Research
First results in preparation.
Trace is building a measurement instrument, and an instrument is only as good as the published evidence that it measures what it claims to. Nothing has been published yet. This page lists the studies that are planned, so the claim can be checked later against what actually appeared.
Publication pipeline
10 planned · 0 published
Nothing on this page is a publication. The pipeline stays drawn this way until something real moves along it.
Publication status
No published studies yet
There are no papers, preprints, technical reports, or DOIs to list. When the first results are ready they will appear here with their data, their methods, and their limitations. Until then this section stays empty rather than being filled with something that reads like evidence.
Planned studies
What is being worked on, stated before the results exist.
Each entry is a question and an approach. None of them has an author list, a date, a venue, or a result, because none of them has been run to completion.
Measuring AI-mediated discoverability of biomedical literature
Question. Can the probability that a given biomedical publication is retrieved for a realistic clinical or scientific question be estimated reliably, and with what error?
Approach. A stratified sample of publications across fields and age bands, measured against versioned query sets, with repeated sampling to estimate variance.
Planned — no results
Stability of scientific citation across generative answer systems
Question. When the same question is asked repeatedly, and asked of different systems, how stable is the set of sources cited?
Approach. Repeat measurement over a fixed query set and a fixed window, reporting agreement statistics rather than a single ranking.
Planned — no results
Retrieval, citation, and evidence absorption are not the same measurement
Question. How often is a paper retrieved but not cited, cited but not used to support the claim, or used without being cited?
Approach. Paired annotation of retrieved sets, cited sets, and the claims an answer actually rests on, with inter-annotator agreement reported.
Planned — no results
Effect of open availability on AI evidence retrieval
Question. Does a lawful open version of an article change how often it is retrieved and used, after controlling for field, age, and journal?
Approach. Observational comparison with explicit confounder handling, and a pre-registered statement of what would and would not count as evidence of an effect.
Planned — no results
Query-set construction for evidence-visibility measurement
Question. What makes a query set representative of how research questions are actually asked, and how sensitive are results to that construction?
Approach. Comparison of query sets built by different procedures against the same corpus, reporting how much the construction choice moves the answer.
Planned — no results
Synthetic versus real stakeholder panels
Question. Where do synthetic stakeholder panels agree with human respondents, and where do they diverge in ways that would mislead?
Approach. Paired studies against human data, reporting the divergences prominently. This is the study most likely to constrain the product, which is why it is on the list.
Planned — no results
Narrative drift in artificial stakeholder societies
Question. How does an interpretation of a scientific finding change as it propagates through a simulated population of agents?
Approach. Controlled simulation with fixed seeds and published parameters, reported strictly as sandbox dynamics rather than as a forecast of real discourse.
Planned — no results
Detecting overstatement in AI-generated summaries of research
Question. Can a summary that overstates a study's population, endpoint, or certainty be detected reliably against the source?
Approach. Annotated corpus of summaries paired with source claims, evaluating detection against human judgement with agreement statistics.
Planned — no results
Longitudinal impact of metadata improvements
Question. When a concrete metadata defect is corrected, does measured discoverability change, and over what timescale?
Approach. Before-and-after measurement on a fixed instrument, with a comparison set, and with explicit acknowledgement of what an observational design cannot establish.
Planned — no results
Field-normalized visibility
Question. What normalization makes visibility comparable across fields without hiding the differences that matter?
Approach. Comparison of normalization schemes against held-out judgements, reporting where each scheme distorts.
Planned — no results
The sequence ends here. There is no next node until a study is finished.
Commitments
How results will be reported when they exist.
- Methods and query-set definitions are published before results are claimed.
- Negative results are published, including measurements that did not work.
- Sampling parameters, agreement statistics, and error estimates accompany every figure.
- No study is cited on this site until it exists and can be read.
Nothing here is a citation