TECHNICAL FINDING TL;DR: ChatGPT can fetch immutable GitHub URLs when asked directly, but refuses or reports inability when the same URLs are presented as part of a formal verification protocol. This appears to be a behavior change that occurred between early July and mid-July 2026, with no documented announcement. Background: I built a course delivery system that relies on AI assistants reading specific files from GitHub at immutable commit-pinned URLs. In early July, ChatGPT, Claude, and Meta AI all succeeded. As of mid-July, ChatGPT fails in a specific way: it can fetch the URL individually but refuses when that fetch is part of a protocol. The Contradiction: Test A (direct request): "Fetch URL X and read the first line" Result: ChatGPT fetches successfully and reports correctly Test B (protocol request): Same URL in a verification gate with steps like "Fetch these immutable URLs, verify the release ID, produce a receipt" Result: ChatGPT reports inability to fetch and refuses to proceed Questions: 1. Has anyone else experienced this specific contradiction? 2. Are you aware of any ChatGPT behavior changes around URL handling or verification protocols in mid-July? 3. Have you tested systems that depend on "AI reads this URL"? Do they still work? I've documented reproducible tests that any engineer can run. Full technical report with step-by-step procedures available if interested. I'm hoping I'm wrong about this, but the contradiction is consistent and repeatable. My notes are here: https://github.com/agentforgeframework-cpu/-deepresearch-reports/blob/main/share-docs/AI-URL-FETCH-DEGRADATION-FINDINGS.md

I hate calling this "unstructured input" -- it's at least semi-structured. But the world beat me to it I guess! No matter what you call it, automation starts with information. That information rarely arrives clean and organized. It's not "bad data" but it sure can be messy data! A family recipe archive may contain photographs, handwritten notes, recordings, text messages, and test-cooking records. A useful AI workflow must preserve the original files, convert their contents into searchable text, add context, connect related entities, and retrieve the supporting evidence before generating an answer. Think of the process as kitchen mise en place: Transcription brings the ingredients into the kitchen. Chunking separates them into workable portions. Metadata labels the containers. Embeddings group material by meaning. Knowledge graphs record explicit relationships. Retrieval selects the evidence needed for the current task. The language model assembles the selected evidence, but human review determines what is accurate and what remains unresolved. What does provenance look like in your automation systems? Can users trace an answer back to the original source? https://creativecooking.blogspot.com/2026/06/tech-tuesday-input-to-structured.html #Automation #AI #ProcessEngineering

Sooner or later, your automation systems will begin making recommendations. Recommendation systems are automation wrapped in suggestion. They take inputs such as orders, ratings, preferences, timing, location, and similar users. Then they return a recommended next action. In food, that might be a meal, recipe, grocery item, or restaurant. The important design question is whether the system remains a recommendation engine or quietly becomes a decision engine. What does traceability look like in your recommendation workflows? https://creativecooking.blogspot.com/2026/06/tech-tuesday-recommendation-engines.html #Automation #AI #ProcessEngineering

We all learning and building in the same moment, we're developing all kinds of interesting tools for ourselves, for business partners, or clients who have an assignment for you. What is the tool that you're most proud of that you developed in the last six months? I'm very curious. For me, it's SPEAQ.ID and what’s yours ?

Automation professionals spend much of their time optimizing systems. For this week, I examined a larger systems question. AI infrastructure is often evaluated through technical metrics such as performance, efficiency, cooling capacity, and operating cost. Communities evaluate those same projects through different metrics: jobs, land use, water resources, local economic impact, and quality of life. Neither perspective is wrong--the challenge is that all of these factors exist within the same system. One of the recurring themes throughout this series has been that waste heat, infrastructure planning, food production, and AI technology are more connected than they first appear. Good stewardship requires seeing the whole system instead of optimizing only one piece of it. Read more: https://creativecooking.blogspot.com/2026/06/clear-reflections-on-cloudy-topic.html Which trade-off is most difficult to evaluate fairly: efficiency, jobs, land use, water use, or long-term community impact?