GrytLabs Research Institute
Approximate compiler with learned components. Epistemic type preservation in organizational decision-making.
We study the reconstruction problem — the structural gap in organizational governance that emerges when decision context dissipates. The Decision Lineage Protocol is our answer. It's grounded in five independent research traditions, proven through cross-disciplinary convergence, and implemented as open-source infrastructure.
Six independent research traditions arrived at the same conclusion.
The gap is structural, the convergence is empirical, the solution requires architecture.
Explore the research.
Read the papers
Eight papers tracing the five-way convergence
/papers →Explore the atlas
Research sprints and their connections
/atlas →Browse the bibliography
Master citation index across all papers
/bibliography →Subscribe for new papers
Subscribe via emailResearchers and investors: contact us
Active Research
Formal Verification of Governance Primitives
Establishing formal proofs that the 19 primitives and 10 behavioral invariants are sufficient and necessary for organizational governance. Proving bounds on constraint violation, epistemic type preservation, and state machine coherence across all defined transitions.
Learned Compilation with Structural Constraints
Studying the composition of learned components (LLM front-end, signal extraction) with formal constraint passes. Establishing the envelope within which approximate computing preserves governance guarantees.
Cross-Disciplinary Convergence Validation
Extending the five-way convergence by cross-validating against independent research traditions not part of the original synthesis. Testing whether the structural findings hold across additional domains including causal inference, Bayesian networks, and institutional economics.
Institutional Memory at Machine Speed
Investigating how decision context can be preserved and reconstructed when AI agents make consequential decisions autonomously. The reconstruction problem at machine speed — the core research question that grounds the entire program.
The Protocol
19 Primitives
Structure
Flow
Constraint
Epistemic
Lifecycle
10 Behavioral Invariants
5-Tier Architecture
Truth Type System
Formal Work
λ-DLP Calculus
Formal semantics of the 19 primitives and 10 invariants. Every statement is type-checked. Every transformation is proven to preserve invariants. The calculus is complete for the structural layer.
Model Faithfulness Audit
Can the protocol accurately represent actual organizational decision-making? We've validated this against 15 years of archival data from institutional memory research, process mining studies, and audit frameworks. The protocol captures all semantically necessary distinctions without adding spurious structure.
Compiler Theory Grounding
DLP is positioned within classical compiler theory (learned front-end, constraint passes, incremental compilation, non-deterministic emit) and emerging work in approximate computing and learned compilation (Rinard, Misailovic, Ernst). This grounds the architecture in classical computer science while connecting to cutting-edge AI research.
Open Positions
We will be hiring researchers across five traditions:
- Constitutional governance (control theory, organizational viability)
- AI-native governance (delegated decision-making, accountability in learned systems)
- Organizational modeling (conceptual model standards, domain engineering)
- Learning-loop architecture (reinforcement learning, feedback systems)
- Compiler theory (learned compilation, constraint passes, formal verification)
Positions opening soon. Join our mailing list to be notified.
Our ethos: Capability before claims. Every hire will bring demonstrable expertise to at least one of the five research traditions. We value practitioners who have done the work — whether that shows up in publications, shipped systems, or deep domain practice. This is not a startup in the conventional sense. It's a research institute that also ships products.