Artificial intelligence, as we know it, has a problem. Most large language models produce different answers to the same question, even under apparently identical conditions. That inconsistency isn’t a bug – it’s a feature of probability. But when AI is used for risk assessment or decision support, that variability becomes an issue.
There may, however, be a way forward. A new research and engineering framework called Decision Physics claims to show that AI can operate deterministically – producing the same output every time, with full lineage and zero drift.
Developed by Martin Lucas, Chief Innovation Officer at Matrix OS and TheaHQ, the system replaces probabilistic inference with deterministic computation. In a live reproducibility test, 1,000 identical inputs reportedly produced 1,000 identical outputs, each verified through audit-grade receipts.
The findings, published in The 30 Per Cent Problem: The World Is Arguing with Its Own Machines, suggest that trillions in potential global AI value may be lost through drift and instability. Each non-reproducible output compounds uncertainty, reduces auditability, and erodes business confidence in automated systems.
“AI has become our new infrastructure,” says Martin Lucas. “But it’s built on probability, not physics. The same model can disagree with itself across two runs – which means trust, compliance, and accountability all start to erode before scaling begins. We needed a new foundation. So we built one.”
The science behind Decision Physics
Decision Physics introduces four proposed laws of deterministic computation:
- DP-1: Replay Invariance – identical inputs and states yield identical outputs.
- DP-2: Symbolic Isomorphism – semantically equivalent statements resolve to identical canonical forms.
- DP-3: Lineage Conservation – every output carries immutable provenance (Λ).
- DP-4: Drift Nullification – output stability persists despite model or corpus updates.
This framework aims to transform AI from stochastic sampling into deterministic function mapping – aligning machine intelligence with the principles of reproducible science.
The research leveraged a proprietary deterministic build environment, the TheaHQ Deterministic Build Pack, implementing deterministic clocks, seeded randomisation, and verification receipts. Across 1,000 test iterations, every output was identical at the bit level, confirmed via SHA digest verification.
The experiment indicates that AI systems may be able to achieve audit-grade reproducibility – a development that could pave the way for regulator-ready verification in finance, healthcare, defence, and government applications.
The implications
Deterministic AI could offer major benefits, including:
- Regulatory compliance: reproducible outputs that meet audit and legal standards
- Scientific reproducibility: research that can finally be replicated identically
- Economic stability: models that do not compound uncertainty
- Operational trust: systems that think consistently, not statistically.
Editor’s note
While the principles of Decision Physics are still emerging and require independent validation, they raise important questions for resilience and risk professionals. If deterministic AI proves viable at scale, it could mark a step change in the reliability, auditability, and trustworthiness of machine-driven decision systems.
