# Enhanced Interaction
# Book: Language
# Chapter URL: https://programmer.ie/books/language/09-chapter/
# Prose fingerprint (sha256, normalized): aa284a0947227a1c16d85164065764ed6406263ba0b2e579ad0df88cc73bc0e9
# Snapshot generated at build time by Hugo. This is the complete rendered prose
# of the chapter; it is not a transcript of the source file.

A doctoral student reads a systems paper that cites, in passing, “the well-known impossibility result.” The citation gives no proof and no reference, just the assertion. She opens a new tab and searches the phrase. She skims three PDFs, finds the actual theorem in the second, and checks its assumptions against the citing paper’s setting. They do not hold there. Forty minutes have gone. The paper is still open where she left it, but the mental stack of its argument has half-evaporated, and she will spend ten more minutes working out where she was.
Some of those forty minutes may have been avoidable, but not all of the work was clerical. Finding candidate papers and locating a theorem can be delegated. Deciding whether the theorem’s assumptions apply may require exactly the student’s expertise. A reading environment could keep those operations alongside the paper, preserving her place and making the evidence easier to inspect. It could not responsibly turn a retrieved theorem into an applicable one by assertion. The source left her without the surrounding material she needed, and the interface offered no way to bring that material in without breaking her reading.
The previous chapters built the pieces that make another arrangement possible. There is an unchanged source; a semantic record, which is a fallible, source-linked interpretation of that source; and a primary rendering, the view the reader actually reads. This chapter adds an optional surface alongside the rendering. It is not a new authority downstream of the rendering. The sidecar may display claims derived from the source or candidates discovered elsewhere, but each item must remain independently accountable to where it came from. This chapter asks what may enter that surface and on what terms:
SOURCE ↓ SEMANTIC RECORD ↓ PRIMARY RENDERING ↓ OPTIONAL SIDECAR Once these pieces exist, what additional material should be allowed to appear beside a source while someone reads it? That is a narrower question than “how can AI help with reading?” The sidecar is not a second author and not a general assistant. It is an optional presentation surface for three kinds of material: source-linked material, external candidates and explanations the reader has requested. It may be empty. The key questions are: what was supplied, what was inferred, how each item is labelled, whether it deserves attention, and whether the reader can get back to the evidence.
Origin and interaction are different questions Three useful labels recur in a sidecar, but they are not three mutually exclusive origins. Derived and retrieved describe where evidence comes from. Interactive describes how the reader requests or changes an item. An interactive response may combine a source quotation, an external paper and a new model inference. Calling the whole response “interactive” cannot settle what supports any of its parts. The distinction is small in wording and consequential in implementation.
Derived. Re-expressed from the current source or its semantic record: a claim map, a list of assumptions explicitly stated in the paper, or a pointer to an earlier definition. The derivation is a proposal about what the source says. A model can omit a negation or infer an unstated assumption even when no external document was used. A derived item therefore needs source spans and a check of the interpretation. A label saying “from this document” is not enough.
Retrieved. Found in material outside the current source: the original theorem, a dataset, a later correction or a paper reporting a different result. Retrieval creates a candidate and a source trail, not a verified fact. The record should identify the external work, the version or edition when available, the exact passage retrieved and the retrieval time when freshness matters. A close match may concern a different theorem, or it may use assumptions that do not apply.
Interactive. Requested or revised by the reader while reading. Suppose she selects the purported theorem and asks, “Which assumptions would have to hold for this result to apply here?” The answer may quote the current paper, retrieve the original theorem and propose a comparison. Those are different epistemic components inside one interaction. A reader’s request authorises an attempt to answer; it does not give the answer authority, and it does not make the system’s unstated background knowledge into evidence.
Each label therefore points to a different boundary. The distinctions below are warnings, not a complete partition of all possible items:
derived ≠ externally supported retrieved ≠ verified interactive ≠ authoritative A derived item can misread its own source or introduce an unsupported interpretation, so its source pointer alone does not establish fidelity. A retrieved item may be authentic and still irrelevant to this claim, out of date or inapplicable to these assumptions. An interactive response may cite both kinds of material while adding deductions that neither source makes. The sidecar must expose these differences at the level of the consequential claim, not just as one badge for an entire card. None of these items is the original author speaking, unless the words are explicitly presented as an exact quotation.
Where it came from, always This leads to the chapter’s central rule, which should survive every change of interface: origin, interpretation and evidential standing must remain separable. A reader should be able to distinguish an exact quotation from a source-linked paraphrase, an external candidate from an assessed piece of evidence, and either of those from a model’s new inference.
The reader should always be able to tell what came from the source, what was derived from it, and what came from elsewhere.
The four visible layers below are disclosure categories, not four exclusive content types. A single interactive answer may need more than one: a quotation labelled SOURCE, an outside theorem labelled EXTERNAL EVIDENCE and a conditional comparison labelled MODEL GENERATION. The interface must keep that separation even when it composes them into one response.
SOURCE what the author wrote SEMANTIC DERIVATION what the system re-presented from it EXTERNAL EVIDENCE what was brought in from other material MODEL GENERATION what the model produced in response The sidecar must never silently merge into the source, and the failure is easy to picture. A retrieved summary adopts the typography of the paper and appears to be the author’s own argument. A conditional model answer is copied into the reader’s notes without its condition. A claim is reused in a later session without the source version or passage it came from. In each case an item changes standing as it crosses an interface boundary. A durable note or export therefore needs its origin and unresolved status attached. A coloured badge that disappears on copy is not enough.
This is how assistance can acquire unearned authority. A clear explanation may become more memorable than the original, and a reader may attribute the explanation’s conclusion to the author even when the source did not draw it. Better prose is not the problem; invisible provenance is. The remedy is not to make sidecar writing deliberately awkward. It is to make quotation, interpretation, external evidence and unresolved inference visibly and persistently different.
Proposed disclosure layers, not four exclusive content types. One interactive response may need a source quotation, an external theorem and a model inference distinguished within it. Open full-size diagram. What problem it solves, and what it costs The sidecar is intended to address a specific interruption in reading: leaving the source to locate supporting material, then reconstructing where the argument stood. It might reduce manual searching, protect the reader’s place and expose relevant evidence. These are hypothesised benefits, not outcomes established by the scripted example later in this chapter. Returning to the student’s search, success is harder to define than speed. She must notice that the theorem’s assumptions do not fit the citing paper’s setting. A fast but unwarranted “verified” label would be a failure, not a success.
The sidecar also carries costs. Each candidate beside the text competes for visual attention, asks to be interpreted and may send the reader down another trail. One relevant theorem can justify several candidate inspections. Ten attractive but inapplicable items can be worse than doing the search manually. Costs include switching, task resumption, verification effort, latency and the possibility of source confusion. Useful discoveries, incorrect suggestions, missed decisive evidence and intrusions should be counted separately. A single satisfaction score cannot establish that the gains outweighed those failures. Research on mixed-initiative interfaces has long treated uncertainty about users’ goals, and the timing and cost of interruptions, as design variables rather than problems solved by adding an intelligent agent (Horvitz, 1999).
Reader preference is worth measuring, but it is not a substitute for the reading outcome. A reader may welcome a convenient margin and still misattribute a quoted theorem or misunderstand its scope. Evaluation therefore needs the task result, the quality of source attribution and the cost of interruption alongside reported preference. The sidecar must earn its value against ordinary reading and manual search, not merely against an intentionally cumbersome workflow.
Three disciplines follow from these costs.
Available, not intrusive. The reader can reveal, dismiss or close the sidecar. Preparing candidates need not seize attention. Automatically opening a panel, flashing a warning or inserting unsolicited text into the paper is a separate intervention that needs its own justification. Horvitz’s mixed-initiative work treats direct invocation, uncertain intent and the cost of poorly timed assistance as design questions. This chapter deliberately tests reader-invoked access first. It does not claim that proactive interruption has been earned. That question belongs to the book’s later attention work.
Anchored in the source and current task, not a durable profile. The sidecar can use the passage the reader selected, the question they explicitly asked and sources admitted for that task. It should not silently build a persistent theory of the reader, borrow another project’s documents or promote a past inference into present authority. Availability, relevance and permission to influence this computation are separate decisions. The context and memory work elsewhere in the book examines those boundaries in detail. No account history is needed to test the mechanism proposed here.
Nothing is a legitimate output. A passage may already be sufficient, a retrieved theorem may be only a name match, or a check may be impossible with the available evidence. The sidecar can remain empty or report “unresolved” rather than manufacture a supporting item. Abstention is not, by itself, evidence of good selection. An evaluation must also count the decisive candidates the sidecar failed to surface.
What may join the communication Underneath these rules lies an admission decision. External material may be relevant enough to inspect without being reliable enough to cite as support. A model may be permitted to display a proposed interpretation without being entitled to declare the source corrected. The sidecar therefore needs bounded permission to show, annotate or withhold candidates. It should not silently treat every candidate as part of the paper’s argument. The following questions form a proposed admission checklist. The implementation has not validated them:
a candidate arrives ↓ is it derived, retrieved or generated? ↓ does it carry its origin and its status? ↓ is it labelled as a candidate, not a finding? ↓ may the reader dismiss it, check it and trace it? A retrieved item begins as a candidate. Before it counts as support for a claim, the system would need to check identity, the quoted passage, the relation to the specific claim and the assumptions under which that relation holds. A citation or a similarity score does not accomplish those checks. An interactive answer should separate source quotations, attributed external statements and new deductions, and say when the comparison is unresolved. These labels describe the evidence available. They are not a certificate of truth. The typed relationships and relevance tests needed for cross-source candidates are taken up next, rather than assumed here.
The reader should also be able to inspect what the system actually attempted: candidates offered, sources opened, items dismissed and corrections made. But a log of shown items does not reveal useful items that were never retrieved. To measure missed evidence, an experiment needs an independently constructed reference set. Recording behaviour also has costs. Instrumentation should be bounded to the task, avoid collecting unrelated browsing history, and not turn transient reading activity into a durable personal profile without a separate decision.
A boundary: explaining is not checking Some sidecar material addresses a claim’s truth or applicability. That is a different operation from making a source easier to read. Automated fact-checking research separates claim identification, evidence retrieval and checking. Success at one stage does not establish the next (Guo, Schlichtkrull and Vlachos, 2022). A status such as supported, contradicted, mixed or insufficient must therefore be scoped to a particular claim, an identified evidence set and the relevant conditions. “Contradicted by this passage” is not a general verdict that a paper is false. An apparent clash may disappear when populations, time periods or theorem assumptions are aligned. The reader can inspect the reasons and disagree. This chapter establishes the display and provenance obligations. It does not describe a fact-checker that has earned those judgements.
A proposed test: help without changing the source The opening theorem example suggests a test, but no reader experiment is reported here. Freeze a small set of papers with deliberately difficult citation gaps, together with the relevant external works. Independently adjudicate each task’s decisive assumptions, correct applicability judgement and acceptable supporting passages before testing an interface. Include negative cases in which the likely retrieved paper has a similar title but different assumptions, and cases where no supported answer is available.
Compare at least three conditions on matched tasks: the unchanged paper with ordinary manual search; the same paper with a conventional searchable reference panel; and the paper with a reader-invoked, source-labelled sidecar. Keep access to the same underlying evidence as comparable as possible. Record the time and number of source switches, and counterbalance task and condition order where people take multiple conditions. A static, authored evidence panel is a useful further control when the question is whether generation adds value beyond good information design.
Score correct applicability decisions and correctly cited decisive assumptions separately from total time, task resumption cost, unnecessary candidate inspection, false attribution to the primary paper and reader preference. Count the cases where the system confidently supplies the wrong theorem, fails to surface the decisive one or should have stayed silent. This design would test whether the sidecar improves a bounded reading task. It would not establish general comprehension, long-term learning or a useful policy for unsolicited interruption. Those remain different questions.
Running it The reference implementation (ernanhughes/project-language) includes npm run example:ch09. As described by the manuscript, it places derived passages, retrieved candidates and a model-labelled answer beside a constructed systems paper. It runs on a Pi host with a scripted stand-in for the model and makes no real model call. The example demonstrates a bounded interface contract under chosen inputs. It is not the reader comparison proposed above, and the fact that its candidate items carry the right labels does not establish that their contents are correct.
The scripted program keeps its three presented categories visibly distinct. It refuses a purported source quotation that does not occur in the source, and it flags an item carrying a forged layer. A derived item’s displayed quotation is checked for occurrence, but the system’s interpretation is labelled separately and is not checked for semantic fidelity. Through the model’s proposal tool, an answer remains model-generated even if the script tries to classify it as a derivation, and citations must name known items. The reader requests the margin, item count is capped, empty output is valid, and the event ledger keeps helpful and costly events in separate columns. These are meaningful mechanical controls. They do not implement the claim-level, cross-source validation proposed in this chapter.
The limits rule shows the distance between a field being present and a claim being justified. In the scripted run, “None.” and stub text are rejected, as is an uninformative limit pasted onto a second answer. The check requires language that looks like a specific limitation. It does not verify that the limitation is true, or even that it limits the conclusion: “may be fully trusted without any checking” would pass. The example also does not establish that a retrieved theorem is authentic, that its assumptions apply, that source attribution survives note-taking, or that reading becomes faster or more accurate. The opening student’s task remains a proposed use case, not a demonstrated result.
Where this leaves us A source, a fallible semantic record and a primary rendering make room for optional material beside the reader. Derived and retrieved name evidence origins. Interactive names a mode of engagement that may draw on either and may introduce model inferences. The sidecar’s job is to keep those components visibly separable, traceable and dismissible, without quietly rewriting the primary source. Its availability and mechanical labelling have a scripted demonstration. Its claimed reading benefits still need a reader study. The right outcome may be an empty panel or an explicitly unresolved question.
The progression is deliberate. The representation was fixed and then made transformable. A structured record gave one place to keep what a rendering was meant to preserve. Now external material can appear alongside the source. But preserving origins is not enough to choose which external candidate deserves the reader’s attention. A theorem can be topically similar yet inapplicable. A paper can contradict a claim under one set of assumptions and support it under another. A later version can supersede an earlier one. The next question is therefore not simply what did retrieval find? It is what relationship does this candidate bear to the claim being read, and what evidence warrants that relationship?
References Guo, Z., Schlichtkrull, M. & Vlachos, A. (2022). A Survey on Automated Fact-Checking. Transactions of the Association for Computational Linguistics, 10, 178–206. DOI 10.1162/tacl_a_00454. (Used for the distinction between claim identification, evidence retrieval and checking; no model in this chapter performs a validated fact-check.) Horvitz, E. (1999). Principles of Mixed-Initiative User Interfaces. Proceedings of CHI ‘99, 159–166. DOI 10.1145/302979.303030. (Used for uncertainty, direct invocation and interruption-cost design principles, not for claims about the present sidecar’s effectiveness.)