Research

Papers from Scaledex Research. Each one is a bounded synthesis of retained public sources that states what those sources can and cannot establish, published with its references, its provenance, and the IEEE PDF and TeX source.

Scaledex Research

Cloudflare Workflows Lifecycle and CPU-Cost Controls: Documented Boundaries for Idle Time, Instance Deletion, and Workflow Deletion

A documentation-review synthesis of Cloudflare Workflows as documented on 2026-09-01. Cloudflare counts active processing, not elapsed time, as CPU time: API-response waits, step.sleep pauses and idle periods incur none, and network, storage and other I/O waits count toward neither CPU time nor Workflows compute. They cover only the documented CPU and compute dimensions. Deleting instances frees storage, with limit updates possibly taking a few minutes; Wrangler's workflows delete removes a Workflow and its own instances; retained-data erasure is not specified.

  • Cloudflare Workflows
  • CPU accounting
  • deletion semantics
  • workflow lifecycle
  • evidence synthesis

Scaledex Research

Four Conditional Governance Controls Relevant to Commissioning and Subsequent Delivery of Business-Critical Custom Software: A Bounded Public-Source Synthesis for SMEs

A bounded public-source synthesis of the governance controls an SME can supportably consider when commissioning business-critical custom software. An independent audit retained four conditional controls: formal change control; named security-program accountability with executive oversight; pre-contract, risk-proportionate supplier due diligence; and, for high-risk supplier arrangements, contractual definition of scope, responsibilities, assets, security expectations and incident notification. The record does not establish product-owner authority, engineering assurance, universal go-live blockers or validated delay rules, so blocker decisions remain provisional.

  • custom software commissioning
  • SME governance
  • supplier due diligence
  • change control
  • security accountability

Scaledex Research

What Public Benchmark Evidence Can Establish About Bangla-Specific and Multilingual Large Language Models

A bounded synthesis of the retained public-study record on Bangla-specific and multilingual large language models across factuality, instruction following and Bangla-English code-mixed evaluation. The retained BenHalluEval evidence does not establish that Bengali-centric pretraining is necessary or sufficient for calibration, and it evaluates judgments against supplied context rather than open-world factual accuracy. No usable direct comparison exists for Bangla instruction following, and two audit limitations, an unreconciled model roster and unverified code-mixing, keep the conclusion study-bounded.

  • Bangla language models
  • multilingual large language models
  • hallucination calibration
  • instruction following
  • code-mixed evaluation