Developing a paper and submission

Choose the argument from the evidence

Begin with the question the paper answers and the result that makes the answer worth reading. Follow its experiment links before outlining the manuscript. A finding can explain how several results fit together; the individual experiment records hold the measurements and their limitations.

Claims link to the quantities being estimated, called estimands, and the work that supports them. The evidence network guide traces these links, including evidence shared by several papers.

Add publication claims to CLAIMS.md: statements that could appear in an abstract, contribution list, result heading, or central figure caption. Give the paper a stable key and use it in the ledger's Paper column. Mark each claim's role in the argument and cite the specific experiment or estimand that supports it. Speculative claims remain provisional.

In the synthetic retrieval example, the paper might investigate improvement on long documents rather than claim improvement on all questions. The overall test, selected subgroup analysis, and independent confirmation have different evidentiary roles. Human promotion review determines which claims the paper can adopt and their wording and scope. A promising figure does not establish that approval.

Create PUBLICATION.md when the paper enters preparation. It connects the stable paper key to its draft, headline claims, unresolved blockers, and possible venues. Put the next action in PRIORITIES.md. If resolving a blocker requires several experiments, give that work a research plan. The publication record tracks the paper's preparation; the plan governs execution.

## Paper: Retrieval on long documents

**Paper key**: retrieval-long-documents
**Draft**: papers/retrieval-long-documents.typ
**Target venue**: not selected

### Headline result

C1 is provisional. Its origin is the found analysis in
EXP-102:E1; EXP-103:E1 is the planned confirmation.

### Blocking issues

| Issue | Severity | State | Action |
|---|---|---|---|
| Confirmation not interpreted | blocks headline claim | open | Review EXP-103 before promoting C1 |
| Artifact package not checked | blocks release claim | open | Verify the declared reproduction path |

Give the manuscript one continuing source

Use papers/ for work this project authors. The manuscript convention gives a new notebook-resident draft a Typst source with a neutral, natural-length layout. Manuscripts share papers/references.bib. For a project with an established external manuscript workspace, record that location and keep the draft there.

The manuscript can hold the complete methods, qualifications, and supporting material before a venue requires cuts. Draft around the selected claims, then trace each quantitative statement and figure back to the experiment that produced it. An analysis needed by the paper but absent from the notebook needs an analysis experiment. Record its inputs and retrospective status.

Keep private evidence links available to the authors without exposing machine-local paths as public citations. The published manuscript needs an appropriate methods description and lasting public artifact references where promised. A notebook citation helps the team recover the evidence; readers still need an explanation of the result.

Build a reading archive from the draft's citations

When checking related work or revisiting a borrowed method, start from the manuscript and its bibliography. The download-research-references skill identifies cited works, finds known authorized copies, and assembles a local archive. It follows supported local includes, so a manuscript split across source files can still provide the reading list.

Use $download-research-references to preview the works cited by this manuscript and its bibliography. Show which copies are available and which are missing before downloading.

Review that preview for missing or misidentified references, then authorize the download. The PDFs go under references/, separate from your own papers/. BIBLIOGRAPHY.md is the annotated reading index. Add what a work contributes, how it relates to this project, and any limitation relevant to your use of it. Keep existing notes and thematic groupings when adding entries.

A paper without an available authorized PDF still gets a source link and a reason the PDF is unavailable. A collaborator can then see what remains to retrieve or read. If source text would help citation tracing, request arXiv source retrieval as well; its inventory is references/SOURCES.md. The archive covers cited literature. It does not search for all related work or substitute for reading it.

The archive reference explains output locations and command-line controls. Downloaded PDFs may stay outside version control while source URLs, annotations, and inventories remain tracked.

Prepare a submission packet for one venue

Once a venue is a serious candidate, verify its current submission instructions from primary sources. Record the source and date checked alongside format limits, anonymity requirements, and the published deadline with its time zone. Separate distinct submission stages if the venue has them. An internal aspiration is not a venue deadline; use a readiness gate and urgency to express what work should happen next.

A submission packet contains the venue-specific manuscript, required supporting files, and any submission-form answers or artifact archive that the venue asks for. Keep it in the external paper workspace and link it from PUBLICATION.md. The continuing manuscript stays in its recorded location. A venue that accepts the manuscript's format directly may need only a frozen build, with no separate source tree.

Keep the complete method and qualifications in the manuscript when cutting a packet to the page limit. The content rule is packet ⊆ manuscript: a packet may omit material but should not hold the only copy of a result or argument. Review anonymized PDFs and linked artifacts as a reader would, including metadata and acknowledgments. Changing the author line alone may leave identifying information elsewhere.

Track which manuscript version each packet reflects

During preparation, edit the manuscript first and port the relevant changes into its packet. Record each packet's location, venue, state, and synced-as-of date in PUBLICATION.md. Include the corresponding source revision when available to identify exactly which source was synchronized.

A submitted packet can lag behind the manuscript. Keep it frozen as the draft develops. A reviewer-driven revision or camera-ready edit may put the packet ahead temporarily. Bring those changes back into the manuscript before ordinary manuscript editing resumes. Note which version contains pending changes until they have been reconciled.

For the retrieval example, a shorter packet might omit a diagnostic breakdown while keeping the narrower main claim. A later correction to the scoring changes the manuscript and any packet still in preparation. A submitted version instead needs an explicit correction or revision decision under the venue's rules; overwriting its local file would erase what readers actually received.

Record readiness separately from submission

Before submission, check that the manuscript's claims match their reviewed evidence, the final rendered files meet the venue's requirements, and promised artifacts are accessible in the intended form. Remove resolved blockers from PUBLICATION.md and name any remaining gate. For a paper that is ready but waiting, record urgency and when readiness was established. Do not invent a deadline.

Preparation does not authorize publication. Obtain approval for the external submission, then preserve its receipt or identifier and the exact submitted packet. Update the publication state, record the milestone in CHANGELOG.md, and leave STATUS.md pointing to that record. A built PDF, a ready packet, and a successful submission are three different states.

Update the evidence and manuscript after review

Responding to review may require clearer prose, an analysis of retained observations, or a new measurement. Trace the request to the affected claim before scheduling more experiments. Use the evidence repair workflow to select the smallest adequate response, and a research plan if the response spans several gated phases.

Correct an experiment's result in its own record, update dependent findings and claims, then carry the correction into the manuscript and active packets. Preserve submitted snapshots and record how each correction was handled. Removing a result from one paper's argument does not remove the underlying experiment.

At handoff, a collaborator should be able to locate the continuing draft and the evidence for its central claims. They also need the exact submitted version and the next unresolved decision. The manuscript and packet reference and claim fields specify the records; the workflow overview connects them to the rest of the research work.