Source-linked AI summary
Who Asked for This? Inline Annotations as Authoring Transactions for Provenance in Agentic Authoring
Chang Xiao
TL;DR
Agentic writing turns paragraphs into the outcome of many requests, yet final documents rarely show which request produced each change. Reactant records typed inline requests as validated transactions linking additions, deletions, and transformations across revisions. The paper demonstrates its revision history and reports colleagues’ use of the transaction record for inquiry and reusable skills.
Problem
Final text often does not reveal which request produced a change, and existing histories can make individual contributions difficult to trace when intentions are intertwined.
Source: §1 · p. 1Method
Reactant uses typed inline annotations as contextual requests and validates structured edit witnesses against recorded before/after states to establish lineage across revisions.
Source: §1 · p. 2, §3.4 · p. 6Results
Takeaways & Limitations
The transaction record can support later inquiry into revision history and extraction of reusable skills from recurring requests.
Source: §1 · p. 3Use cases center on experienced researchers’ academic writing, and no formal comparative usability study establishes whether Reactant is easier, more efficient, or preferable to other workflows.
Source: §6 · p. 9
Abstract
from arXiv · showhide
Writing with AI agents turns a paragraph into the outcome of many requests, yet the finished document rarely explains which request produced which change. We introduce Reactant, an interaction paradigm in which authors place typed inline annotations in their original documents. A verified transaction protocol records each request, skill identity, and the correspondences that identify additions, deletions, and transformations. The kernel validates the witness against the recorded states to establish word-level longitudinal lineage. We demonstrate Reactant through this paper's revision history, and report four months of three colleagues' self-directed use. Their uses include conversational inquiry into history and deriving reusable skills from recurring requests, illustrating how the transaction record serves as an extensible substrate for agentic authoring.
1 Introduction
Agentic writing can turn a paragraph into the outcome of many requests, but the final text does not explain which request produced each change. Reactant records contextual inline requests as validated transactions, preserving lineage across revisions and supporting extensions beyond revision inspection.
1 Introduction
- Final text alone cannot answer who requested an unfamiliar sentence, while post hoc diffs make intertwined changes difficult to trace to individual requests.The paper connects this gap to misremembered human–AI attribution [38] and writers’ desire for control over their contributions [27].Source: §1 · p. 1
- Reactant lets authors place typed, scoped requests beside the text in their existing documents, making each annotation a contextual unit of delegation.Compatible agents read the annotation in context and carry out the requested revision.Source: §1 · p. 2
- The kernel records each request, skill identity, and before/after state, then validates structured edit correspondences to establish lineage across revisions.The correspondences identify additions and deletions and link transformed wording to its predecessor.Source: §1 · p. 3
- Reactant keeps drafting and revision in the same document, works independently of a particular editor, and allows authors to define extensible annotation operations.Built-in operations include proofreading, filling placeholders, and adding citations; custom operations use the same transaction mechanism and contribute to provenance history.Source: §1 · p. 3
- Four months of use and user-developed extensions demonstrate transaction histories supporting conversational inquiry and reusable skills derived from recurring requests.The paper presents the history as an extensible substrate beyond revision inspection.Source: §1 · p. 3
2 Related Work
Prior AI-writing tools record authorship, activity, prompts, or revisions, but many depend on particular environments or group multiple decisions into one history entry. Reactant instead treats each inline request as a transaction and validates explicit transformation links across revisions.
2 Related Work
- Many AI-writing capabilities depend on particular editors or platforms, potentially requiring writers to switch tools or transfer text from established workflows.Related systems include in-app writing assistance and interfaces for direct manipulation and layered editing [22] [30].Source: §2 · p. 3
- Provenance systems capture varied traces, including typed versus AI-assisted text, revision states, phrase-level provenance, prompt-adopted text, and generated content.Examples include Grammarly Authorship, DraftMarks [29], HaLLMark [17], and InkSync [18].Source: §2 · p. 3
- A conversational turn or trigger can group several writing decisions into one revision, so retained history alone does not guarantee a distinct record for each request.Claude Code checkpoints associate snapshots with prompts, while aider can group inline instructions and retain changes through Git.Source: §2 · pp. 3–4
- Reactant executes each typed inline annotation separately and validates an agent-submitted edit witness against before/after states before establishing longitudinal lineage.The shared protocol retains the request and skill identity and can be extended by compatible agents and custom skills.Source: §2 · p. 4
3 System Implementation
Reactant combines inline, editor-independent writing requests with a transaction protocol that records changes and links revisions. Its kernel validates agent-submitted edit witnesses, while Replay lets authors inspect word origins and passage histories.
3 System Implementation
- Reactant records each request, skill identity, exact before/after states, and validated witness to establish word-level longitudinal lineage.The implementation preserves the declared request and observed change, but does not establish that the agent understood the request or produced correct text.Source: §3 · p. 4
3.1 Writing Requests in the Document
- Annotations place a skill and optional natural-language instruction in the document; placement supplies context, while markers can restrict editable or protected text.Double-backtick references can identify additional files to read or write.Source: §3.1 · p. 4
3.2 A Holistic Authoring Example
- The report-writing example shows inline requests alongside report updates and supporting files, with each request recorded separately even when several run together.The example includes citation, prose revision, discussion, and artifact-generation requests.Source: §3.2 · p. 4, Figure 2 · p. 5
3.3 Routing and Extending Skills
- The router runs the first actionable annotation in source order, commits its witness and states, rereads the file, and then selects the next annotation.Authors can add custom skills under .reactant/skills/ without changing the registry.Source: §3.3 · p. 4
3.4 The Transaction Kernel
- Each transaction retains the request, skill definition, agent metadata, context hashes, version identifiers, and validated witness; writes are inspectable but not crash-atomic across files.The integrity command reports inconsistencies rather than reconstructing missing history.Source: §3.4 · p. 6
- The kernel validates agent-declared edit correspondences against before/after snapshots, linking transformed wording across requests while retaining snapshots for restoration.The witness can represent additions, deletions, and transformations; consistency with snapshots does not establish semantic correctness.Source: §3.4 · p. 6
- Snapshots alone cannot distinguish a correction from deleting a word and independently generating its replacement; explicit witness links preserve the declared transformation.Git may store transaction records, but Git alone does not provide the protocol that elicits, validates, and retains links at each request boundary.Source: §3.4 · p. 6
3.5 Inspecting History with Reactant Replay
- Replay displays snapshots and word origins, and its Trace view follows a word or passage through revisions and links each step to its request and comparison.Initial text and unbound changes remain labeled when no recorded request explains their origin.Source: §3.5 · p. 6
4 Tracing This Paper’s Introduction
The authors use Reactant Replay to inspect the recorded history of a passage in this paper’s introduction. The trace links requests to successive changes, including drafting, citation insertion, source correction, and reframing.
4 Tracing This Paper’s Introduction
- The revisions combine drafting, citation insertion, source correction, and reframing; the final editor-independence passage connects retaining a preferred editor to prior findings.The history shows the argument, source, and source role developing through separate requests.Source: §4 · p. 6
- The selected passage’s ancestry contains 62 recorded states; its interval from the first selected request to the endpoint has 25 states: 14 transaction-backed and 11 unbound.The endpoint was saved before this section was drafted; each request preserves its input state.Source: §4 · p. 6
- 38 successful skill invocations appear in the introduction’s full recorded ancestry: 18 edit, 8 proofread, 6 ph, 4 cite, and 2 plan.Twenty used Codex (with GPT-5.6-sol) and 18 used Claude Code (with Opus 5); counts include revisions outside the selected passage.Source: §4 · p. 6
- Opening the current “faster” claim reveals R4’s earlier formulation, where the passage described LLM-for-X and its popup interface.R4 asks to omit the paper’s name and connect the claim to prior findings.Source: §4 · p. 7
- R1’s request for three advantages becomes prose, including the initial editor-independence sentence.The trace connects the inline instruction to the resulting claim.Source: §4 · p. 7
- R2 adds an in-context description of LLM-for-X and citation, linking the prose change to a bibliography update.Before R2, the passage contains the editor-independence argument without that supporting example.Source: §4 · p. 7
- R3 changes the citation key to the original LLM-for-X paper [33] while preserving the descriptive sentence and updating the bibliography.Source: §4 · p. 7
- The trace distinguishes R3’s source correction from R4’s later reframing of what the publication should support.The sequence shows the rationale acquiring a cited example, a corrected source, and a different role for that source in the finished argument.Source: §4 · p. 7
5 Reactant in Practice
Three experienced academic researchers used Reactant over four months, describing how its history supported recovering decisions, continuing work across agents and tasks, and developing new uses. These are situated examples, not a summative evaluation.
5 Reactant in Practice
- Three researchers each reported at least 100 transactions across three or more projects during four months of self-directed use.The authors present their accounts as situated examples rather than a summative evaluation of usability or effectiveness.Source: §5 · p. 7
5.1 Recovering the Request behind a Revision
- Reactant fit users’ existing habit of leaving document comments, and E1 valued placing agent requests in the writing document without switching windows.E1 also created a keyboard shortcut to run /execute, which the authors later incorporated into Reactant.Source: §5.1 · pp. 7, 9
- Replay helped E1 trace a puzzling passage to a vague earlier instruction that had also changed content in another file.E1 also used the history to distinguish their own writing, AI-proofread text, and passages generated entirely by AI.Source: §5.1 · p. 9
5.2 Continuing Work across Agents and Tasks
- E2 valued Reactant’s model-agnostic history because rate limits or model-capability differences often led them to switch agent platforms.Comparing non-adjacent versions helped E2 inspect the cumulative effect of several prompts and skills on a passage.Source: §5.2 · p. 9
- E2 defined task-specific skills, including one that converted technical-paper concepts into explanations suited to a course audience.E2 used this skill while preparing course materials.Source: §5.2 · p. 9
5.3 Building New Uses from the History
- E3 used an agent to identify recurring style instructions in transaction records and generate a reusable writing-style skill.E3 also regularly asked an agent to explain a sentence’s history, then used Replay to confirm the explanation against the record.Source: §5.3 · p. 9
6 Discussion and Conclusion
Reactant preserves request boundaries and explicit transformation links across agents and custom operations, supporting later inspection and reuse of document history. The reported experiences are concentrated in experienced academic writers, and the authors have not conducted a formal comparative usability study.
6 Discussion and Conclusion
- Reactant’s recorded requests, snapshots, and validated edit witnesses make delegated operations and their declared correspondences available for later inspection.The authors describe this as an extensible substrate for revisiting decisions, inspecting how passages developed, and supporting subsequent agent work.Source: §6 · p. 9
- The use cases center on academic writing by experienced researchers, and no formal comparative usability study establishes ease of learning, efficiency, or preference over other workflows.The authors therefore limit the demonstrated contribution to retaining request boundaries and explicit transformation links across compatible agents and custom operations.Source: §6 · p. 9
- Future work includes extracting reusable skills from recurring instructions, supporting conversational inquiry into document history, and studying custom-skill development and sharing.The authors also plan to evaluate effects on authoring and retrospective review and explore applications beyond manuscript writing.Source: §6 · p. 9