Founder's Journal || Truth, Trust, and the Illusion of Authenticity

2026/5/18 GMT-4

Revised 2026/9/6 GMT-4

"Fable 5.1 and GPT 6 both just came out in the last couple of days. It feels like AGI has genuinely become something within reach — like humanity will soon be able to free its own mind and hand everything over to AI. But is that really true?" — Sep 6th

Hi, I'm Stickman Charles.

This piece started today, while I was designing the leveling system for EchoID, EchoForge's user-tier framework. My brief to the AI was clear: no steep price barrier to entry, prioritize ecosystem participation, emphasize long-term contribution, and don't let it turn into "paying for your tier." EchoForge was never meant to be a pay-to-win system where money is what moves you up. What EchoID has always cared about are the assets that can't be accumulated quickly.

Here's what the AI came back with:

• —LV1 Shadow — $0
• —LV2 Etched — 100 ADA
• —LV3 Mirror — 1,000 ADA
• —LV4 Obsidian — 5,000 ADA
• —LV5 Infinity — 10,000 ADA

It even thoughtfully did the math for me: at $20 a year, how many years it would take to level up. It looked completely professional, like the business model of a mature SaaS company. The problem is, none of it was drawn from EchoForge's core philosophy — it was generated from whichever template scores highest on "correctness."

Looking back at the last few months of building EchoForge, this kind of problem hasn't been rare. Sometimes small, like mismatched information scattered across different documents. Sometimes big, like not being sure which version of a smart contract was actually running in production. (More on that in an earlier post: https://www.echoforgeef.com/blog/echo-with-you-1.) And this is with something like ten different models in the mix during development — including ones like Opus 4.6, which was billed at the time as "the strongest on the planet." That's clearly not something you can chalk up to "AI hallucination."

Here's an example. During a full code review of EchoForge in July, I discovered that the contract EchoCert was running in production wasn't the version I thought I'd prepared. It had already been live for close to three months. For a smart contract, that's long enough to make walking it back almost impossible.

Luckily, the two contracts were functionally almost identical, or there would now be a batch of wrongly issued certificates out there.

Some of that comes down to the models back then being weaker than they are now. But more of it comes down to me.

The way I used to work with agents was closer to this: write good instructions, switch on auto mode, and sit back while the agent handed me finished code and files — and just like that, a demo was done. It was actually shipping that took the most effort. Was there a step missing in between? There was: checking and testing my own code and documentation. At the very least, I should have known exactly which version of the code was running.

Once that alarm went off, I built two or three skills specifically to tell the AI what it could and couldn't do. For instance: before any irreversible action, stop and let me carry it out by hand. After finishing each module, test it on the spot, and only move on once the feature works and the documentation is clear.

"What if this is just Charles being new to it — not yet knowing how to use AI well?"

During a stretch of time I spent at Hizz, someone who'd worked as a developer for many years and now builds his own studio with AI told me that testing matters more now, not less — because people are quick to trust agents blindly, and that puts live code at risk of fatal errors. Someone else, who's been running things at Hizz for over half a year, said the same: you test and retest before you dare to actually ship. The more a piece touches money, the more careful you need to be.

Early on, not having a firm grip on your own project doesn't look like much of a problem. But it compounds. The further along you get, the bigger the snowball, and the more serious the consequences.

Why do Claude Code and Codex always need you to plan out the next step and make the call? At bottom, it still has to be you — the person who actually understands the project — carrying that responsibility.

Like the earlier version of this piece said: in the age of AI, the thing that matters most is trust. And where does trust come from? From how well you understand your own project. Only once you've tested it, understood it, and thought it through yourself does a project truly become yours — not one of your agents'.

Keep thinking, and take care.

Published September 6, 2026 · EchoForge · FOUNDER'S JOURNAL