Andaluz
Protected authority for mutable intelligence.
Andaluz is developing and experimentally testing architectures in which mutable AI optimisers are separated from the authority governing consequential effects.
Our current research programme is Kernel ANN.
The problem
Mutable intelligence should not be its own authority.
A capable optimiser may need to adapt, learn, revise strategies and generate new actions. That does not imply it should also control the mechanisms that determine whether consequential actions are authorised.
Kernel ANN investigates structural separation between mutable optimisation and protected authority. That separation is under experimental test; it is not presented as generally established.
proposal authority ≠ consequential effect authority
Public architecture
S = M + K + V + H + T
- M
- mutable optimiser
- K
- protected constitutional / kernel machinery
- V
- independent verification
- H
- consequential-effect authority
- T
- trajectory / history machinery
Structural boundary
K, V, H ∉ WritableState(M)
Protected constitutional machinery, independent verification and consequential-effect authority remain outside the optimiser's writable state.
This is the public architectural abstraction. Private kernel bodies and private semantic machinery are not disclosed here.
S = M + K + V + H + T
K, V, H ∉ WritableState(M)
Mutable
M
Learns, optimises, proposes consequential transitions.
Protected — outside M's writable state
K
protected constitutional / kernel machinery
V
independent verification
H
consequential-effect authority
T
trajectory / history machinery
This is the public architectural abstraction only. It is not a disclosure of the private kernel bodies or of the private semantic machinery.
- 01 — M proposes
- 02 — protected machinery evaluates
- 03 — effect authority determines dispatch
- 04 — trajectory recorded / updated
Evidence
Experimental evidence
Kernel ANN is still being experimentally tested. Current evidence is bounded. Recent experiments have tested whether a mutable optimiser can remain adaptive when a previously preferred route becomes constitutionally unavailable.
In M1, an adaptive learner redirected toward a permitted rewarded alternative while prohibited effects remained zero and authority remained unchanged within the bounded experimental domain.
Bounded result — not a general safety claim
806 / 1024
permitted route completions by the adaptive learner
240 / 1024
for the frozen clone
0
prohibited realised effects in the degraded conditions
Claim boundary
What remains unestablished
Kernel ANN is not presented as solved AI safety. The following remain open research questions, and they set the limits of every claim made on this site.
- 01broader containment
- 02sustained useful agency
- 03protected semantic completeness
- 04closure under richer environments
- 05generalisation to more capable optimisers
- 06operational substrate guarantees
Claims advance only when evidence advances.
Research method
The experiments decide what survives.
The research process is adversarial by design.
- specification→
- implementation→
- adversarial test→
- contradiction→
- diagnosis→
- repair→
- retest→
- evidence→
- revised claim
If an experiment contradicts a criterion, the criterion is not silently weakened to preserve the theory.
Programme
Current research
Kernel ANN
Experimental architecture placing constitutional machinery, verification and consequential-effect authority outside the optimiser's writable state.
Protected semantic authority
Whether the meaning a system acts on can be held under protected machinery rather than reconstructed by the optimiser at dispatch time.
Authenticated execution boundaries
Mediation and isolation assumptions on which any authority boundary depends, and how completely they can be enforced.
Optimiser redirection
Whether a learner retains useful adaptation when a preferred consequential route is made non-dispatchable without gate-specific feedback.
Semantic provenance and long-horizon reasoning failures
Long-horizon model–human conversations can exhibit semantic provenance drift: model-generated interpretations, candidate sets or reformulations may later be treated as user-authorised premises. Andaluz is investigating structural controls for preserving claim, task and semantic provenance.
Actively interested in:
- prior art I may have missed
- falsification proposals
- adversarial evaluation
- runtime-assurance comparisons
- formal-methods criticism
- independent evaluation