TBASS research record
January–March 2021 to June 2026
The digital record of this branch, listed from the most recent entry backwards.
June 20262 researches
- T132TBASSJune 2026
Strategic Representation With Provenance Guarantees: Auditing US, UK and EU Tibet Synthesis at the End of the Series
By June 2026, strategic representation with provenance guarantees had become measurable in Tibet largely through remote sensing and administrative records rather than dense ground observation. Working from 2023–2026, the study audits US, UK and EU Tibet synthesis at the end of the series through an evidence base combining matched US, UK and EU public policy corpora, separate Chinese/Tibetan comparison sources, immutable provenance records, agent syntheses, challenge logs, citations and expert frame labels. Estimation relies on institution-specialist agents with immutable source provenance, claim-evidence graphs, cross-agent challenge rounds, disagreement-preservation scoring, citation verification and expert comparative adjudication, with results reported alongside their uncertainty rather than as point claims. Sparse and uneven observation is modelled explicitly, because in Tibet the absence of a measurement is frequently informative about access rather than about the phenomenon. 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 June 2026 baseline for strategic representation with provenance guarantees that later work can extend without re-specifying the evidence base.
Method notes are released to subscribers.
- T131TBASSJune 2026
Reproducible Deep Research for Tibetan Studies: Can Independent Agents Reconstruct the Same Evidence Trail?
Studying reproducible deep research for Tibetan studies in Tibet means working with sparse observation, extreme terrain and strong seasonality at the same time. The article examines can independent agents reconstruct the same evidence trail?, drawing on bounded research questions, versioned multilingual source pools, repeated agent runs, tool logs, evidence sets, claims, citations and reproducibility metrics for 2026 and recording provenance and missingness for each variable. The estimator is replicated deep-research agents with fixed tool permissions, stochastic-run comparison, evidence-set overlap, claim consistency, citation reproducibility and expert adjudication, with validation structured around the way the data are actually generated rather than random splitting. Elevation, slope and seasonality are treated as structural constraints rather than controls, since on the plateau they set the limits within which everything else varies. 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 June 2026 benchmark for reproducible deep research for Tibetan studies.
Method notes are released to subscribers.
May 20262 researches
- T130TBASSMay 2026
Multimodal Evidence Arbitration: Resolving Conflicts Between Plateau Imagery, Sensor Series and Written Reports
On the Tibetan Plateau, multimodal evidence arbitration is governed by altitude, hydrology and a thin, seasonal infrastructure network, so patterns visible at the county scale often disappear at the regional average. The study examines resolving conflicts between plateau imagery, sensor series and written reports using six classes of evidence — satellite imagery, environmental series, public reports, aligned events, intentionally conflicting cases, provenance metadata and expert resolutions — assembled for 2022–2026 and linked throughout to collection date and source. Analysis uses multimodal arbitration agents with modality-specific encoders, provenance-weighted evidence graphs, contradiction detection, calibrated adjudication and expert environmental review; holdouts follow the geographic, temporal or institutional structure of the evidence so that performance is not inflated by leakage. Sparse and uneven observation is modelled explicitly, because in Tibet the absence of a measurement is frequently informative about access rather than about the phenomenon. Disagreement between specifications is preserved instead of resolved, as it marks the boundary of what the evidence can currently support. The output is a May 2026 reference point for multimodal evidence arbitration, defined tightly enough that subsequent periods can be compared against it.
Method notes are released to subscribers.
- T129TBASSMay 2026
Federated Tibetan Language Adaptation: Improving Domain Models Without Centralising Institutional Text Collections
Federated Tibetan language adaptation matters on the plateau far beyond Tibet itself, because the headwaters, corridors and grid connections concerned extend downstream and across borders. This research examines improving domain models without centralising institutional text collections across a purpose-built evidence base covering 2026: partitioned licensed/open Tibetan-domain corpora, shared evaluation tasks, silo metadata, adapter updates, privacy budgets and held-out test sets. The modelling strategy is federated parameter-efficient fine-tuning with secure-aggregation simulation, domain adapters, privacy accounting, cross-silo evaluation and centralised baseline comparison, checked against simpler baselines to confirm that the added complexity earns its place. Elevation, slope and seasonality are treated as structural constraints rather than controls, since on the plateau they set the limits within which everything else varies. 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 May 2026 account of federated Tibetan language adaptation and an evidence architecture that later batches can reuse.
Method notes are released to subscribers.
April 20262 researches
- T128TBASSApril 2026
Adaptive Geospatial Agents for Plateau Hazards: Choosing Resolution, Tools and Evidence Under Compute Budgets
By April 2026, adaptive geospatial agents for plateau hazards had become measurable in Tibet largely through remote sensing and administrative records rather than dense ground observation. Working from 2021–2026, the study examines choosing resolution, tools and evidence under compute budgets through an evidence base combining satellite imagery, terrain, weather, hazard events, infrastructure layers, tool costs, agent traces, verified calculations and uncertainty labels. Estimation relies on budget-aware geospatial agent with learned tool routing, multi-resolution raster analysis, uncertainty-aware stopping, cost-quality evaluation and event-held-out testing, with results reported alongside their uncertainty rather than as point claims. Sparse and uneven observation is modelled explicitly, because in Tibet the absence of a measurement is frequently informative about access rather than about the phenomenon. 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 April 2026 baseline for adaptive geospatial agents for plateau hazards that later work can extend without re-specifying the evidence base.
Method notes are released to subscribers.
- T127TBASSApril 2026
Evidence Contracts for Tibetan Research Agents: Requiring Source Coverage Before Claims Are Released
Studying evidence contracts for Tibetan research agents in Tibet means working with sparse observation, extreme terrain and strong seasonality at the same time. The article examines requiring source coverage before claims are released, drawing on bounded multilingual research tasks, public/licensed sources, claim contracts, gold evidence, agent trajectories, blocked/released claims and evaluator labels for 2026 and recording provenance and missingness for each variable. The estimator is contract-constrained research agents with claim schemas, minimum evidence rules, citation entailment, contradiction checks, release gates and expert audit, with validation structured around the way the data are actually generated rather than random splitting. Elevation, slope and seasonality are treated as structural constraints rather than controls, since on the plateau they set the limits within which everything else varies. 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 April 2026 benchmark for evidence contracts for Tibetan research agents.
Method notes are released to subscribers.
March 20262 researches
- T126TBASSMarch 2026
Strategic Representation Under Verifiable Agents: Comparing US, UK and EU Tibet Frames With Evidence Graphs
On the Tibetan Plateau, strategic representation under verifiable agents is governed by altitude, hydrology and a thin, seasonal infrastructure network, so patterns visible at the county scale often disappear at the regional average. The study compares US, UK and EU Tibet frames with evidence graphs using six classes of evidence — matched US, UK and EU public policy corpora, separate Chinese/Tibetan comparison sources, claim graphs, citations, agent syntheses and expert-coded frames — assembled for 2023–2026 and linked throughout to collection date and source. Analysis uses institution-specialist agents with isolated retrieval, claim-level evidence graphs, comparative synthesis, disagreement-preservation metrics, counterfactual source removal and expert frame adjudication; holdouts follow the geographic, temporal or institutional structure of the evidence so that performance is not inflated by leakage. Sparse and uneven observation is modelled explicitly, because in Tibet the absence of a measurement is frequently informative about access rather than about the phenomenon. Disagreement between specifications is preserved instead of resolved, as it marks the boundary of what the evidence can currently support. The output is a March 2026 reference point for strategic representation under verifiable agents, defined tightly enough that subsequent periods can be compared against it.
Method notes are released to subscribers.
- T125TBASSMarch 2026
Collaborative Scholar Agents Across Tibetan, Chinese and English Sources: Measuring Context Loss in Delegation
Collaborative scholar agents across Tibetan, Chinese and english sources matters on the plateau far beyond Tibet itself, because the headwaters, corridors and grid connections concerned extend downstream and across borders. This research measures context loss in delegation across a purpose-built evidence base covering 2026: bounded research tasks, Tibetan/Chinese/English source pools, agent handoffs, retrieval evidence, citations, context annotations and evaluator scores. The modelling strategy is federated specialist-agent workflow with language-specific retrieval stores, structured handoffs, context-loss metrics, provenance checks, source-boundary tests and bilingual expert review, checked against simpler baselines to confirm that the added complexity earns its place. Elevation, slope and seasonality are treated as structural constraints rather than controls, since on the plateau they set the limits within which everything else varies. 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 March 2026 account of collaborative scholar agents across Tibetan, Chinese and english sources and an evidence architecture that later batches can reuse.
Method notes are released to subscribers.
February 20262 researches
- T124TBASSFebruary 2026
Causal Representation Learning for Plateau Ecology: Separating Weather, Habitat and Observation Bias
By February 2026, causal representation learning for plateau ecology had become measurable in Tibet largely through remote sensing and administrative records rather than dense ground observation. Working from 2021–2026, the study models weather, habitat and observation bias through an evidence base combining permissioned species observations, camera/acoustic effort, habitat, weather, terrain, accessibility proxies and observation-process metadata. Estimation relies on causally regularised multimodal representation learning with explicit observation-process covariates, environment interventions, site holdouts and sensitivity analysis, with results reported alongside their uncertainty rather than as point claims. Sparse and uneven observation is modelled explicitly, because in Tibet the absence of a measurement is frequently informative about access rather than about the phenomenon. 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 February 2026 baseline for causal representation learning for plateau ecology that later work can extend without re-specifying the evidence base.
Method notes are released to subscribers.
- T123TBASSFebruary 2026
Permission-Aware Archive Agents: Enforcing Collection Rules During Autonomous Tibetan Research
Studying permission-aware archive agents in Tibet means working with sparse observation, extreme terrain and strong seasonality at the same time. The article examines enforcing collection rules during autonomous Tibetan research, drawing on synthetic/licensed collection policies, bounded archive tasks, source metadata, permission states, agent tool traces, violations and compliant reference trajectories for 2026 and recording provenance and missingness for each variable. The estimator is permission-aware agentic RAG with action masks, collection-policy retrieval, access-control simulation, violation detection, safe fallback behaviour and curator/expert audit, with validation structured around the way the data are actually generated rather than random splitting. Elevation, slope and seasonality are treated as structural constraints rather than controls, since on the plateau they set the limits within which everything else varies. 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 February 2026 benchmark for permission-aware archive agents.
Method notes are released to subscribers.
January 20262 researches
- T122TBASSJanuary 2026
Multimodal World Models for Plateau Hydrology: Learning Bounded State Transitions Across Snow, Lakes and Rivers
On the Tibetan Plateau, multimodal world models for plateau hydrology is governed by altitude, hydrology and a thin, seasonal infrastructure network, so patterns visible at the county scale often disappear at the regional average. The study models bounded state transitions across snow, lakes and rivers using six classes of evidence — multispectral imagery, snow/lake masks, river networks, weather, elevation, streamflow where available and dated hydrological states — assembled for 2020–2026 and linked throughout to collection date and source. Analysis uses multimodal geospatial world model combining raster foundation embeddings, hydrological time series and graph connectivity, with rollout uncertainty and basin-held-out backtesting; holdouts follow the geographic, temporal or institutional structure of the evidence so that performance is not inflated by leakage. Sparse and uneven observation is modelled explicitly, because in Tibet the absence of a measurement is frequently informative about access rather than about the phenomenon. Disagreement between specifications is preserved instead of resolved, as it marks the boundary of what the evidence can currently support. The output is a January 2026 reference point for multimodal world models for plateau hydrology, defined tightly enough that subsequent periods can be compared against it.
Method notes are released to subscribers.
- T121TBASSJanuary 2026
Verifiable Research Agents for Tibetan Studies: Building Claim-Level Provenance Graphs Before Synthesis
Verifiable research agents for Tibetan studies matters on the plateau far beyond Tibet itself, because the headwaters, corridors and grid connections concerned extend downstream and across borders. This research examines building claim-level provenance graphs before synthesis across a purpose-built evidence base covering 2026: bounded multilingual research tasks, licensed/open sources, atomic claims, gold evidence, provenance graphs, agent trajectories and evaluator labels. The modelling strategy is planner-retriever-verifier agents with claim graphs, source entailment checks, citation completeness, contradiction flags, trajectory audits and bilingual/expert adjudication, checked against simpler baselines to confirm that the added complexity earns its place. Elevation, slope and seasonality are treated as structural constraints rather than controls, since on the plateau they set the limits within which everything else varies. 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 January 2026 account of verifiable research agents for Tibetan studies and an evidence architecture that later batches can reuse.
Method notes are released to subscribers.