How DTRI does research
The Deep Time Research Institute is an AI-native research program by design, not by shortcut. AI is used as core research infrastructure under direct author supervision. All factual claims, statistical results, and citations are independently verified before any manuscript leaves DTRI.
This page documents the methodology so that reviewers, readers, and other researchers can hold DTRI's work to its stated standards. If you find a discrepancy between this page and the published record, please tell us — we will publish a correction.
Every empirical study at DTRI is pre-registered on the Open Science Framework before data collection or analysis begins. Pre-registration documents the analytic plan, the predicted effect, and the criteria for confirmation or falsification — locked in time-stamped form before results are seen.
Two examples currently in the public record:
Pre-registered studies that fail confirmation are reported as nulls. See the Null Results page for the standing list.
Every manuscript that leaves DTRI runs through a three-layer reference audit before submission — and again after every substantive revision.
An automated audit script (audit_manuscript_citations.py) resolves every cited reference against CrossRef, Open Library, Semantic Scholar, and PubMed. DOIs, author lists, journal names, years, and page ranges are checked against the canonical metadata returned by each registry. Discrepancies are flagged for manual review.
Citations to high-profile authors (where misattribution is most consequential) get a second pass via direct web verification of the author's publication record. This layer exists because programmatic registries occasionally return plausible-looking but wrong matches for common names, and because reviewer trust is most easily lost on the citations a reader is most likely to recognize.
A compliance script (manuscript_compliance_check.py) reads the final manuscript for venue-specific submission requirements: prior-submission disclosure (mandatory at every DTRI target venue), conflict-of-interest statements, declarations of AI assistance, ethics statements where applicable, and word-count limits. The author-contributions statement and the AI declaration must be internally consistent — the compliance check enforces this.
Errors identified by the three-layer audit are described in cover letters as "discrepancies identified during a pre-submission reference audit." This is DTRI's standard phrasing — it accurately describes the process and the source of any correction.
Every DTRI paper is paired with a Zenodo deposit containing the analytic data, the analysis code, and where applicable the simulation scripts and Monte Carlo seeds. DOIs are visible on the papers page next to each paper.
Statistical results in DTRI papers are computed programmatically from the deposited data and code. Numbers do not appear in manuscripts unless they are produced by a script that is itself part of the deposit. This rule is enforced by the audit pipeline: any numerical claim that cannot be located in a JSON output file is flagged.
If a reader wants to reproduce a result, the deposit + code + a Python environment are sufficient. We have no private working data; what you see is what we used.
Elliot Allan is the sole author of all DTRI manuscripts to date and is responsible for all content — including any errors. AI assistance is disclosed accurately in every submission. When AI was used for generative drafting, the disclosure reflects generative drafting, not copy-editing. The author-contributions statement is required to align with the AI declaration; the compliance check enforces this.
The principle: AI is a research tool under direct supervision. It does not substitute for author judgment, and it does not absolve the author of responsibility for the final product. Where AI was used, that use is documented; where the work depended on AI in non-routine ways, the methodology section reflects that.
DTRI's view is that AI-native research methods are part of the scientific record, not a footnote to it. Hiding their use is what creates concern; documenting their use is what makes the work auditable.
Every paper runs through an in-house adversarial review pass before submission. Multiple AI models — Claude (Anthropic), Perplexity, Gemini (Google), Grok (xAI) — each independently critique the manuscript, looking for unsupported claims, statistical errors, citation problems, framing weaknesses, and reviewer-bait. The author then responds to each round in writing, either tightening the manuscript or recording why a critique was rejected.
This is not a substitute for peer review — it's a pre-submission cleanup pass that catches issues which would otherwise burn reviewer time. The actual scientific judgment of whether DTRI's findings hold up is for the peer-review process at each venue to decide.
Work that touches Indigenous knowledge systems is held to a separate gate. For research drawing on Aboriginal Australian, Native Californian, West African, or other Indigenous traditions, DTRI engages AIATSIS-equivalent consultation pathways before submission. Some manuscripts are held indefinitely pending appropriate cultural review.
Three current examples are held on this gate:
These will remain on hold until the consultation process closes. "Held" is a deliberate research decision, not a delay.
DTRI publishes null results alongside positive results. The Null Results page catalogs hypotheses that DTRI has tested and falsified — including the 108° sacred-longitude spacing hypothesis (falsified), monument-orientation alignments (null), and three competing causal mechanisms for the great-circle corridor pattern (agricultural, lithological, groundwater — all null).
Independent research institutes that hide their nulls cannot be trusted. We make ours visible.