Our Work

Added added circle check to energy flow diagram in myems-api and myems-admin added duplicate check for binding meter to space in myems-api and myems-admin added duplicate check for binding meter to s…

Added added circle check to energy flow diagram in myems-api and myems-admin added duplicate check for binding meter to space in myems-api and myems-admin added duplicate check for binding meter to shopfloor in myems-api and myems-admin added duplicate check for binding meter to combined equipment in myems-api and myems-admin added duplicate check for binding meter to equipment in myems-api and myems-admin added duplicate check for binding meter to store in myems-api and myems-admin added duplicate check for binding meter to tenant in myems-api and myems-admin added myems_production_db.tbl_equipment_hourly to database added modification protection for super administrator in myems-api and myems-admin added yAxisScale attribute to the MultipleLineChart component to control the scale configuration of the ECharts Y-axis in myems-web added ID to equipment list in combined equipment reports added space dashboard to myems-api and myems-web added min/max/avg to trend chart in myems-web Changed changed component import method from static to dynamic imports in myems-web added ErrorBoundary component to capture rendering errors in myems-web upgraded echarts version in myems-web updated meter realtime report in myems-web updated deepseek api url and model Fixed fixed Overwritten property issue in myems-web fixed datetime picker color theme issue in myems-web fixed parameters_data issues in myems-api fixed issues of combinedequipmentdashboard, equipmentdashboard, shopfloordashboard, storedashboard and tenantdashboard in myems-api fixed issue of combined equipment load excel exporter in myems-api Removed removed print(req.params) from myems-api

Orchestrating workloads across heterogeneous multi-cloud infrastructure requires placement policies that trade accelerator throughput, monetary cost, data locality and service-level objectives&nb…

Orchestrating workloads across heterogeneous multi-cloud infrastructure requires placement policies that trade accelerator throughput, monetary cost, data locality and service-level objectives against one another, and that keep doing so as the workload mix drifts. Hand-tuned heuristics are static; reinforcement learning needs large interaction budgets and generalises poorly off-distribution; and using a large language model (LLM) as the online scheduler is economically impossible, because a production cluster takes 104–105 placement decisions per hour and each decision would cost an LLM invocation. We present AutoOrch-LLM, a two-timescale closed loop that removes the LLM from the decision path entirely: a fast loop executes a compiled symbolic policy in microseconds per candidate, while a slow loop invokes an LLM a handful of times per deployment to rewrite that policy. Three technical components make the loop work. (i) A Diagnostic Residual Report (DRR) converts a scalar objective into dense causal feedback by attributing per-decision counterfactual regret—against the best feasible alternative that existed at that instant — to six interpretable failure modes with quantified objective shares. (ii) A feasibility projection filters candidate sets by hard constraints before any learned score is consulted, which makes constraint satisfaction independent of LLM correctness. (iii) A risk gate admits a proposed policy only when a paired stationary bootstrap on held-out trace windows rejects the no-improvement hypothesis, and a structural trust region on abstract syntax tree edit distance bounds how far one round may move. Evaluated on the SHA-256-verified Alibaba PAI GPU cluster trace (1,261,050 task records, 1,897 heterogeneous machines, six accelerator classes) mapped onto a four-provider multi-cloud testbed, AutoOrch-LLM reaches J = 0.5333±0.0052 on eight held-out windows using 60 window simulations and only 3 LLM calls, against 0.5471 for a matched random search that consumes 168 simulations and 0.5953 for a strong hand-tuned expert policy. Removing the DRR degrades the objective by +29.2%. Under fault injection, eight adversarial or degenerate score functions produce 0 hard constraint violations, and the static gate rejects 10/10 code-injection attempts.

We introduce Semantic Compression Work: the drop in the next-token Shannon entropy of a language model across a token sequence. Each sequence either compresses the model's uncertainty (a funnel), leav…

We introduce Semantic Compression Work: the drop in the next-token Shannon entropy of a language model across a token sequence. Each sequence either compresses the model's uncertainty (a funnel), leaves it approximately unchanged (equilibrium), or expands it (a fan-out). The framework treats the model's probability space not as a black box producing outputs, but as a measurable system whose uncertainty evolves sequence by sequence.

Automated release bot to push zenodo

deepisles_1.1_20260715

Background & Summary, condensed; state plainly that 124 IMF-sourced indicators are not redistributed and are regenerated by the reconstruction script

 This article presents a comparative pragmatic and linguistic analysis of compliment expressions in three typologically diverse languages: Russian, English, and Uzbek. Compliments, as positively …

 This article presents a comparative pragmatic and linguistic analysis of compliment expressions in three typologically diverse languages: Russian, English, and Uzbek. Compliments, as positively evaluative speech acts, serve as powerful tools of social bonding, politeness, and identity negotiation across cultures. Drawing on the theoretical frameworks of Austin (1962), Searle (1969), Brown and Levinson (1987), and Wolfson (1983), the study examines the structural, lexical, and semantic features of compliments in each language, identifying both universal tendencies and culture-specific patterns. The findings reveal that English compliments are typically formulaic and syntactically fixed, focusing on individual achievement and appearance; Russian compliments show greater grammatical flexibility and emotional depth owing to the inflectional morphology of the language; and Uzbek compliments reflect collectivist cultural values, age- and status-based hierarchies, and a preference for indirect or metaphorical praise. The article also considers how gender, social distance, and communicative context shape complimenting behavior in each linguistic community. The results have implications for intercultural pragmatics, language pedagogy, and cross-cultural communication research.

Data files to run PM6 and Y6 MD simulations