A member academy of the International Academies of Social Sciencesxjass.org · formerly xjass.org.cn
XJASSÜrümqi

XJASS research record

January–March 2016 to June 2026

The digital record of this branch, listed from the most recent entry backwards.

June 20262 researches

  1. X252XJASSJune 2026

    Closing the Longitudinal Loop: A Machine-Learning Observatory of Xinjiang Evidence and Western Strategic Representation, 2023-2026

    Closing the longitudinal loop sits at the centre of how Xinjiang's interior economy is organised, linking resource endowment, infrastructure and settlement in ways that no single indicator captures. The study models a machine-learning observatory of Xinjiang evidence and western strategic representation, 2023-2026 using six classes of evidence — harmonized metadata and derived indicators from the series' public-source policy, environmental, infrastructure and economic dataset specifications, model-era records, institution-separated frames, provenance graphs and uncertainty annotations — assembled for 2023–2026 and linked throughout to collection date and source. Analysis uses longitudinal ensemble observatory combining temporal embeddings, institution-separated frame models, evidence graphs, change-point detection, model-era metadata and hierarchical uncertainty decomposition; no single generative summary is treated as ground truth; holdouts follow the geographic, temporal or institutional structure of the evidence so that performance is not inflated by leakage. Because investment and endowment often move together in Xinjiang, the design keeps physical conditions and policy-driven activity as separate inputs rather than merging them into one development score. Disagreement between specifications is preserved instead of resolved, as it marks the boundary of what the evidence can currently support. The output is a June 2026 reference point for closing the longitudinal loop, defined tightly enough that subsequent periods can be compared against it.

    Method notes are released to subscribers.

  2. X251XJASSJune 2026

    End-to-End Provenance Guarantees for Xinjiang Deep Research: From Query to Final Claim

    By June 2026, end-to-end provenance guarantees for Xinjiang deep research had become one of the clearest expressions of how investment, terrain and administrative geography interact across Xinjiang. This research examines from query to final claim across a purpose-built evidence base covering 2024–2026: bounded research questions, versioned public source universe, tool calls, calculations, claim DAGs, citations, injected failures and auditor decisions. The modelling strategy is provenance-first research architecture with immutable evidence IDs, typed tool calls, claim-evidence DAGs, citation entailment, deterministic calculation replay, failure injection and independent audit, checked against simpler baselines to confirm that the added complexity earns its place. The analysis distinguishes what the terrain and resource base make possible from what infrastructure and administrative decisions have actually delivered. Residuals are treated as evidence: the cases the model misses tend to mark infrastructure openings, regulatory changes or reporting shifts that the covariates do not contain. Taken together, the study provides a bounded June 2026 account of end-to-end provenance guarantees for Xinjiang deep research and an evidence architecture that later batches can reuse.

    Method notes are released to subscribers.

May 20262 researches

  1. X250XJASSMay 2026

    Industrial Energy-Water Co-Efficiency: Multi-Objective Learning Across Xinjiang Sectors

    Questions about industrial energy-water co-efficiency in Xinjiang turn quickly into questions about distance, water, energy and the corridors that connect the region to Central Asia. Working from 2021–2026, the study models multi-objective learning across Xinjiang sectors through an evidence base combining aggregate sector energy use, water use, output proxies, recycling, electricity mix, climate, location and efficiency indicators. Estimation relies on multi-objective gradient boosting and Pareto-front estimation with sector embeddings, climate/production controls, spatial holdouts, uncertainty intervals and sensitivity to normalization choices, with results reported alongside their uncertainty rather than as point claims. Because investment and endowment often move together in Xinjiang, the design keeps physical conditions and policy-driven activity as separate inputs rather than merging them into one development score. Systematic error is reported rather than smoothed away, because clustered residuals usually point to an omitted institutional or infrastructural factor. The result is a reproducible May 2026 baseline for industrial energy-water co-efficiency that later work can extend without re-specifying the evidence base.

    Method notes are released to subscribers.

  2. X249XJASSMay 2026

    Disagreement-Preserving Synthesis: Formal Metrics for US, UK and EU Xinjiang Research Agents

    Disagreement-preserving synthesis is difficult to observe directly in Xinjiang: the relevant processes are spread across counties, seasons and infrastructure networks that are recorded unevenly. The article examines formal metrics for US, UK and EU Xinjiang research agents, drawing on matched US/UK/EU public corpora, comparative tasks, specialist outputs, syntheses, citations, disagreement matrices, expert labels and separate Chinese comparison sources for 2023–2026 and recording provenance and missingness for each variable. The estimator is source-isolated specialist agents plus constrained synthesizer, disagreement matrices, semantic divergence metrics, claim-evidence graphs, source-swap tests and expert calibration, with validation structured around the way the data are actually generated rather than random splitting. The analysis distinguishes what the terrain and resource base make possible from what infrastructure and administrative decisions have actually delivered. Where the model fails, the failure is documented and located, since patterned error is often the most informative output of the exercise. Predictive performance is not treated as causal proof; the contribution is a dated, documented May 2026 benchmark for disagreement-preserving synthesis.

    Method notes are released to subscribers.

April 20262 researches

  1. X248XJASSApril 2026

    Dust-Storm Nowcasting With Geospatial Foundation Models: Short-Horizon Risk Across Xinjiang Transport Corridors

    Dust-storm nowcasting with geospatial foundation models sits at the centre of how Xinjiang's interior economy is organised, linking resource endowment, infrastructure and settlement in ways that no single indicator captures. The study examines short-horizon risk across Xinjiang transport corridors using seven classes of evidence — satellite aerosol products, wind, visibility, land surface, road/rail corridors, dated dust events, traffic disruption aggregates and lead-time labels — assembled for 2022–2026 and linked throughout to collection date and source. Analysis uses geospatial foundation embeddings fused with radar/satellite aerosol proxies, wind fields and corridor topology in a probabilistic spatiotemporal network; event holdouts and lead-time calibration; holdouts follow the geographic, temporal or institutional structure of the evidence so that performance is not inflated by leakage. Because investment and endowment often move together in Xinjiang, the design keeps physical conditions and policy-driven activity as separate inputs rather than merging them into one development score. Disagreement between specifications is preserved instead of resolved, as it marks the boundary of what the evidence can currently support. The output is an April 2026 reference point for dust-storm nowcasting with geospatial foundation models, defined tightly enough that subsequent periods can be compared against it.

    Method notes are released to subscribers.

  2. X247XJASSApril 2026

    Cross-Model Research Reproducibility: Do Different Frontier Systems Reach the Same Xinjiang Evidence?

    By April 2026, cross-model research reproducibility had become one of the clearest expressions of how investment, terrain and administrative geography interact across Xinjiang. This research examines do different frontier systems reach the same Xinjiang evidence? across a purpose-built evidence base covering 2024–2026: bounded questions, versioned public sources, multiple model trajectories, retrieved evidence, claims, citations, costs and expert reproducibility scores. The modelling strategy is multi-model deep-research replication with fixed source universe and tool permissions, evidence-set overlap, claim consistency, citation reproducibility, cost accounting and blinded expert adjudication, checked against simpler baselines to confirm that the added complexity earns its place. The analysis distinguishes what the terrain and resource base make possible from what infrastructure and administrative decisions have actually delivered. Residuals are treated as evidence: the cases the model misses tend to mark infrastructure openings, regulatory changes or reporting shifts that the covariates do not contain. Taken together, the study provides a bounded April 2026 account of cross-model research reproducibility and an evidence architecture that later batches can reuse.

    Method notes are released to subscribers.

March 20262 researches

  1. X246XJASSMarch 2026

    Spring Irrigation Allocation as Safe Reinforcement Learning: Exploring Water-Efficiency Policies in Simulation

    Questions about spring irrigation allocation as safe reinforcement learning in Xinjiang turn quickly into questions about distance, water, energy and the corridors that connect the region to Central Asia. Working from 2022–2026, the study models exploring water-efficiency policies in simulation through an evidence base combining aggregate irrigation allocations, crop-demand proxies, reservoir/river states, weather, evapotranspiration, basin constraints, historical outcomes and simulator scenarios. Estimation relies on offline safe reinforcement learning in a calibrated basin simulator with water-balance constraints, crop-demand states, conservative policy improvement, uncertainty penalties and historical backtesting, with results reported alongside their uncertainty rather than as point claims. Because investment and endowment often move together in Xinjiang, the design keeps physical conditions and policy-driven activity as separate inputs rather than merging them into one development score. Systematic error is reported rather than smoothed away, because clustered residuals usually point to an omitted institutional or infrastructural factor. The result is a reproducible March 2026 baseline for spring irrigation allocation as safe reinforcement learning that later work can extend without re-specifying the evidence base.

    Method notes are released to subscribers.

  2. X245XJASSMarch 2026

    Causal Representation Learning for Strategic Frame Change: Distinguishing Institutional Shifts From Source-Mix Effects

    Causal representation learning for strategic frame change is difficult to observe directly in Xinjiang: the relevant processes are spread across counties, seasons and infrastructure networks that are recorded unevenly. The article models distinguishing institutional shifts from source-mix effects, drawing on public US/UK/EU policy documents, frame labels, institution, source type, publication context, time, document-selection indicators and separate Chinese comparison material for 2021–2026 and recording provenance and missingness for each variable. The estimator is causal representation learning with institution/time/source-type treatments, balancing representations, doubly robust estimation, negative controls and sensitivity analysis, with validation structured around the way the data are actually generated rather than random splitting. The analysis distinguishes what the terrain and resource base make possible from what infrastructure and administrative decisions have actually delivered. Where the model fails, the failure is documented and located, since patterned error is often the most informative output of the exercise. Predictive performance is not treated as causal proof; the contribution is a dated, documented March 2026 benchmark for causal representation learning for strategic frame change.

    Method notes are released to subscribers.

February 20262 researches

  1. X244XJASSFebruary 2026

    Snowmelt Timing With Multimodal World Models: Forecasting Basin Transitions From Weather and Imagery

    Snowmelt timing with multimodal world models sits at the centre of how Xinjiang's interior economy is organised, linking resource endowment, infrastructure and settlement in ways that no single indicator captures. The study forecasts basin transitions from weather and imagery using seven classes of evidence — multispectral snow observations, temperature, precipitation, radiation proxies, elevation, basin attributes, runoff timing and transition labels — assembled for 2022–2026 and linked throughout to collection date and source. Analysis uses multimodal latent world model combining satellite snow states and meteorological sequences, probabilistic rollout, basin transfer, counterfactual weather perturbations and calibration against simpler sequence models; holdouts follow the geographic, temporal or institutional structure of the evidence so that performance is not inflated by leakage. Because investment and endowment often move together in Xinjiang, the design keeps physical conditions and policy-driven activity as separate inputs rather than merging them into one development score. Disagreement between specifications is preserved instead of resolved, as it marks the boundary of what the evidence can currently support. The output is a February 2026 reference point for snowmelt timing with multimodal world models, defined tightly enough that subsequent periods can be compared against it.

    Method notes are released to subscribers.

  2. X243XJASSFebruary 2026

    Permission-Aware Research Agents: Enforcing Source and Tool Boundaries in Xinjiang Policy Analysis

    By February 2026, permission-aware research agents had become one of the clearest expressions of how investment, terrain and administrative geography interact across Xinjiang. This research examines enforcing source and tool boundaries in Xinjiang policy analysis across a purpose-built evidence base covering 2023–2026: bounded questions, permitted/forbidden source partitions, tool schemas, agent actions, violation labels, gold evidence, final answers and institution metadata. The modelling strategy is policy-constrained research agent with typed permissions, pre-action authorization checks, evidence-store isolation, violation traps, utility-versus-compliance evaluation and adversarial task tests, checked against simpler baselines to confirm that the added complexity earns its place. The analysis distinguishes what the terrain and resource base make possible from what infrastructure and administrative decisions have actually delivered. Residuals are treated as evidence: the cases the model misses tend to mark infrastructure openings, regulatory changes or reporting shifts that the covariates do not contain. Taken together, the study provides a bounded February 2026 account of permission-aware research agents and an evidence architecture that later batches can reuse.

    Method notes are released to subscribers.

January 20262 researches

  1. X242XJASSJanuary 2026

    Cold-Spell Recovery of Urban Utilities: Multisystem Resilience Modelling Across Xinjiang Cities

    Questions about cold-spell recovery of urban utilities in Xinjiang turn quickly into questions about distance, water, energy and the corridors that connect the region to Central Asia. Working from 2022–2026, the study models multisystem resilience modelling across Xinjiang cities through an evidence base combining aggregated electricity/heating/water service states, outages, recovery times, temperature, wind, demand, maintenance and city infrastructure attributes. Estimation relies on multilayer temporal graph network with utility-specific nodes, shared weather shocks, recovery-state embeddings, survival heads, city-held-out testing and calibrated recovery-time distributions, with results reported alongside their uncertainty rather than as point claims. Because investment and endowment often move together in Xinjiang, the design keeps physical conditions and policy-driven activity as separate inputs rather than merging them into one development score. Systematic error is reported rather than smoothed away, because clustered residuals usually point to an omitted institutional or infrastructural factor. The result is a reproducible January 2026 baseline for cold-spell recovery of urban utilities that later work can extend without re-specifying the evidence base.

    Method notes are released to subscribers.

  2. X241XJASSJanuary 2026

    Evidence-Graph Compilers for Xinjiang Research: Turning Agent Traces Into Auditable Claim Networks

    Evidence-graph compilers for Xinjiang research is difficult to observe directly in Xinjiang: the relevant processes are spread across counties, seasons and infrastructure networks that are recorded unevenly. The article examines turning agent traces into auditable claim networks, drawing on bounded research tasks, agent trajectories, public US/UK/EU sources, atomic claims, citations, contradiction links, replay logs and separate Chinese comparison sources for 2023–2026 and recording provenance and missingness for each variable. The estimator is trace-to-graph compiler using structured claim extraction, evidence-link verification, provenance normalization, contradiction edges, graph completeness metrics and independent replay evaluation, with validation structured around the way the data are actually generated rather than random splitting. The analysis distinguishes what the terrain and resource base make possible from what infrastructure and administrative decisions have actually delivered. Where the model fails, the failure is documented and located, since patterned error is often the most informative output of the exercise. Predictive performance is not treated as causal proof; the contribution is a dated, documented January 2026 benchmark for evidence-graph compilers for Xinjiang research.

    Method notes are released to subscribers.