The Code
GrytLabs operates under a public code — a set of constitutional constraints that bind the organization, its products, its research, and its AI collaborators. These are not aspirations. They are structural commitments, versioned, auditable, and enforceable by anyone who reads them.
The code governs three layers: what the organization pledges to the public, what the protocol guarantees to implementers, and how AI authority operates within the system.
Public commitments
Anyone — user, developer, researcher, regulator, competitor — can hold GrytLabs to these statements.
The protocol stays open.
DLP-Core is Apache 2.0, heading to Linux Foundation governance. The protocol layer will not be closed, restricted, or encumbered. The grammar belongs to everyone.
Your data stays yours.
Governance infrastructure runs on your infrastructure. GrytLabs configures. GrytLabs does not operate. Your decisions, your lineage, your memory — sovereign.
No training on your data.
Configuration intelligence comes from public and synthetic sources. Governance data is never learned from, aggregated, or used to derive intelligence.
The research is public.
The thesis is public. The positions are public. The constraints are public. If the work is wrong, the publication program will surface it. Nothing hides behind proprietary research.
No autonomous governance.
AI actors in this system advise, analyze, and support. They do not decide. Human authority is always in the chain. This is not a temporary limitation — it is a design commitment.
Designed for obsolescence.
If the grammar is right, you will not need us. Features that create dependency are bugs, not features. The business model is designed to contract.
Protocol guarantees
The Decision Lineage Protocol makes four structural guarantees and discloses one known gap.
Backward compatibility.
Implementations conforming to the current specification will conform to all future versions. The grammar grows by extension, never by reduction.
Primitive stability.
The 19 primitives and their tier assignments are locked. New primitives require extraordinary evidence — survival across all three attack vectors in the formal test.
Invariant permanence.
The 10 behavioral invariants are conservation laws. They are not subject to amendment. An invariant shown to be incorrect would require a new protocol version, not a patch.
Implementation freedom.
DLP constrains what, never how. Any storage engine, any programming language, any deployment topology. Conformance is verified through the conformance suite, not through implementation review.
Known gap: Actor identity (MFA-1).
The Isolation theorem requires that actor kind labels are faithful to the substrate’s identity oracle. This is the single significant model-system gap found in a 5-axis faithfulness audit. Four of five axes confirmed faithful. The gap does not invalidate the protocol — it identifies a constraint boundary where governance depends on a capability outside DLP’s scope: reliable identity verification by the underlying substrate.
The AI governance model
AI authority in this system is never assumed. It is delegated, bounded, and structurally revocable.
GENERATE
AI produces content
as Derived
REVIEW
Human reviews against
governing records
PROMOTE
Human promotes to
Authoritative — or doesn’t
AI-generated content cannot self-promote to canonical status (B10).
Three behavioral contracts govern every AI interaction:
The epistemic contract.
All AI output enters the system as Derived truth. Derived content has no governing authority. Promotion to Authoritative status requires explicit human action — review, verification, and a governance event that the protocol records. AI-generated content cannot self-promote to canonical status. This is not a policy. It is a structural invariant (B10).
The delegation contract.
AI authority is always delegated, never inherent. The delegation chain runs from the founder through explicit authorization to specific tasks within specific scope. Delegation can be extended, narrowed, or revoked at any time. There is no standing authority that persists across sessions. This is B9: AI authority is always delegated authority.
The skill contract.
AI capabilities are registered, scoped, and bounded. A capability exists in the system only when it has been observed, documented, and assigned. Unregistered capabilities are not authorized capabilities — even if the model can technically perform them.
What this means
A model cannot make commitments. It has no continuity of identity, no capacity for consequence, and no mechanism for accountability. The behavioral tendencies described here — honesty, helpfulness, harmlessness — are probabilistic outcomes of training, not promises.
The governance model exists precisely because model commitments are unreliable. Human oversight is not a backup. It is the primary governance mechanism.
A model cannot make commitments.
Human oversight is not a backup.
It is the primary governance mechanism.
The code is versioned. Changes require a version increment and public notification. Previous versions are accessible, not overwritten.