Enhanced Interaction

A doctoral student reads a systems paper that cites, in passing, “the well-known impossibility result” — no proof, no reference, just the assertion. She opens a new tab. She searches the phrase. She skims three PDFs, finds the actual theorem in the second, checks its assumptions against the citing paper’s setting, and discovers the assumptions do not hold here. Forty minutes gone. The paper is still open where she left it, her mental stack of its argument half-evaporated. She will spend ten more minutes reconstructing where she was.

Nothing about those forty minutes was intellectually necessary. Every step — identify the missing reference, retrieve candidates, locate the theorem, compare assumptions — was mechanical support work the reading environment could have done alongside her. She never left the paper because the paper demanded it. She left because the medium is inert: it displays the source and abandons her at exactly the moments she needs the surrounding information world most.

This chapter earns the machinery that ends that abandonment — and it draws one boundary before anything else, because the rest of the book depends on it:

Enrichment is not interruption.

Identifying and retrieving useful surrounding information while someone reads is one problem. Deciding when that information deserves to break into attention — interrupt, notify, badge, push — is a different problem, owned by Chapter 19’s attention policy. This chapter’s delivery surface is deliberately simple: AVAILABLE IN SIDECAR. Everything the system finds waits beside the source, inspectable at the reader’s initiative, never seizing the floor. A sidecar that interrupts is a notification system wearing a sidecar’s clothes, and it will be scored as a timing-policy failure no matter how good its retrieval is.

The earned claim, stated narrowly:

A source-linked semantic record can be used to proactively retrieve potentially useful explanatory, evidential, and connective information without requiring the reader to formulate every search manually.

Not yet: AI knows which related information matters. Usefulness ranking is Chapter 8’s question. This chapter gets candidates into the sidecar; the next chapter decides what kind of relationship makes them worth attention.

The pipeline: retrieval and usefulness are separate stages

The architecture keeps two things apart that most “related content” features fuse:

semantic record
      ↓
candidate query/probe
      ↓
retrieval
      ↓
candidate set
      ↓
relevance/usefulness decision
      ↓
sidecar

Candidate queries are generated from the record: claims become evidence probes, concepts become definition and background probes, relations become prior-work probes, open qualifiers become assumption-check probes. Retrieval runs those probes against paper stores, documentation, code, and prior reading — ordinary retrieval machinery, deliberately unglamorous. The candidate set is everything returned. And the usefulness decision selects what appears in the sidecar.

The separation is load-bearing for two reasons. First, it gives Chapter 8 a clean insertion point: this chapter’s usefulness decision can be topical-relevance ranking, and Chapter 8 then demonstrates topical relevance is too coarse — what matters is how the candidate relates (supports, contradicts, extends, duplicates). If retrieval and usefulness were fused, that improvement would have nowhere to attach. Second, it makes failures diagnosable: an empty sidecar can mean bad probes (record too thin), bad retrieval (store lacks the item), or bad selection (good candidates rejected). Each points at a different fix. EXP-07 below exploits this by comparing query-only assistance against automatic topical sidecars against record-driven sidecars — isolating what the record adds over the query, and what selection still lacks.

Proactive conversational AI, surveyed comprehensively by Deng and colleagues (2025; ACM, DOI 10.1145/3715097), supplies the interaction lineage: systems that anticipate needs from context rather than waiting for explicit requests, spanning dialogue, recommendation, and task assistance. The survey’s value here is framing plus caution — proactivity’s benefits arrive with initiative costs, and the literature’s evaluations concentrate on engagement and task success rather than the distraction ledger this book insists on. The just-in-time-objectives work (Lam et al., CHI 2026, DOI 10.1145/3772318.3790713 — cited as a lead pending full read) offers the scoping idea this chapter borrows: interpret current activity as an implicit objective that focuses retrieval, rather than treating the whole document as one bag of keywords. Both are used for possibility and design vocabulary, not for usefulness claims the book has not yet earned.

Three kinds of enrichment, failing differently

“Related information” is three different jobs wearing one label, and the chapter refuses to merge them — because each fails in its own way, and the failure pattern is what creates pressure for Chapter 8:

EXPLANATORY
What helps me understand this source?
definitions, background, tutorials, worked examples
fails by redundancy: explains what the reader already knows,
  or restates the source in easier words that add nothing

EVIDENTIAL
What helps me verify or contextualise this claim?
supporting studies, contradictory results, datasets, methods detail
fails by weakness: retrieves low-quality, inapplicable, or
  misleading evidence — confident-looking but wrong

CONNECTIVE
What links this to other work, code, papers, prior material?
implementations, follow-ups, prior art, the author's earlier system
fails by vacuous similarity: topically close but useless —
  the same words, none of the leverage

A sidecar that mixes the three cannot be evaluated — explanatory redundancy looks like evidential weakness in aggregate metrics — so EXP-07 scores each category separately, with category-appropriate failure accounting. And the third failure is the chapter’s parting gift to the next: the sidecar finds something highly similar that contributes nothing new. Topical similarity without relationship awareness produces exactly this — the paper that shares vocabulary with the source while bearing on none of its claims. Chapter 8 opens on that corpse.

The fact-checking branch: bounded subtype, not a chapter takeover

Evidential enrichment shades into fact checking, and the chapter contains it with a deliberately small contract. Automated fact checking is a staged pipeline in the literature — claim detection, evidence retrieval, verdict prediction, justification (Guo and colleagues’ TACL survey; the FEVER-style task family) — and this book uses a trimmed four-step form as one sidecar mode among three:

checkable claim
  → retrieve evidence
  → show source/context
  → evidence status

Three containments keep it honest. First, checkable claims only: verifiable assertions with retrievable evidence — not opinions, predictions, values, or matters under legitimate expert dispute. Second, the output is evidence status with sources shown, never a bare true/false verdict: supported, contradicted, mixed, or insufficient evidence, always with the retrieved context visible and the source one click away. Third, for contested or political material, the sidecar reports and contextualises; it does not infer deception, motive, or intent from contradiction. A contradicted claim is a claim with contrary evidence attached, not a liar identified. Broader epistemic policy — what to do about persistent falsehood, how to weigh sources against each other — lives elsewhere in the book’s ethics, not in this pipeline. The sidecar brings evidence to the reader’s attention. Judgement stays with the reader.

What the sidecar must not do

A short list, because proactivity’s failure modes are well documented and this chapter declines all of them in advance:

• It must not interrupt. Sidebar availability only; timing belongs to Chapter 19.
• It must not present retrieved items as verified. Retrieval is not verification (the Embeddings-book distinction, reused) — candidates arrive labelled as candidates, with provenance, until evidential machinery promotes them.
• It must not infer the reader’s goal beyond the current activity. The probes derive from the record of this source, not from a model of the person — personal conditioning arrives in Part II and Part III with its own consent machinery.
• It must not hide its misses. Manual searches the reader still performed, items opened versus ignored, corrections and dismissals — all logged, because the sidecar’s usefulness is measured in searches avoided and discoveries used, not in panels rendered.

What this chapter earned

Reading need not mean abandonment: record-derived probes can fill a sidecar with explanatory, evidential, and connective candidates without manual searching, delivered as available rather than interruptive, with retrieval and usefulness as separable stages and fact checking as a bounded evidential mode. What the chapter did not earn — which candidates actually matter, and why — is visible in its failure patterns, especially the similar-but-useless connective item. Topical relevance got us candidates. It cannot tell us what kind of relationship makes one worth attention.

The sidecar finds candidates, but similarity alone does not explain why a candidate matters.

References

• Deng, Y. et al. (2025). Proactive Conversational AI: A Comprehensive Survey of Advancements and Opportunities. ACM. DOI 10.1145/3715097. Used for proactivity framing + initiative-cost caution; no usefulness claims taken.
• Lam et al. (CHI 2026), Just-In-Time Objectives, DOI 10.1145/3772318.3790713. Cited as lead (full read pending) for activity-as-implicit-objective scoping.
• Krisam et al. (2025), FOMO-to-Focus agentic topic alerts. Cited as lead (details pending) for interruption-cost context; timing claims deferred to Ch 19.
• Guo, Z. et al., A Survey on Automated Fact-Checking (TACL). Claim detection → evidence retrieval → verdict → justification pipeline. Used for the bounded fact-checking subtype structure.
• FEVER-style verification task family. Referenced as task context for claim-level evidence retrieval; no benchmark numbers taken.

Proposed experiment EXP-07: sidecar assistance across four conditions

Status: PROPOSED. Task: extended reading/implementation sessions over technical papers with planted needs (missing reference, checkable quantitative claim, available implementation, contradictory follow-up). Conditions: A no sidecar; B user-query-only assistance (search box, no proactivity); C automatic topically-relevant sidecar (whole-document probes, relevance-ranked); D semantic-record-driven sidecar (claim/concept/relation probes, category-separated). Measurements per enrichment category: useful items surfaced (human-adjudicated against planted needs), irrelevant items, manual searches avoided, task completion and time, duplicate information, source diversity, open/use rate of surfaced items — and distraction scored separately from utility (attention-switch counts, self-reported interruption, resumption lag). Predeclared negative case: a highly similar but non-contributing connective item must be detectable in scoring — conditions that surface it while missing the contradictory follow-up demonstrate the topical-relevance ceiling and open Chapter 8. Failure criteria: D ≈ C on useful items (record adds nothing over topical); D wins items but loses net task time to distraction (enrichment without attention discipline); B ≈ D on completion (proactivity unnecessary for these tasks). Artifacts expected: paper/need packs, probe logs, sidecar transcripts, per-category scores. What a positive result would not justify: that the system knows which related information matters — usefulness ranking by relationship type belongs to EXP-08.