A post-Transformer architecture that keeps living memory across sessions, carries causality inside the forward pass, and imports knowledge the way you install an app. And it is already trained from scratch.
A memory that outlives the session
On screen
Real screenshots of the application
What it is
Crisol is the sovereign cognitive organism: a complete architecture, not an incremental improvement on the Transformer. It replaces quadratic attention with HoloBinder, a single context vector updated token by token at constant cost that survives the session. It anchors representation in a holographic space with separate functional regions — one axiomatic, one of physical-causal invariants, and a free subspace filled by the experts — instead of an undifferentiated continuum. And it puts causality inside the forward pass itself, with counterfactual do() operators, rather than leaving it to correlation. Above all: this is not a paper. The first Crisol Mini is trained from scratch.
Who it's for
Anyone who needs their intelligence to be theirs: running on their hardware, remembering across sessions, able to say who decided what, and not dependent on anyone renewing a licence.
How it works
HoloBinder — constant-cost context
A single 4096-dimension h_ctx vector updated token by token by Hadamard product with 0.95 decay. There is no attention matrix growing with the square of the sequence: cost per token is constant, and the state does not die when the conversation closes.
Living memory
Three layers and eight memory types. The organism does not start from zero each session: it keeps what it has lived and carries on from there.
A space with regions, not a continuum
NAR carries axiomatic reasoning over 2048 axes shared by every Crisol — which is why an imported expert routes correctly from the first token — and NOE anchors the physical-causal invariants, 256 active in v1.0 within a reserved space of 512. Neither one costs capacity: they are orthogonal projections of the full vector, not slices of it, so the expressive space remains all 4096 axes.
Native causality
Counterfactual do() operators inside the forward pass, not a prompt layer on top. The system tells what correlates from what causes, and can simulate the intervention.
Agentic, with someone responsible
Four Z-agents with traceable responsibility over each decision, plus six integrated sub-agents. Facing an output you can ask who decided it and why, not only what came out.
Custodia ecosystem
Signed .crisolpkg expert containers, installed the way you install an app. Adding a domain does not force a retrain of the whole model, and knowledge becomes a licensable asset.
Frequently asked
Isn’t a loss starting at 45 a bug?
No: that is the expected starting point. The model begins uncalibrated, worse than chance, crosses the uniform baseline at step 478 and falls to 3.5 (perplexity ≈34) over 5,000 steps, without a single NaN. And it does so having seen 164 million tokens — less than 1% of one epoch. The figure that matters is not the final perplexity but that a new architecture converges cleanly, measurably and reproducibly from scratch, on a corpus of its own that is still small. It is the foundation to scale it on.
None of this has to be taken on trust. Crisol has a site of its own, with the full architecture, the training figures and every phase of the project.
icrisol.comCrisol is trained from scratch on a corpus of its own, still small, and the architecture has already shown that it converges cleanly. The figures on this sheet are those of the Crisol Mini run, its first instance. The next step is to scale it with a large training run, and for that we are looking for compute: if you can provide it, let's talk.