Category
Flagship cross-domain · generation 1
Models · Generation 1
A flagship model
with the substrate underneath.
Flagship post-transformer model trained on the ReasonLoom substrate.
- Model ID
RL-X1.G1.2026- Substrate
- Stamen runtime · Heddle memory
- Training
- Atelier developmental loop · typed verifier in-loop
- Context ceiling
- there isn’t one
Substrate
Stamen + Heddle
Trained with
Atelier developmental loop
Best for
Long-horizon reasoning without context limits
What RL-X1 is
A model with the substrate underneath
RL-X1 is the first generation of our flagship cross-domain line. It does not live inside a context window. It reads, binds, and composes through Heddle, runs on Stamen, and is reared by Atelier. The result is reasoning quality on long-horizon tasks that comes from architecture, not from prompt engineering.
The structural shift
Why the difference is not “a bigger window”
Conventional models scale by extending an attention buffer. RL-X1 does not have one to extend. The work that the window used to do is done by the substrate instead.
Conventional
Token window
- × Memory ceiling = buffer length.
- × Recall is a scan of attention.
- × Composition re-derived per turn.
- × Provenance lives in prose.
RL-X1
Structured substrate
- + No buffer to overflow.
- + Recall is a substrate primitive.
- + Composition is bind/walk, not re-read.
- + Provenance is structural.
What changes versus a transformer
Three structural differences
RL-X1 is interesting because of what it is not — not a bigger attention model, not a tokens-in-tokens-out model, not a one-shot decoder.
X1
No context-window ceiling
Memory lives in structured binding, not in a buffer the decoder has to scroll. Long-horizon tasks stop being a token-budget problem.
X2
Composition is a primitive
Reasoning over analogy, counterfactual, and multi-hop chains uses the same bind/recall surface. The model does not have to re-derive structure from language each turn.
X3
Grounded by training
The model is reared by Atelier, with a typed verifier in the loop. What it knows, it can defend; what it does not know, it defers on.
Where it sits
Internal evaluation
Numbers are internal — the suites and conditions are documented in the evaluation programme. The pattern, not any single value, is what we report.
| Task family | RL-X1 | Conventional baseline | Δ |
|---|---|---|---|
| Long-horizon multi-hop | P@5 1.00 | P@5 ~0.62 | +0.38 |
| Cross-document binding | 0.94 | 0.71 | +0.23 |
| Compositional analogy | 0.88 | 0.56 | +0.32 |
| Defer-on-unknown | 0.96 | 0.41 | +0.55 |
| Context-window overflow | 0 | frequent | n/a |
P@5 1.00
Multi-hop retrieval through the stack
End-to-end retrieval through the model and memory bridge.
+0.65
Lifelong retention vs amnesiac control
Inherited from the Atelier developmental loop.
0
Context-window failure modes
There is no context window to overflow.
A reasoning trace
What a multi-hop question looks like
A question that would force a conventional model to scroll its window becomes a sequence of substrate operations.
rl-x1 · trace · multi-hop
- 01 READ
perceive(corpus)Inputs land as structured evidence — not as a token buffer.
- 02 BIND
bind(claim_a, source_a)Claim is tied to where it came from. Provenance is structural, not appended.
- 03 BIND
bind(claim_b, source_b)A second piece of evidence is bound. No re-derivation from prose.
- 04 WALK
walk(claim_a → claim_b)Multi-hop is a substrate operation. The decoder does not have to scroll.
- 05 COMP
compose(answer | evidence)The answer is composed from bound evidence. What is asserted is defensible.
- 06 EMIT
emit(answer, audit_trail)Output ships with the audit trail attached. Through Mnemo, this is enterprise-ready.
The X-line
Where RL-X1 sits in the generation roadmap
G1
RL-X1
shipped
Flagship cross-domain · long-horizon reasoning without context limits.
G2
RL-X2
planned
Multi-modal substrate native. Perception and binding share the same surface.
G3
RL-X3
research
Self-revising recall. The model edits its own memory under typed verification.
Where RL-X1 is being used
Cross-domain reasoning
Long-horizon analysis
Tasks that span hundreds of inputs and need structured recall across all of them. The model is not asked to fit them in a window.
Scientific reading at scale
RL-X1 reads collections, binds claims, and composes inferences across them. The work product is structured, not narrative.
Memory-aware decision support
Used through Mnemo, RL-X1 reasons over multi-tenant memory with the audit trail attached.
