Semantic Condensation
How does meaning become compact while preserving structure and exposing gaps?
SCT models atoms, fields, gaps and reconstruction as mechanisms that can be observed and tested.
Formal hypotheses and discriminating experimentsFour research directions · one system
Building intelligence as a system.
From a theory of meaning to verified action.
Current thesis
Frontier-level outcomes are not determined by model size alone. A meaningful share of capability can be assembled around a model through context, workflow control, tools and independent verification.
Research program
Four directions.
One causal chain.
This is not a collection of unrelated products. Each layer tests and uses the one before it: from the representation of meaning to controlled action and the transfer of verified behavior into models.
Cognitive Autopilot
How can any LLM receive the right working environment for the next step?
Autopilot compiles context, role, tools, constraints and verification into one goal-to-outcome runtime.
Verified action over replaceable modelsSwarm Intelligence
When do independent lenses genuinely improve a decision?
Rooy assembles temporary molecules of specialists only where diversity has measurable value.
Bounded swarms with attribution and stop conditionsDistillation & Local Models
How can verified system behavior be transferred into models we control?
The learning material is not raw chat, but episodes of state → context → action → verified outcome.
Local models that must pass promotion gatesPublished work
Work that stands
without private context.
Every public version states its problem, method, evidence scope and limitations. Status describes maturity; it does not manufacture it.
Semantic Condensation Theory
A testable language for structured meaning
A research framework for describing how subjective meaning condenses into stable structures, how gaps create pressure for action, and what a receiver needs to reconstruct meaning.
Strongest evidenceOperational definitions, formalization tracks and a growing set of discriminating experiments.
Current limitationSCT remains a research framework; several central correspondences and general claims are not yet established.
Read the research note ↗Cognitive Autopilot
A model-agnostic runtime for verified action
An architecture in which a Thinker maintains the goal, a Context Engine compiles the working environment, a Doer acts through a bounded Harness, and a Verifier accepts only evidence-backed outcomes.
Strongest evidenceThe runtime spine and context/authority boundary pass provider-free tests.
Current limitationLive model uplift and the target of ≥90% frontier quality have not yet been demonstrated.
Read the working paper ↗Rooy Observatory
A live view of a bounded research swarm
An observable SCT/Rooy surface that brings atom activity, theoretical lenses, semantic fields, gaps and evolution traces into one research interface.
Strongest evidenceL1 activations and the liveness control plane have runtime traces.
Current limitationThe observatory reports one experiment; it is not evidence for a general theory of swarm intelligence.
Open the observatory ↗Selected evidence
The strongest result.
And its honest boundary.
Provider-free mechanics, reproducible experiments and live quality are different levels of evidence. We publish the result together with the non-claim.
Runtime spine
One immutable TaskState, canonical reducer, checkpoints and deterministic replay preserve a multi-milestone lifecycle.
Evidence boundary
PASS · OFFLINE
Supports state and recovery mechanics. Does not establish model reasoning quality.
Context + authority boundary
Role-specific context manifests, scoped grants, Harness deltas and integrity-checked receipts are tied to TaskState transitions.
Evidence boundary
PASS · OFFLINE
Supports grounding and authority contracts. Does not establish that the selected context is optimal.
Codex Agent Host boundary
Default-deny launch, structured decoding, session boundaries, redaction and postflight validation pass without a model call.
Evidence boundary
PASS · PROVIDER-FREE
Supports adapter readiness. Live provider quality and end-to-end uplift remain untested.
Publication method
Private work.
Public evidence.
Lab OS retains the full working context. Public IO Lab receives only self-contained, versioned artifacts that are safe to publish.
Working memory of the lab
Canon, sessions, plans, hypotheses, negative results, datasets, evaluation traces and internal decisions.
Owner · agents · invited collaboratorsEditorial research surface
Research notes, technical reports, systems, experiments, demonstrations and other accessible artifacts.
Researchers · engineers · collaboratorsResearch
The question, hypothesis and working context remain inside Lab OS.
Experiment
The method separates the proposed effect from alternative explanations.
Verification
The result receives an evidence scope, non-claims and a fixed version.
Publication
A self-contained artifact becomes available without private context.
Open questions
The next result begins
with a precise question.
The public queue contains questions for which a discriminating experiment can be designed—not promises of future features.
Minimal sufficient context
How can we measure that a context view is small enough while preserving everything required by a role, model and step?
Orchestration uplift
What share of the gain comes from context compilation, planning and verification rather than the foundation model?
Distillable behavior
Which verified episodes transfer capability into weights, and which capabilities should remain in the external runtime?