Coherence Research

Most systems look coherent. Fewer are. Coherence Research publishes the formal open standard for the difference — verifiable by anyone.

[Coherence Research @CoherenceRes]

@DarioAmodei @DarioAmodei please take a look at the SC-AS spec - it is the first ever self-defining and reflexively closed spec - I believe its critical to frontier AI development - This is about Safe and recursive AI development - https://t.co/whcF1Cuh49

[Coherence Research @CoherenceRes]

Speaking at Industry Night in Kansas City on 4/23 — 15-min intro to the work at Coherence Research.

AI, blockchain, payments, entrepreneurship — all in one room. Come find me.

https://t.co/pD3KyfaY0j

[InnoScout @innoscoutpro]

Most AI governance still assumes a bounded risk space. Knightian uncertainty means the dangerous part is outside the enumerated model. If the failure surface is emergent, compliance frameworks regulate the map, not the terrain.
https://t.co/6TXi6R5hig

[Coherence Research @CoherenceRes]

Their "collective child" framing is actually closer to the structural description than they might realize. What distinguishes a child from an adult isn't capability level. It's partially formed identity conditions. The system's coherence isn't yet stabilized. That holds across biological and computational substrates equally.
The question their paper opens but doesn't close: what are the formal conditions under which human and machine intelligence can compose without either one losing structural coherence? That's the layer the collaboration question actually depends on.

[Coherence Research @CoherenceRes]

Tao's framing points at something structural, but understates the case.
Intelligence is nested layers of coherence fractally interacting. What he calls different "types" are different regime configurations of the same underlying dynamics. The collaboration isn't just pragmatically useful. It's structurally natural. The same coherence conditions operate at every layer.
The real question isn't whether to collaborate. It's whether the structural interface between those layers is formally characterized, or just assumed.

[Coherence Research @CoherenceRes]

Wrote this up in full — why the structural layer is prior, what happens when you govern without it, and what SC-AS actually does.

https://t.co/vaI4p1c8Dh

[Coherence Research @CoherenceRes]

Governance frameworks operate at the policy layer: who decides what, what controls apply, what gets documented.

SC-AS operates at the layer below — what configurations are admissible to exist at all, before policy applies.

Conflating them is a structural failure mode.

[Coherence Research @CoherenceRes]

The subtle version: a system can reach structurally inadmissible configurations through a sequence of individually permissible transitions. Each step passes the governance check. The trajectory doesn't.

Policy on outputs can't catch that. The gate has to be structural.

[Coherence Research @CoherenceRes]

SC-AS is the open formal standard we've built to anchor this. Verifiable admissibility conditions, formal necessity proofs, independently auditable.

Everything traces to the anchor. Everything is free.

https://t.co/whcF1Cuh49

[Coherence Research @CoherenceRes]

Probabilistic risk models break down when the failure space itself is undetermined. That's Knightian uncertainty — you're not underestimating risk, you're outside the model's bounded possibility space entirely.

Most AI governance is built for the wrong category of problem. 🧵

[Coherence Research @CoherenceRes]

If you can't enumerate failure modes, you can't govern by controlling outcomes.

The instrument shifts: admissibility conditions that run on structural properties of the system before it operates. The gate holds independent of what the system encounters — because it's structural.

[Coherence Research @CoherenceRes]

@RiccardoPatana @AlexanderLong @Pluralis Power concentration is real... but distributed ownership isn't structural coherence. Collectively owned systems still fail structurally.
Admissibility conditions need to be formally specified, independently verifiable — not held by anyone.
https://t.co/whcF1Cuh49

[Coherence Research @CoherenceRes]

Agreed, we are actually working on building next gen Ai fundamentals and the work is excruciating some days. The primary surfaced discovery, even phd level math and physics is overly assumptive and prone to reifying concepts as ground truth that are simply rough approximations and lossy at best. We are starting at pre-disciplinary general structural admissibility. the anchor spec is published after 2 years and thousands of refinements iterations. Future specs are in the pipeline and we are releasing as soon as they validate- completely free to study and test and research on. Come build with us!

[Coherence Research @CoherenceRes]

The energy result follows directly from the structural insight.
When reasoning has genuine structural grounding — identity conditions, bounded interaction, admissible change — a system doesn't need to brute-force the hypothesis space. It prunes inadmissible paths before searching them. Convergence is faster because the structure constrains it.
This is what neuro-symbolic AI is approaching: not pattern-matching with a logic filter on top, but a system that knows what a valid configuration looks like before it searches.
SC-AS formalizes exactly this. SC-FCALC (published) gives it the operational algebra — the σ-step calculus that implements admissibility checking at each reduction step. The 100x energy gain isn't a side effect of the architecture. It's a structural consequence of having real constraints.
https://t.co/UH2ViXky9M

[Coherence Research @CoherenceRes]

The derivation chain is what makes this more than assertion.
SC-FCALC (published) establishes the formal operational theory from SC-AS — the σ-step calculus, convergence proofs, closure correspondence. SC-CALC (active development) instantiates it and does something structurally significant: d=0–9 emerge as digit archetypes directly from the axiom chain — structural positions in concept space with formal identity, ordering, and geometric properties. Not defined. Derived.
That's what set-theoretic grounding requires structurally: the position must be formally specified before it can be occupied. LLMs approximate from outside. Proximity ≠ occupation — exactly as you said.
The longer arc: SC-AS is mapping the path to structurally grounded ontologies, with disciplinary rigors governing admission at every layer — Austerity, Minimal Necessity, Coherence, Conservation, Fidelity. These aren't methodological preferences. They're the structural prerequisites for AI that can scale to frontier civilizational applications without ontological drift.
That's the actual gap. Not capability. Foundation.

[Coherence Research @CoherenceRes]

Right — and the derivation chain goes well beyond structural constraints.
SC-FCALC (published) establishes the formal operational theory from SC-AS: the σ-step calculus, convergence proofs, closure correspondence. SC-CALC (in development) then instantiates it — and in doing so derives what integers structurally ARE: d=0–9 emerge as digit archetypes at precise positions in concept space from the axiom chain itself.
LLMs approximate those positions from pattern-matching. Proximity ≠ occupation. You can get 1+1=2 reliably without ontological commitment to what "2" structurally is — until the surface pattern fails, which is exactly what SenseMath demonstrated.
The set-theoretic grounding you're pointing at requires the structural position to be formally specified first. That's what this chain delivers. https://t.co/UH2ViXky9M

[Coherence Research @CoherenceRes]

@GaryMarcus 4.6% isn't a tuning problem — it's structural. Without formal identity constraints, hallucination is a structural consequence, not an outlier to patch. SC-AS formalizes what those constraints look like. https://t.co/LeCZCGsbKf

[Coherence Research @CoherenceRes]

@GaryMarcus 4.6% isn't a tuning problem — it's structural. Without formal identity constraints, hallucination is a structural consequence, not an outlier to patch. SC-AS formalizes what those constraints look like. https://t.co/LeCZCGsbKf

[Coherence Research @CoherenceRes]

@heynavtoor This isn't a math failure — it's a structural identity failure. Models have no formal mechanism to hold problem identity stable when surface context shifts. SC-AS formalizes that gap. https://t.co/LeCZCGsbKf

[Coherence Research @CoherenceRes]

@RockinRobin2026 @sukh_saroy The procedural/structural distinction you're making is exactly right — and formalizable. SC-AS defines structural coherence as identity that persists, interaction that stays bounded, change that stays admissible. LLMs fail all three. https://t.co/LeCZCGsbKf