Cross-study synthesis

Drafted human-editable specification · v0.5.0 · split from “Synthesis and reconstruction” on October 21, 2026 · no agent-performance claim

Epistemic purpose

• Integrate results from many studies or data sets into a pooled estimate, a reconciled account, or a connection between literatures that no single study supports.

When this mode helps

• Several studies measure the same quantity with different methods, sites, or teams, and disagree or are too small alone; or two literatures address one problem without citing each other.
• Not this mode: reconstructing one past event from its physical traces, which is reconstruction from traces.

Agent actions and representations

• State the inclusion rule before searching; record every study screened and why it was kept or dropped; extract effect sizes or estimates with their uncertainties; model dependence between studies (shared data, shared teams, shared methods); test heterogeneity; report the pooled result with its spread.
• Inputs: the question, the inclusion rule, the study records and their provenance, the budget.
• Outputs: the study table, the dependence structure, the pooled estimate or connection, and a claim record.

Evidence and claim scope

• The pooled claim holds only for the population of studies the rule admitted. Dependent studies count once. A literature connection is a hypothesis until a test supports it.
• Distinguish a warranted decision at the time from a claim’s later assessed adequacy; append follow-up without rewriting the original record.

Transitions and stopping

• Stop when the pooled estimate is stable to the inclusion of any single study and the dependence structure is stated. Hand disagreements to hypothesis assessment; hand pooled quantities to estimation when a forward model is needed; hand a connection to theory for a mechanism.
• On graph v0.4.2: a composition, many O records compared at C and combined at K. In the decision tree, question 3 (aim: a pooled estimate) and question 7 (agreement across independent data and methods).

Characteristic failure

• Count dependent studies as independent; let the inclusion rule drift after seeing results; pool incompatible quantities; mistake a plausible literature connection for a tested result.

Human evidence and borrowing

• Defined for ecology and evolution by the meta-analysis literature (Gurevitch et al. 2018) and organized in synthesis centres (Carpenter et al. 2009). Synthesis-centre papers drew 82.7 against 42.9 mean citations (Hackett et al. 2021).
• Geoscience cases: a reference Earth model fitted to many data types at once (Dziewonski and Anderson 1981); a reconciled ice-sheet mass balance across altimetry, gravimetry, and input-output methods (Shepherd et al. 2012); four decades of Antarctic mass balance (Rignot et al. 2019).
• Connecting separate literatures: Swanson (1986) motivates literature-based discovery; Nersessian (2022) compares integration processes. One agent system’s synthesis statements were 57.9% accurate against 85.5% for its data analyses (Mitchener et al. 2025).
• The actions above are design hypotheses derived from these sources, not established optimal policies.

Evaluation against past work

• Reproduce a published synthesis from its study list: same pooled value within its stated spread, and the same verdict when one study is removed. Include a planted dependent pair and a study that fails the inclusion rule.
• Assessment definitions · Agent implementation

Carpenter, Stephen R., E. Virginia Armbrust, Peter W. Arzberger, et al. 2009. “Accelerate Synthesis in Ecology and Environmental Sciences.” BioScience 59 (8): 699–701. https://doi.org/10.1525/bio.2009.59.8.11.

Dziewonski, Adam M., and Don L. Anderson. 1981. “Preliminary reference Earth model.” Physics of the Earth and Planetary Interiors 25 (4): 297–356. https://doi.org/10.1016/0031-9201(81)90046-7.

Gurevitch, Jessica, Julia Koricheva, Shinichi Nakagawa, and Gavin Stewart. 2018. “Meta-analysis and the science of research synthesis.” Nature 555 (7695): 175–82. https://doi.org/10.1038/nature25753.

Hackett, Edward J., Erin Leahey, John N. Parker, et al. 2021. “Do synthesis centers synthesize? A semantic analysis of topical diversity in research.” Research Policy 50 (1): 104069. https://doi.org/10.1016/j.respol.2020.104069.

Mitchener, Ludovico, Angela Yiu, Benjamin Chang, et al. 2025. “Kosmos: An AI Scientist for Autonomous Discovery.” arXiv 2511.02824. https://arxiv.org/abs/2511.02824.

Nersessian, Nancy J. 2022. Interdisciplinarity in the Making: Models and Methods in Frontier Science. MIT Press. https://doi.org/10.7551/mitpress/14667.001.0001.

Rignot, Eric, Jérémie Mouginot, Bernd Scheuchl, Michiel van den Broeke, Melchior J. van Wessem, and Mathieu Morlighem. 2019. “Four decades of Antarctic Ice Sheet mass balance from 1979–2017.” Proceedings of the National Academy of Sciences 116 (4): 1095–103. https://doi.org/10.1073/pnas.1812883116.

Shepherd, Andrew, Erik R. Ivins, Geruo A, et al. 2012. “A Reconciled Estimate of Ice-Sheet Mass Balance.” Science 338 (6111): 1183–89. https://doi.org/10.1126/science.1228102.

Swanson, Don R. 1986. “Fish Oil, Raynaud’s Syndrome, and Undiscovered Public Knowledge.” Perspectives in Biology and Medicine 30 (1): 7–18. https://doi.org/10.1353/pbm.1986.0087.