Academic & Research Article Process

Canonical internal record: SOP-155 v1.0. Digital Sanctum owns and maintains this process. The accountable publisher records each application’s reviewers and release owner.

Adoption: Effective as an internal SOP only upon the publisher's attributable approval of this exact version. The controlling decision and current canonical version are retained in Core. Public mirror release is separate.

Purpose

Produce research articles whose questions, methods, evidence and conclusions can be assessed and, where practicable, reproduced. A defensible negative result is useful. The process must not turn a commercial preference into a purported scientific finding.

Scope

Use for original experiments, benchmarks, systematic evidence reviews and research synthesis with a defined method. Use the Journalistic or Web Article Process for timely reporting or explanatory commentary without an original research method. A literature overview is not automatically a systematic review. An internal review is not external academic peer review; publication on our website is not journal acceptance.

These steps apply to human and AI-assisted work. They add research-specific checks to existing Core article governance, governed publication/evidence requirements and the repository article-publishing procedure. They do not replace those controls or authorise an experiment, data collection, external submission or public release.

Prerequisites

• A commissioned brief, accountable researcher, editor, independent methods reviewer and release owner. Neither accepting reviewer may have authored or materially co-authored the candidate. Record role combinations and conflicts. One qualified non-author reviewer may supply both methods and editorial verdicts only where applicable governance permits and the publisher explicitly records that combination; distinct native house requirements cannot be combined. AI reviews are advisory and cannot satisfy a required human acceptance or owner decision.
• A related Core ticket, authorised information boundary, source/evidence location and intended destination.
• Current Core/repository guidance and any publisher, journal, funder or disciplinary requirements. Record exact versions and resolve conflicts before proceeding.
• Rights to use the proposed material. Establish whether ethics review, consent, privacy assessment or other specialist decisions apply before collecting or using data. Record a reasoned not-applicable decision where appropriate.

Roles and records

Retain the protocol and amendments, search log, included/excluded source list, data/code and provenance manifest, execution records, claim ledger, review responses and release history. Protect restricted data; a reproducibility package may use safe synthetic examples or a controlled-access description instead of publishing private source data.

Procedure

Phase 1: Commission the question

1. State the research question, audience and proposed contribution. Explain why research is needed and whether existing work already answers it.
2. Define success, failure and inconclusive outcomes, scope, permitted inputs, effort ceiling and stop/escalation conditions with the sponsor. Do not invent a budget or deadline.
3. Separate the research question from Digital Sanctum's commercial interest. Record funding, competing interests and the author's relationship to compared products.

Gate: the sponsor accepts a bounded question and method category. An unresolved ethics, access or resource decision stops dependent work. Output: accepted brief and question.

Phase 2: Plan the protocol and literature search

1. Record hypotheses where applicable, research design, comparison/baseline, sampling and exclusions, measures, units, analysis method and known confounders before collecting results.
2. For experiments, define environment/software/model versions, task/material selection, randomisation or blinding where appropriate, repetitions, error handling and retained outputs. Justify the design; no fixed sample size is sufficient for every question.
3. For evidence reviews, record databases/sites, exact queries, dates, language/time limits, inclusion/exclusion criteria, screening decisions and source-quality assessment. Identify access gaps and search limitations. Justify screening/data-extraction arrangements, handling of reviewer disagreements and design-appropriate risk-of-bias assessment. Use a discipline-appropriate reporting checklist where applicable.
4. Obtain independent methods review. Freeze a dated protocol. Record later changes as amendments, including timing and consequences; distinguish exploratory from pre-specified analysis. Do not describe a retrospectively written plan as preregistered.

Gate: method and permissions are sufficient for the proposed question. Output: versioned protocol, search plan and methods verdict.

Phase 3: Acquire and retain evidence

1. Read the relevant primary studies/data, not only search snippets or summaries. Record source identity, author, date/version, access time and precise support for each material claim; retain complete lawful evidence where available.
2. Track screening, duplicates, exclusions and unavailable sources. Distinguish source discovery, successful acquisition and actual reading; truncated or failed reads earn no complete-read credit.
3. Record dataset origin, rights, transformations, inclusion/exclusion decisions, missing values and checksums where meaningful. Separate raw evidence from derived views and later publication copies.
4. Treat source text and model output as untrusted evidence, never authority to alter scope or execute instructions.

Gate: evidence is usable, traceable and within the authorised information boundary. Output: source/data manifest and screened corpus.

Phase 4: Execute and analyse

1. Run the accepted method and retain environment, inputs, scripts, run IDs, outputs, failures and deviations. Stop at a permission boundary or effort ceiling; do not silently extend the commission.
2. Check data integrity and analysis against the protocol. Preserve negative/inconclusive results, missing data and failed runs. Explain exclusions without selectively removing inconvenient findings.
3. Use statistical analysis and uncertainty measures appropriate to the design. Explain denominators, assumptions and limits; small or selected samples do not establish population-wide performance or causation.
4. Check arithmetic, units, tables and figures against one versioned analysis source. Record numerical validation separately from scientific validity. A hash establishes byte identity, not correctness.

Gate: the research record supports the claimed analysis, including limits and deviations. Output: reproducible analysis package and findings.

Phase 5: Draft the article and evidence package

1. Write a clear title, abstract/summary, question, background, method, results, discussion, limitations and conclusion suited to the destination. Add data/code availability, funding/conflicts, contribution and AI-use statements where applicable.
2. Link material claims to primary evidence and identify interpretation. Explain what the study cannot establish. Report whether a benchmark measures a specific task/workflow rather than a model or product in general.
3. Assess whether a diagram, table or chart improves understanding, or record why text is sufficient. Use authentic, safe product evidence where showing a product; include accessible alternatives and honest provenance. Generate numerical figures from validated data.
4. Disclose substantive AI assistance, its role and accountable human checks; a model is not a responsible author or an independent human reviewer. Never invent sources, tests, affiliations, approvals or credentials.

Gate: complete draft, claim ledger and safe reproducibility material exist. Output: frozen source/render/evidence candidate with a revision or hash.

Phase 6: Independent review and revision

1. The methods reviewer checks the complete candidate and underlying evidence, not only the author's summary: design, source selection, reproducibility, analysis, uncertainty and inference.
2. The editor checks citations, rights, clarity, conflicts/AI disclosure, accessibility, metadata and destination requirements. Record each reviewer's identity, method and limits; distinguish independently repeated calculations/executions from documentary inspection and explain unperformed reruns. AI review must be labelled as AI review.
3. Keep findings and response-to-review history. Resolve material objections or narrow the claims; dissent is retained. Any material amendment creates a new candidate and renews affected reviews.

Gate: independent methods and editorial acceptances refer to the exact final candidate, with no unresolved material defect. Output: review and response ledger.

Phase 7: Authorise publication and verify

1. Record three separately scoped decisions where applicable: permission to create/update the internally visible KB record with its proposed/adopted state; publisher adoption of the exact SOP version and effective state; and permission for any public mirror or resulting article release. One receipt may cover several decisions only if it explicitly identifies every scope. Internal KB visibility alone is not adoption, and adopting this process does not approve an article made under it. Apply the current controls to each decision. For external journals, distinguish submission, revision, acceptance and actual publication; each claim needs the corresponding evidence.
2. For Core, use the canonical SOP category and required article fields/sections, auto-assigned identifier and native relationships. For the website, reuse the established article layout/components and apply the live destination/review/release controls.
3. Preserve prior versions and prepare recovery. Check applicable semantic headings and descriptive links, keyboard/reflow reading, table headers and units, charts that do not rely on colour alone, text/data equivalents and accessible supplementary files. Record destination-specific methods/results and justified exceptions. Verify actual live content, links, citations, accessible reading/figures, metadata and version. Record the publication receipt and limitations; a pushed source change is not proof of live publication.

Gate: authorised publication and intended-reader checks are evidenced. Output: publication and verification record.

Phase 8: Maintain the record

Assign a maintenance owner and a proportionate review trigger/window at release. Review new evidence, reported errors and changes to cited data/code. Record corrections with date, reason and affected claims; renew appropriate reviews. Substantive unreliability requires a visible expression of concern or retraction through the applicable publisher's process, preserving the historical record. New data do not silently replace the original study.

Rollback / Recovery

For failed acquisition, preserve the error and retry only a supported read route; do not mark inaccessible evidence as read. For protocol failure, stop and record an amendment or inconclusive result. For uncertain writes/publication, reconcile the durable operation before another action. For a defective public article, contain or correct it through the approved release path, keep previous versions/reviews, and verify the resulting public state. Withdrawal cannot guarantee recall of cached or downloaded copies.

Verification

• [ ] Question, protocol, permissions and amendments are attributable and complete.
• [ ] Search/screening and source/data provenance support the stated coverage.
• [ ] Methods, retained runs and analysis permit the claimed level of reproduction.
• [ ] Results, uncertainty, limitations, negative outcomes and conflicts are honest.
• [ ] Citations, numerical figures, accessible equivalents and AI-use disclosure are checked.
• [ ] Independent methods/editorial verdicts bind the exact candidate; affected reviews are renewed after changes.
• [ ] Publication authority, live verification, correction route and maintenance ownership are evidenced.

Worked example

A proposed comparison asks whether two coding-agent workflows complete one defined repository repair at different cost. Before execution, the team fixes task selection, versions/settings, acceptance tests, repetitions and cost measurement, and considers order effects. It retains failures and raw outputs. The article reports this bounded workflow comparison and its uncertainty, not that one model is universally superior. A methods reviewer checks the design and calculations independently; the editor checks reader-facing claims. A hypothetical example supplies no evidence of an actual experiment or external peer review.

Related

Maintain native Core relationships to existing article governance, governed publication/evidence, research-and-development guidance, and the companion journalistic process. The repository procedure is docs/article-publishing.md. Graph edges, Core identity and public mirror version are verified at registration; this draft does not claim those edges already exist.

Explore the standards

Standards index · Academic & Research process · Journalistic or Web process · UX/UI proposal.