We Built What Anthropic Just Published—Two Months Ago
It started with a blink. Then a pause. Then that familiar mix of validation and quiet disbelief.
Scrolling through Google News last night, I came across an article on The Decoder, featuring a headline about Anthropic’s new method for training AI models using something called Internal Coherence Maximization (ICM). This technique enables language models to teach themselves, without relying on external human labels.
And I thought: “Wait a minute. We’ve been building that into our agent system for the past two months.”
This isn’t a humblebrag. It’s a testament to what happens when you build in the open, design from first principles, and trust the internal logic of your product to evolve alongside cutting-edge research, sometimes even ahead of it.
The Power of Self-Tuning Systems
ICM is built on a simple yet powerful idea: what if models could determine the quality of their answers based on internal consistency, rather than relying on a human to label them?
Anthropic’s researchers defined two key principles:
- Mutual Predictability – Can the model reliably predict answers based on its previous reasoning?
- Logical Consistency – Do those answers contradict each other?
These are the very same principles that we’ve embedded into our system design, from our MetaScorer Agent to our role maturity tracking to how agents handle review queues.
Instead of relying on human supervision to validate output, our system cross-references role outputs, detects internal contradictions, and routes tasks based on coherence, not gut instinct. It’s all part of the Edge framework we’ve been refining to power tools like TrustSignal, RiseKit, and InvestWise-AI.
How We Built ICM—Before We Knew Its Name
We didn’t call it ICM. We just knew our agents needed to evaluate each other, grow more intelligent over time, and stop shipping conflicting or incoherent outputs.
So we built:
- A MetaScorer Agent to audit answers across labs and assign alignment scores.
- A Role Maturity Tracker to weigh decisions based on historical trust and performance.
- A Review Queue System that routes contradictory or incoherent responses to reprocessing, rather than escalation.
The result? Our agents don’t just work—they evolve. And the evolution is based entirely on internal coherence, not labeled data.
This system now powers:
- Off-page SEO message scoring in TrustSignal
- Strategic planning simulations in InvestWise-AI
- Behavioral funnel alignment in RiseKit
And all of it’s driven by agent self-evaluation and reward refinement, without human judgment calls slowing things down.
Why This Matters for Solo Founders
Reading that article from Anthropic wasn’t intimidating—it was energizing. Because it proved something we’ve believed since day one:
You don’t need a research lab to build systems that think for themselves.
If you understand logic, consistency, and behavior-driven reward systems, you can start building self-tuning AI today. Not as a paper. Not as a prototype. As a working production system that makes your SaaS smarter with every run.
And if you're building solo? Even better. These systems work best when they're lean, interpretable, and directly connected to how your product grows.
What’s Next: Numbers, Papers, and Public Evolution
We’re not done. With our first round of AI agent evaluations underway, we’re about to start measuring:
- Internal coherence scoring accuracy
- Contradiction frequency across roles
- Maturity progression and self-resolution success rates
Those numbers will shape our upcoming research paper. But right now, the most crucial part is this:
The future of AI is already here—and you can build it too.
We’ll continue to publish as we progress, illustrating how these concepts evolve through real-world product use. If you're curious about agent workflows, scoring systems, or building your own ICM logic into practical tools, follow along.
Let’s build AI that continually improves—and do it in public.