Governed Adaptation

The change-control layer for AI that learns.

Lumenais can preserve context and propose bounded improvements. Gnosis decides what is allowed to influence future behavior.

Why it exists

Learning without governance becomes hidden drift.

Governance without learning becomes static software. Gnosis is the middle path: candidate changes can be proposed, but durable influence requires scope, review, testing, signed decisions, and rollback.

If Lumenais learns, the reviewer question is simple: what changed, why was it allowed, and how can it be reversed?

Promotion Pipeline

Proposed changes do not become hidden priors by default.

The public claim is bounded adaptation with auditable gates. Sensitive changes move through a governed pipeline before they can affect future retrieval, routing, memory promotion, or selected system behavior.

1

Static review

Candidate changes declare scope and are screened for dangerous patterns before execution.

2

Isolated testing

The proposal is evaluated outside production with bounded runtime and resource controls.

3

Coherence review

Expected behavioral delta is checked against policy, continuity, and safety constraints.

4

Signed decision

Promotion creates an audit record and rollback path; failure keeps the signal out of durable influence.

What it governs

The gates sit where learning can change future behavior.

Gnosis is not a separate product bolted onto the side. It is the governance spine for the parts of Lumenais that adapt: memory promotion, routing hints, specialist proposals, and selected bounded module changes in reviewed workflows.

Memory and routing influence

Reviewed context can shape later answers, but noisy or rejected signals should remain auditable without becoming hidden priors.

Candidate-improvement review

Proposals are evaluated as candidates, not assumptions. A rejected candidate is retained as evidence, not silently promoted.

Trust-tiered autonomy

Higher-impact actions require stronger evidence, stricter review, and more explicit rollback guarantees.

FieldHash-compatible records

Selected proposal and decision records can be bound to tamper-evident provenance for qualified technical review.

What it prevents

The point is not more autonomy. The point is safer promotion.

The strongest safety signal is prevention: useful learning can be delayed, rejected, scoped down, or rolled back when the evidence is weak or the risk is high.

Hidden drift from unreviewed memory or routing updates

Stale corrections becoming durable after better evidence arrives

Policy-conflicting shortcuts that improve a metric while weakening governance

High-risk changes bypassing human approval or rollback requirements

Boundaries

What Gnosis is not.

Not autonomous self-evolution

The system starts with limited autonomy and must demonstrate alignment before gaining more.

Not base-LLM training

Normal use does not update the underlying LLM weights; adaptation lives in governed orchestration layers.

Not a universal safety guarantee

The claim is inspectable change control, not proof that all future behavior is safe.

Not a bypass around humans

Core values, new capabilities, and high-risk deployment changes require explicit human approval.

Implementation Posture

Implemented primitives, selected workflows, qualified review.

The repo includes proposal logging, hierarchy validation, specialist trust metrics, proposal and implementation certificates, governed-action records, and rollback-ready deployment paths. Public materials summarize the control plane; full traces and implementation details are reserved for technical diligence.

Candidate changes pass static review, isolated testing, and policy checks before promotion.
Proposal, decision, implementation, and rollback records are retained for audit.
Trust tiers gate sensitive capabilities; access is earned, not assumed.
Selected Gnosis records can be FieldHash-bound where configured.

The safety story is the product story: governed continual learning is only credible when every durable change has a reason, a scope, a review trail, and a rollback path.

See how Gnosis fits into the full learning loop.

The whitepaper places Gnosis alongside memory, Deep Synthesis, FieldHash, and the benchmark evidence for governed continual learning.