+0.65 ± 0.02
Lifelong retention advantage
Consolidating runs vs amnesiac controls, multi-seed.
Cognitive · Active research
A trainer that rears cognition, not one that fits it.
A developmental trainer for cognition substrates — grounding, not memorisation.
“como mi madre” — ground first, generalise second.
What Atelier is
A curriculum, not a fine-tune.
Atelier is built on the conviction that the right way to train a substrate is the way a child is reared: by perceiving, producing, being corrected, binding, and consolidating, all on a two-timescale schedule. It is the conductor that wires the substrate, the structured memory, the fine-tuning framework, and the cognitive gym into a single curriculum — and the layer where we measure whether that curriculum actually produces grounded behaviour rather than a lookup table.
The conductor wires substrate, structured memory, the fine-tuning framework, and the cognitive gym into a single rearing loop — and the verifier sits in the middle so progress is never faked.
The loop
Five steps, every episode.
The constants are the steps. What changes is the world the student is reared in.
- AM1
Perceive
Evidence comes in from one or more modalities — including a live resource channel.
- AM2
Produce
The student acts under a grounded production objective. No shortcut paths.
- AM3
Correct
A typed entailment verifier checks the production. No faked rewards.
- AM4
Bind
Successes bind into structured memory with role-swap and multi-hop recall.
- AM5
Consolidate
A two-timescale CLS step folds the binding into lifelong identity.
Milestones
What we have measured.
- Phase 1
Decisive head-to-head landed
Three rearing pathways compared in isolation. Traditional grounded training works; a distillation shortcut collapses to lookup; the developmental loop grounds the world cleanly.
grounded 1.00shortcut 0.00lookup 1.00 - Phase 2
Honest verifier built
A typed entailment verifier with no fallback masking. Reward goes NaN when entailment goes NaN. Live arXiv path opt-in and auth-gated.
no faked NLINaN propagationauth-gated - Phase 4
Lifelong identity measured
Consolidating runs retain all prior worlds. Amnesiac controls forget catastrophically. The gap holds across seeds.
forgetting +0.00retention 1.00amnesiac forget +1.00 - Today
Multi-seed, twelve improvements landed
A six-improvement batch closed with multi-seed error bars. Slot-factored relational binding wins cleanly. Architectural-priors claim falsified honestly.
n=5slot vs byte +0.36arch-priors falsified
Measured (multi-seed, n=5)
Grounded, not memorised.
1.90 ± 0.14 ×
CLS sample efficiency
Two-timescale CLS schedule vs single-timescale baseline.
+0.36
Slot-factored vs byte-level binding
Held-out relational retrieval, role-swap held out.
Decisive head-to-head
The B-collapse, in one chart.
Three rearing pathways under matched compute. The distillation shortcut looks attractive on lookup tasks and disappears on grounded production.
Note: C−B advantage on grounded production = +0.79 ± 0.18 across seeds. C>A is not clean (±0.24). The robust win is rearing-method, not architecture.
What we falsified
Negatives we publish anyway.
Because the verifier never fakes a signal, Atelier is also the place where we publish what does not work.
-
“Post-transformer architecture wins on its own” — falsified at the same fair-fight conditions. An external vanilla Transformer reached 1.00 on naming under matched compute. The robust value lives in the objective and in slot-factored binding, not in the architecture.
-
Imprevisto recovery shrank from a romantic +0.25 to +0.13 ± 0.09 under multi-seed evaluation. We publish the smaller number.
-
Distillation shortcut paths look attractive in early runs and collapse to a lookup table under held-out evaluation. Reported as a paradigm-level warning.
“We rear cognition. We do not fit it.”
Where the curriculum runs
What Atelier is for.
Training the flagship lines
RL-X1 is reared inside Atelier. The loop is what turns the substrate plus structured memory into a usable model — not a fine-tune script.
See RL-X1 →Learners that do not forget
CLS-style two-timescale schedule is the basis for the continual line. Lifelong retention is measured, not assumed.
See RL-C1 →A platform for honest negatives
Two paradigm-sized falsifications have already gone through. The verifier is the reason the publishing bar stays high.
See evals →Related research
The work behind Atelier.
Available through