became a rule with consequences. That is “use,” not “association.”

Why this matters for walks. Agents often land on the concept because the user’s words matched the definition. The decision-relevant material may live on the framework that applied the concept under pressure. Without reverse, the walk can stop at a tidy definition and never learn that the house already spent a book operationalising it. With reverse, the peculiar larger thing that considers this concept relevant becomes a candidate next hop — not because it is popular, but because it is where the concept earns its keep.

A practical check: ask whether the reverse neighbour is “related in spirit” or “depends on this concept to function.” Only the second is type 2.

3. A distant project where the idea became operational

What it looks like. Doctrine pages and project pages often live in different regions of a graph. A concept about selective resolution at structural hotspots may be discussed abstractly in an investigation architecture, then implemented as behaviour in a specific tooling surface — a wiki search package that returns layered evidence, convergence counts, and bloom only where the result subgraph already implicated a node. The project page points at the doctrine. The doctrine page may never list every project that later embodied it, especially once the project count grows.

What reverse surfaces. From the doctrine, reverse answers “where did this leave the whiteboard?” That is operational consequence: not another essay restating the principle, but a place where probes, fusion and attention budgets were actually wired. Walkers doing implementation questions need that hop. Walkers doing pure theory may not. The graph should still make the hop available, because the model cannot reliably invent the project id from a doctrine abstract.

Why “distant” matters. Distance here is graph distance and institutional distance, not word count. The project may not share vocabulary with the doctrine page. Outbound similarity from the doctrine may prefer other doctrine. Reverse is how peculiar operationalisations re-enter the walk when the agent is standing in the wrong room for the user’s real intent.

A practical check: if the reverse neighbour is a project, repo, or deployment page whose job is embodiment rather than restatement, tag type 3. Do not demote it because it “doesn’t look like a knowledge page.”

4. An unexpected cross-domain relationship

What it looks like. The same principle is restated or applied in software architecture language, then in strategy language, then in publishing operations, then in governance. A similarity model may or may not notice. The surrounding neighbourhoods disagree even when the core proposition is recognisable. Semantic refraction names a related discipline from the publishing side: usable meaning rises when closed units form joins the whole was too coarse to hold — averages destroy edges; addresses create them. Cross-domain reverse is that idea facing the walker: you landed in software, and something in governance points back because the constraint is the same constraint under a different lens.

What reverse surfaces. Not the restatement you already linked outbound as applies, but the unexpected domain page that treated this node as evidence or dependency without living in your default path. The peculiarity is the signal. If every reverse neighbour is the bidirectional twin of an outbound you already shipped, you have not found type 4; you have found dense bidirectionality (see the rarity paradox below).

Why this is easy to discard. Cross-domain reverse hits look noisy on a dashboard optimised for “top related.” They are exactly the hits that re-anchor a walk that was about to overfit one domain’s vocabulary. For route-resilient substrates — many paths, same genba — those odd entrances are not decoration. They are how a question asked in the wrong dialect still finds the territory.

A practical check: if the reverse neighbour’s neighbourhood would never appear in a one-hop outbound expansion from here, and the join is still load-bearing for some real question family, treat it as type 4 and protect it in the payload even when it is rare.

5. The newer canonical consolidation the older page cannot point at

What it looks like. An original page states a claim. Later work corrects a boundary, supersedes a recommendation, or consolidates several near-claims into one current authority. The later page carries a supersedes or updates edge to the earlier one. The earlier page, unless someone performed a careful back-edit, still reads as if it were current. It does not know it has been superseded. Only the reverse direction — or a dedicated as-at walk down a supersession chain — carries the correction to a walker standing on the old node.

This is the case that makes the “same edge read twice” objection look careless. The stored edge may be one directed arc from new to old. From the new page, outbound answers “what did I replace?” From the old page, reverse answers “am I still current?” Those are different questions with different decision consequences. Chronological stacking and typed supersession already exist as maintenance doctrine for keeping disagreement and succession navigable rather than averaged away. Direction is how that doctrine reaches a walker who did not enter through the new page.

What reverse surfaces. Authority correction. Not “related reading,” but “stop treating this as present doctrine.” If your agent answers a present-tense policy question from a superseded page because reverse was omitted from the package, you did not save tokens. You bought a fluent wrongness with a tidy payload.

A practical check: for any page that has been superseded even once, ask whether a present-doctrine walker landing on the old page is offered the new authority without already knowing its title. If not, reverse (or an equivalent as-at mechanism) is load-bearing, not optional polish.

The rarity paradox

Reverse-only relationships are uncommon in a mature, carefully edged graph for a boring reason: most relations are already represented in both directions, or the payload already showed the outbound twin. Dense bidirectionality is a sign of health, not of wasted engineering. Maintenance cultures already treat missing inbound as a defect — orphan pages with no inbound links show up in lint checklists for exactly that reason.²

Here is the non-obvious half. Precisely because reverse-only residuals are rare, a reverse relation that is genuinely absent from the visible result payload carries unusually high information value. Rarity and value are the same fact here, not a trade-off. Teams make the opposite mistake twice:

• Mistake A: “Reverse is rare, so don’t build it.” You just discarded the residual class that survived an overlap audit for being non-redundant.
• Mistake B: “Reverse is valuable, so flood the payload with every reverse edge.” You just recreated richness-as-proxy — spending context on bidirectional twins the model can already see from outbound lines.

The design move that respects the paradox is selective: keep reverse as a first-class orientation, then protect the reverse-only residual the way the novelty-preserving carve-out protects novelty in general. Dense bidirectionality explains why the residual is small. The five reveals explain why the residual is not optional.

Rarity is why the reverse residual is informative — not why you can ignore it.

Kind is not direction

The canon already invests heavily in relation kind. Typed edges such as develops, applies, updates, supersedes, compresses and originated-in tell the walker why two pages are connected. That inventory is load-bearing. It is also incomplete if you treat kind as a substitute for direction.

Kind answers: what is the nature of this link? Direction answers: which end am I standing on, and therefore which question am I asking? The same supersedes arc is “what did I replace?” from the new page and “what replaced me?” from the old. The same applies arc is “what concept do I operationalise?” from the framework and “who operationalised me?” from the concept. Kind without direction is a labelled undirected graph wearing directed syntax.

This is also where this piece parts company with a neighbouring doctrine without fighting it. Wiki redundancy as error-correcting structure defends multiple semantic entrances, progressive development and cross-domain restatement against naive merge — a corpus-shape argument. Those entrances can be reached by many paths. Reverse is not the same claim. You can have multi-entrance redundancy and still only ever expand outbound. You can ship reverse and still foolishly merge type-1 entrances. Keep both rules; do not swap them.

Quote Satellite already owns reverse for a different job

If you have lived in this canon for a while, reverse index may already sound familiar. It should. Quote Satellite — developed in Quotes Without Canonical Authority — uses a deterministic reverse index so exhibits can be retrieved from a canonical concept without materialising dozens of quotes as primary graph neighbours. The operating rule there is crisp: traversal both ways, authority one way. Quotes remain type-tagged exhibits; they do not gain citizenship over truth.

That is reverse-as-access-structure for quote-carding and exhibit retrieval. This article is reverse-as-epistemic-orientation for graph design and walk payloads. The mechanism family overlaps; the purpose does not. Re-deriving Quote Satellite here would be theft from a sibling doctrine and a distraction from the five reveals. The only design transfer worth taking is the discipline: reverse can be first-class without granting every reverse neighbour equal authority. Access without citizenship is a pattern you can reuse for other non-canonical exhibits. It is not a substitute for asking what inbound means when both ends are already citizens of the graph.

Honesty: what was not measured

Load-bearing honesty: there is no dedicated graph-directionality study behind this argument. The novelty-preserving replay measured a context budget and an overlap residual. It did not measure edge direction as an independent experimental factor. The stabilising effect of reverse on pathing is felt and inferred from walk behaviour — “can make pathing feel more stable” is the honest register. It must not be upgraded to a measured result in slides, sales decks, or memory.

This honesty is not a disclaimer stapled on at the end. It is the reason this piece almost was not written. The natural editorial move was to fold the reverse-edge paragraph into the carve-out article and move on. That would have been thrifty and wrong. The meaning of direction is not a corollary of a byte-count chart. Separating the articles forces the forward contribution into the open: if we do not have a directionality study, we should specify the experiment that would settle it, tightly enough that a team can run it, rather than smuggle “feels more stable” into the house as if it were already scored.

What we do have, from the field transcript that produced both pieces:

• A co-design loop that proposed reverse neighbourhoods after looking at walk histories.
• A qualitative observation that pathing felt more stable and that peculiar still-relevant connections appeared.
• An economic residual class in which inbound-only entries were disproportionately non-redundant under an overlap audit (#184).

What we do not have:

• A controlled ablation that removes only inbound-only payload entries.
• Before/after scores for evidence invariance or answer invariance attributable to direction.
• Any percentage that quantifies “how much more stable.”

If your organisation cannot live with that gap, run the experiment below before you promote reverse from design doctrine to reliability claim. If you can live with provisional design advice, ship reverse protection as a novelty class and keep the stability language in the felt register until the numbers exist.

The experiment that would settle it

This section is the forward contribution that earns the article its own number. It is a protocol, not a results report. Do not read the steps as outcomes. Read them as a design you can execute on your own walk corpus.

The measurement instruments already exist next door. Route-Invariant Grounding splits groundedness into four quantities scored across a family of walks of the same question under deliberately varied wording, routing parameters and traversal order:

Path identity within one walk is the wrong quality score. Evidence invariance across many walks is the load-bearing property. Answer invariance is necessary and insufficient — answers can agree while floating free of evidence. Traversal efficiency is an engineering constraint, not the hero metric. Those distinctions are owned by #182; this experiment only borrows them as the scoring surface for direction.

1. Freeze a question family before you touch payloads. Choose six to twelve questions that already produce walks through regions where reverse-only residual has appeared at least once in historical logs. Include present-doctrine questions, origin questions, and at least two adversarial rewordings that avoid house vocabulary. Freeze the text. If you write the questions after seeing ablation results, you are not running an experiment; you are writing a story.

2. Define “inbound-only entry” operationally. For each model-facing package, mark an entry as inbound-only when (a) it is present because something outside the primary hit set points at a visible node, and (b) it is not already serialised as an outbound neighbour of any page the package already includes. Bidirectional twins of outbound edges do not count. If your instrumentation cannot make that cut, fix instrumentation first. Ablating “all reverse-looking strings” will trash signal and noise together.

3. Build two packages for every walk seed.

• Control (A): current production package, including protected reverse-only residual under whatever novelty-preserving rule you already ship.
• Ablation (B): identical package with inbound-only entries removed. Leave outbound, halo gists, and non-reverse novelty residual intact. Do not “compensate” by expanding Top-K outbound. That confounds the factor.

4. Replay a family, not a single walk. For each question, run multiple traversals under varied wording and routing parameters — the same spirit as route-invariant grounding’s family-of-walks design. Capture pages opened, claims cited, edges used, final answers, and whether source-descent still reached the genba chapters. Prefer external cognitive provenance over the model’s post-hoc story about what it “must have read.”

5. Score all four quantities on A and B. Do not collapse them into one green/red. Especially watch:

• Evidence invariance (B vs A): did load-bearing chapters or claims drop out when reverse-only residual disappeared? Type-5 supersession failures often show here first: the family still answers fluently from an older page because the new authority never entered the package.
• Answer invariance of conclusions and qualifications: two answers that share a headline but drop opposite caveats are not invariant. Prose similarity is a weak proxy.
• Path variance: rising path chaos is interesting only if evidence also moves. High path variance with stable evidence can still be health.
• Traversal efficiency: record it; do not let thrift win the argument if evidence invariance collapsed.

6. Map families onto the variance diagnostic table. #182’s four-row table is the decision surface: path varies / evidence stable; path and evidence vary / answer stable; evidence stable / answer varies; both vary. The dangerous cell for this ablation is “answer still looks fine while evidence set lost the reverse-supplied authority or operationalisation.” That is the family-level cousin of bad path + right answer — model prior or residual local prose wearing a search UI.

7. Stratify reverse-only hits by the five reveals before you interpret. Tag a sample of removed entries as absorption, application, embodiment, cross-domain, or supersession. If the ablation hurts only when type-5 entries were removed, your design conclusion is different from a world where type-3 operational projects were doing the work. An unlabelled residual will produce uninterpretable “sometimes reverse helps” folklore.

8. Report nulls and partials. A legitimate outcome is: reverse-only residual was economically novel under overlap audit and still did not move evidence invariance for this question family. That would demote the stability claim without demoting the novelty claim. Another legitimate outcome is: stability moved only for supersession-heavy regions. That is still a result — a scoped one. A protocol that can only celebrate reverse is not a test.

9. Do not smuggle Quote Satellite metrics into this study. Exhibit retrieval precision for quote-carding is a different purpose. Measure it in its own lane. This ablation is about citizen-graph walk packages and epistemic orientation, not about secondary indexes for immutable excerpts.

Instrumentation note: agent evaluation platforms already capture full trajectories — tool selection, arguments, retrieval paths, latency and cost — not only final strings.³ That is necessary capture. It is also how teams accidentally optimise path identity. This experiment inherits #182’s warning: do not treat a changed page sequence as failure if the genba still holds, and do not treat a fluent answer as success if the reverse-supplied authority never entered the evidence set.

Design consequences once the meaning is stated

If direction is an epistemic orientation, several design moves stop being optional taste.

Payload composition. Treat inbound-only residual as a protected class under the novelty-preserving rule, not as a second bloom that re-prints every reverse twin. Cap already-visible reverse the way you cap already-visible salience. Protect what the model cannot obtain from the outbound neighbourhood it was handed.

Authoring discipline. When you write a supersedes, updates, or “this framework depends on that concept” edge, assume walkers may be standing on either end. Prefer explicit kind labels over bare “related,” because reverse without kind is how peculiar becomes noise. Do not rely on future you to back-edit every old page with a forward pointer; reverse is the correction channel when back-edits fail.

Evaluation. Add direction ablation to the weekly battery once the four quantities are instrumented. Until then, keep stability language in the felt register. Promote reverse from “interesting idea the co-design loop found” to “reliability claim” only when step 8 above has something real to say.

Mental model for reviews. In design reviews, ban the phrase “just more links” as a dismissal of reverse. Replace it with: “Which of the five reveals does this reverse neighbour supply, and is it already visible outbound?” If neither answer is clear, you may be looking at noise. If a reveal is clear and outbound-invisible, you are looking at the residual the carve-out was built to protect.

Sibling routing, restated cleanly.

• #184 owns measurement of the context budget and the novelty residual.
• #182 owns the four quantities and the family-of-walks measurement doctrine.
• #183 owns which repetitions are load-bearing entrances versus free bytes.
• This piece owns what reverse means, the five reveals, the rarity paradox, and the direction ablation that would close the honesty gap.

What you should do on Monday

Take one recent walk package from production or from a replay harness. Highlight every reverse-looking entry. Split the highlights into two piles: bidirectional twins of outbound the model already had, and reverse-only residual. Tag the residual pile with the five reveals. You will learn faster from twenty tagged residual edges than from another abstract debate about whether reverse is “worth building.”

Then pick one question family that has at least once landed on an older page in a region with known supersession. Ask, without looking at the new title first: would today’s package tell the walker that newer authority exists? If the answer is no, reverse (or as-at) is not a research nice-to-have. It is a present-tense correctness problem.

Finally, schedule the ablation as a real experiment with a pre-registered question list. Score the four quantities. Report nulls. Keep “feels more stable” out of any sentence that wants a number until a number exists.

Outbound edges answer where you might go. Inbound edges answer what found you relevant later. That is a different question — not more links.

After this, you should be able to classify your own edges by the question each direction answers, name the five jobs reverse can do, and treat rarity as information rather than as an excuse. The measurement that would turn the felt stabilising effect into a reliability claim is specified. Until it runs, the honest house line is simple: reverse is a different epistemic orientation; protect the reverse-only residual; do not claim a study you did not perform.