Hi everyone, Sharing the abstract below for feedback before I take this paper further. I'd especially welcome blunt reactions on whether it reads as a genuine prior layer or just a relabeling of what existing frameworks already do — and whether the opening lands on a first read. Abstract Before evidence can be weighed, it must first be constituted as an evidential object. Every evidential evaluation rests on three components: a focal claim, the observations admissible as its instances, and what each instance contributes. Specifying them jointly — a prior methodological step I call evidential instantiation — is typically left implicit.I call their joint specification — a prior methodological step — evidential instantiation. This step is typically left implicit. Examined through this tripartite structure, seven recurrent methodological difficulties — among them pseudoreplication, conflicting conclusions from the same data, and incommensurable meta-analyses — become intelligible as manifestations of a common structure. The same structure provides a common basis for comparing established methodological frameworks: because they were developed for different methodological tasks, each makes some components explicit while leaving others implicit or presupposed — rarely the whole. The present account specifies that step, complementing these frameworks. I provide an initial formalization of this tripartite structure, making all three components explicit and inspectable at once, in terms that hold across domains. Making them explicit does not guarantee correct evaluation; it makes recurrent evidential difficulties systematically locatable — and thus diagnosable — in the claim, the admissible instances, or their contribution. This is not a calibrated instrument. It specifies the evidential object that existing inferential frameworks presuppose before evidence can be weighed, leaving domain-specific calibration to future work. Keywords: evidential instantiation; metascience; research methodology; measurement; evidence synthesis; reproducibility

Hi Community, I would like to test a talk topic with you for an upcoming event later this year. The deadline for the Call for Papers (CFP) closes tomorrow: https://compute.events/paris2026/cfp.html How does the proposal read to you? Is there an angle I missed, and would you want to spend 30 minutes listening to this topic? As with all abstract submissions, it needs to be short and to the point to get the message across to a reviewer in 30 seconds. Your feedback would be incredibly valuable in helping me tweak the text and land an invitation to speak! Thanks in advance for your thoughts and feedback! Gijs ============================================ Title: Navigating the Vanguard: A Practical Guide to Selecting Geospatial Foundation Models Brief Summary (Abstract) The recent explosion of Earth Observation Foundation Models (EO-FMs), such as AlphaEarth, TerraMind, and AnySat, has created unprecedented capabilities, but also severe "model fatigue". This talk provides a practical guide for data practitioners to make model selection easier and more transparent. We will map the current landscape of GeoAI models, demonstrate how to evaluate their predictive power using accessible open-source tools, and share a modular pipeline architecture to scale processing on standard cloud GPUs without requiring a high-performance computing (HPC) cluster (for testing the models' local accuracy). Finally, we will unpack the critical "semantic and temporal cautions" to ensure attendees understand the hidden uncertainties in their geospatial embeddings before making operational decisions, because all models are built for a specific purpose, and while they are called “Foundation Models”, there is not one model that fits all use cases - project domains are often very specific, and will need special tuned models to produce a solution. Description Objective & Central Thesis: The geospatial domain is currently experiencing a vanguard of multi-modal, general-purpose foundation models. The central thesis of this talk is that successfully navigating this landscape requires a structured approach to selecting and evaluating models, rather than treating them as magic black boxes. This practical guide aims to simplify model selection while emphasising scalable compute and rigorous uncertainty quantification.

You paste in your drafts and ask it to "make it better." That's not using AI. That's outsourcing your thinking (and slowly handing your job to a machine.) Here's how I think about it differently: AI is my writing advisor and sounding board I can talk to 24/7. Not my ghostwriter. When I use AI in my writing, I'm not asking it to replace what I do. I'm using it to: → challenge my arguments before a reviewer does → pressure-test my logic when I'm too close to the work → help me see my own weaknesses faster The goal is to come out of that process with better quality output, not to produce something I couldn't have written myself. Here's the uncomfortable truth: If AI can fully replicate your academic voice, your reasoning, your expertise... what exactly are you bringing to the table? Your unique value isn't your ability to write sentences. It's the decades of domain knowledge, the hard-won intuition, the ability to ask questions no one else is asking. AI should amplify that. Not replace it. Now, as a lifelong learner, I'm curious how you are using AI in your academic writing or research workflow?

Tell me in the comments or send me a direct message.

Your paper comes back rejected after a year of work. The review isn't brutal. It's vague - "The contribution isn't clear." No fatal flaw in the methods. Nothing to point at and fix. Just: unclear. I've seen it in nearly every field I've supervised, examined and mentored in - engineering, social science, medicine. The surface details change; the rejection sentence barely does. The science is rarely the problem. It's a framing problem wearing a rejection letter. Here's what most people don't realise about the other side of the submission portal: Between a 30% and 70% of papers are rejected before peer review even starts. And the single most common reason isn't bad methods — it's "we can't see what's new here." That's because reviewers don't read your paper the way you wrote it. They skim. A reviewer forms their initial opinion from the abstract and the first figure, and the rest is mostly confirming that first impression. So your contribution has to survive such a 90-second skim. Three ways to make sure it does: 1. Write your one-sentence contribution first. "This paper shows that ___." One clause. If you need four, you haven't chosen yet. And no, reviewer will choose for you. 2. Plant that sentence on the skim path. The four places a reviewer looks first: abstract, last paragraph of the introduction, the figures, the conclusion. Your one idea, in plain words, in all four. Most papers state it once and imply it three times. That reads as "unclear." 3. Show the gap, don't assert it. "Little research exists on this" is a sentence reviewers distrust. A comparison table (prior studies down the side, the open column as your gap) does the work very well. None of this gets weak work past good reviewers. If the science is broken, fix the science. But for the pile of solid papers bounced for "limited novelty," the work is already done. You've just buried the point. That's a writing problem. Which is the good kind that you can fix in an afternoon. What's the one sentence your last paper was trying to say? Try writing it in the comments. It's harder than it looks.