Research · 01
Errloom
The post-training layer. Current RL pipelines are brittle glue code around a handful of scalar rewards — and that ceiling is what keeps well-trained small models rare. Errloom rebuilds the toolkit around four ideas.
Musical dynamics integration
During rollout a structured periodic signal — a “score” — is injected alongside the token stream and the sampling temperature is modulated in time with it. The premise is that reasoning has cymatic structure: certain rhythms of exploration and consolidation produce steadier chains of thought than a flat temperature. Errloom treats the score as a tunable object and lets RL discover which tempos help which task classes.
Mosaicware programming language (.mos)
A small declarative DSL in which a training run is written as a program: reward terms,
gates, curricula and rollout topologies are first-class expressions that compose. The
same .mos file that trains a model also specifies exactly how it was
trained — which is what makes all weights released auditable
rather than a slogan.
Redefined RL abstractions
The vocabulary is deliberately physical. Rewards act as gravity (a field the policy falls toward), rubrics as attractors (stable points the trajectory orbits), environments as looms (a fixed warp the policy weaves against), rollout sets as tapestries (the finished weave you inspect for defects). The metaphors carry real intuitions — restoring force, resonance, superposition — that transfer to how the optimiser behaves.
Procedural learning architecture
Rather than baking skills into weights, Errloom encodes learning procedures as small programs that attach to the model kernel and run at inference. A procedure is a reusable unit of “how to get better at X” that the model can call, compose and refine. This is the mechanism behind Thauten’s operator library and, downstream, behind the valuation engine improving its own estimators over time.
In the protocol
Errloom updates the engine’s estimator weights wₖ by multiplicative
weights against realized error. When ground truth arrives — a round, a tender,
an IPO — estimators that were right gain weight.
wₖ(t+1) = wₖ(t) · exp( −η · ℓₖ(t) ) / Z(t)
ℓₖ(t) = ( vₖᵢ(t*) − P_settle )² / P_settle² // scored only at settlement t*