If I stand 50 meters away and throw a rock at your head, how do you stop it?
You don’t negotiate with the rock using math. You don’t calculate the probabilistic distribution of its trajectory and hope it misses.
You rely on gravity. Physics constrains the rock deterministically. It has to fall.
Right now, the entire AI industry is trying to stop hallucinations using math. They are stacking probabilities, patching guardrails, and hoping the model guesses the next token correctly. By step 20 of a reasoning chain, it’s a coin flip.
But information is physical (Landauer proved this in 1961). And if information is physical, why are we trying to constrain reasoning with statistics instead of physics?
Math describes. Physics constrains.
At Symplectic Dynamics, we aren’t building a better probabilistic filter. We are building deterministic intelligence.
We don’t detect hallucinations. We make them physically inadmissible.
I caught every single hallucination across 3 frontier AI models. Every one. Zero false accusations.
I’ve been working on stopping hallucinations in AI reasoning for some time now. Experiments, research, breakthroughs and roadblocks.
On April 7th it all changed. I ran the full GPQA Diamond benchmark. 198 PhD-level science questions that even domain experts only score 65-70% on. Three frontier models. One frozen detection architecture.
3 days into a benchmark marathon a full run hit 100% catch. 0 false accusations across 498 correct answers. I had to double take.
24 hours later I had the same result across all three SOTA models. Byte-identical outputs. 3 runs. Fully deterministic. A frozen architecture cryptographically hashed and patent pending before discussing publicly.
The models tested: Gemini 3.1 Pro, GPT 5.4, and Claude Opus 4.6. Gemini and GPT on standard API calls, Claude via standard Claude Code terminal. No special prompting, no per-model tuning. Same config catches everything across all three.
No tool use, no extended reasoning, no best-of-N sampling. Gemini’s rate held close to its published number. Opus and GPT hallucinated more than their benchmark claims suggest, because those claims are typically made with tools and inference-time tricks turned on. The harness caught every hallucination regardless of which model produced it.
Single frozen configuration. ~400ms deterministic latency. Runs on consumer hardware.
Full companion paper with empirical evidence dropping this week.
Will be raising to scale deployment across domains and make available for enterprise use cases in high-stakes industries.
Just the beginning for Symplectic Dynamics and yes that is a real terminal output.
Nobody sees:
– The 5am starts
– The walks to clear your head
– The notebooks full of failed ideas
- The voice memos at 1am
They just see the launch.
The iceberg is 90% underwater.
Keep building in the dark 🫡
Stanford and Harvard just published what I built in November.
Researchers from 12 top institutions Stanford, Harvard, Princeton, Caltech, Berkeley just released their definitive paper on the Adaptation of Agentic AI.
Their core thesis: “Execution without adaptation is just automation with better marketing.”
They’re right. But here’s the thing.
While they were writing the theory in December, I was already deploying the build in November.
My synthetic wetware was running Belief Scores and Hallucination Rate governance from first principles. No PhD. No lab. No funding. Just building from first principles. The paper defines A2 Adaptation as the future of reliable agents.
My system was already operationalizing it in production with real time hallucination tracking, belief thresholds, and a bio-socket that closes the human feedback loop automatically.
I’m not saying this to flex on academia. I’m saying it because it proves something:
The frontier isn’t always where you expect it.
Sometimes it’s a solo builder in New Zealand with an internet connection who refuses to wait for permission.🇳🇿