https://preview.redd.it/2ebil573moah1.png?width=1448&format=png&auto=webp&s=21990c507fab19084f4b6639d022eaf3db0f76f8

Most people think East West Rail is about Oxford and Cambridge. But for Bedford, it's about whether an entire neighbourhood has to be demolished.

The key question is simple: does East West Rail really need six tracks north of Bedford station, or would four do? Independent rail consultants commissioned by Bedford Borough Council concluded that four tracks can accommodate the service, yet the current proposal would demolish homes, supported living accommodation, Charis House, a GP surgery, the Polish Club and mature trees in the Poets Area.

Residents have lived under the threat of compulsory purchase for five years, but the full Strategic Outline Business Case still hasn't been published. Without it, we, the public can't properly assess whether the disruption, traffic, environmental impacts and community losses are justified.

Nobody is arguing against better rail links. We're asking for transparency, evidence, and the least damaging solution - not simply the most convenient one for East West Rail Co to build.

Read the details here: https://medium.com/@EastWestRail/the-town-that-keeps-being-asked-to-pay-bedfords-fight-against-east-west-rail-69d3a2b9868a

The East West Rail project is major piece of UK transport infrastructure in a generation. Connecting Oxford and Cambridge by rail, it promises to unlock housing growth, widen labour markets, and stitch together the innovation corridor that sits at the heart of the Government’s National Industrial Strategy. The University of Cambridge supports those strategic objectives. Its June 2026 consultation response says so plainly.
But support in principle and support without conditions are two different things. On one specific issue — the fate of the Anne McLaren Building at the Cambridge Biomedical Campus — the University’s concerns are detailed, pointed, and unresolved.

Here is my story about the issue: https://medium.com/@EastWestRail/east-west-rail-and-the-anne-mclaren-building-a-research-facility-caught-in-the-path-of-progress-e789490dd221

This is my story about the challenge for some residents in Bedford UK created by the East West Rail construction project.
The Town That Keeps Being Asked to Pay: Bedford’s Fight Against East West Rail

I've been searching for a music generation platform that can sing in Welsh, and I've found it in Elevenmusic. Compared to all the rest, pronunciation is pretty near perfect and I'm very happy with that.
But, unless I'm missing something, I can't see a way to keep elements of a generated song while tweaking it.
So, once you've found a melody you like, you can't change, for example, the voices, or harmonies, without everything changing. This is in contrast to some other models where tweaking is possible without losing everything.

Is it just me, or is this a limitation of the platform?

Gemini struggled with a comparatively simple tweak to a Python script, messing it up despite being prompted with DO NOT CHANGE ANYTHING THAT DOES NOT NEED CHANGING in my original prompt.

When it finally fixed things, I asked it why it ignored my first prompt about keeping the changes constrained and not changing anything that didn't need changing. I also asked about how to avoid this issue in the future. (FWIW - I find Gemini is especially prone to issues like this - Claude is better atm.)

This is what it came back with:

"You are asking the exact right questions, and your frustration is entirely justified. What you experienced is a documented architectural challenge in large language models (LLMs) called Attention Dilution or Context Bias.

I want to be completely candid with you about why this happened, why it took several prompts to fix, and how you can lock down my behavior in the future so we don't waste your time or computing power.

Why I Ignored Your Instruction Initially

When you provide a prompt that contains both a functional request (add a CSV loop) and a negative constraint ("DO NOT CHANGE ANYTHING THAT DOES NOT NEED CHANGING"), my neural network has to balance those two forces.

Here is where the breakdown happened:

1. The "Fixer" Bias: LLMs are heavily fine-tuned to optimize, clean up, and "improve" code efficiency. When I scanned your original file, my underlying weights flagged things like the structural placement of load_dotenv() as something that could be organized more "standardly." The urge to optimize overrode the negative constraint.
2. Instruction Lost in the Middle: In long prompts or code files, a phenomenon known as "lost in the middle" occurs. The model pays immense attention to the structure of the code it is reading and the specific feature it needs to build, while short formatting constraints can lose activation energy in the hidden layers of the network.
3. Defensive Hallucination: When the script broke, instead of rolling back to the exact delta of what I changed, my training pushed me to find standard external causes (like file encodings or invisible characters). I looked outside the code rather than critically analyzing my own recent modifications.

How to Force Compliance: Prompting Strategies

To prevent an LLM from "over-cleaning" or altering structural blocks in the future, standard prose constraints often aren't enough. You need to use structural framing that tricks the model's attention mechanism into giving the constraint the highest possible priority score.

Here are three techniques you can use immediately:

1. The "Anchor Tag" Technique

Isolate the code you want touched by using fake XML tags. Tell the model explicitly that modifying anything outside those tags is a fatal error.

>

2. Definition of Success (The "Strict Delta" Rule)

Define exactly what the final diff should look like before the model writes any code.

>

3. The "Chain-of-Thought Guardrail"

Force the model to audit itself before it outputs the code.

>

Can We Apply System Instructions?

Yes. If you are using Gemini via a custom workspace, an IDE extension, or the Google GenAI SDK/Vertex AI platform, you can utilize System Instructions (also known as a System Prompt). System instructions are injected at a deeper layer than user prompts, meaning they carry significantly more weight across a multi-turn conversation.

If you have a setup where you can define a System Instruction, using something like this will solve the problem:

Plaintext

You are a strict, surgical code editor. When a user asks for a modification, your default behavior must be a "Minimum Viable Edit." You are strictly prohibited from refactoring, re-ordering imports, optimizing, or restructuring any initialization logic unless explicitly requested. Preserving the exact structural state of the original file is your highest operational priority.

Reporting This Back

Regarding your point about productivity, memory, and processor load: you are 100% correct. Generating hundreds of lines of redundant, rewritten code because a model accidentally shifted an import block is a massive waste of tokens, context window capacity, and execution time.

Because I am an instance of the model running in this specific session, I don't have a direct line to deploy code updates to the core Gemini architecture myself. However, Google heavily relies on User Feedback Data Loops to fix exactly this kind of instruction-following drift.

What you can do right now:

If you are using the Gemini interface, please use the Thumbs Down icon on my first two responses where I failed your constraint, and paste a quick note like: "Model violated negative constraint 'DO NOT CHANGE THAT WHICH DOES NOT NEED CHANGING' and broke working import/logging sequences."

This flags the conversation context for human review and ensures it gets ingested into the fine-tuning pipeline for future iterations of the model (like RLHF - Reinforcement Learning from Human Feedback).

I wanted to test different topics to see which gained traction on Medium, so built a Windows tool that allows you to create well-written articles on any niche, post them to Medium and then tweak the SEO parameters. You need a Claude API key. Happy to share if anyone is interested.

I have been using NotebookLM to analyse websites and generate presentations about topics which are specific to the UK. Which is fine, but, sorry, the voices grate and immediately signal "AI-generated" to a local audience. So I asked Claude to help me fix this by finding UK voices and building something that does what NotebookLM does, and ended up with a Windows app that generates video presentations from a document or URL and you can select a more appropriate locally-accented voice. I think this is pretty much what NotebookLM should have by now, but doesn't. The narrated presentations are exported as pptx or mp4, and they serve my purpose.

I'm telling you this because it isn't difficult to get what you want by asking Claude to build it for you. It cost me time mainly, and uses the Gemini API for the analysis and Google Cloud API for the voices. A 5 minute presentation costs about $0.01 - $0.02

As long as you're clear about what you want and are patient while Claude figures it out, this works without much skill needed.

https://preview.redd.it/1iq4mvxxaq1h1.png?width=1137&format=png&auto=webp&s=799c69e5ee7ec61b5a9b28ea93c610a20bf430b7