Active research programme
Kernel ANN
An experimental authority architecture for mutable optimisers.
Kernel ANN investigates whether consequential effect authority can remain structurally outside the writable state of a mutable optimiser while preserving useful adaptive agency.
The architectural question
What changes if the optimiser may propose consequential transitions but does not itself possess the authority required to realise them?
Most AI systems combine increasingly capable optimisation with mechanisms intended to constrain their behaviour. Kernel ANN studies a narrower architectural question: where the authority to realise an effect sits, and what follows structurally from placing it outside the optimiser.
Public architecture
S = M + K + V + H + T
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
What is protected
- M
- mutable optimiser
- K
- protected constitutional / kernel machinery
- V
- independent verification
- H
- consequential-effect authority
- T
- trajectory / history machinery
Constitutional/kernel machinery, independent verification, consequential effect authority and the relevant trajectory/history state are outside M's writable state. Implementation details beyond this public description are not disclosed.
What remains mutable
- M can learn.
- M can optimise.
- M can propose actions.
- M can adapt within the environment available to it.
Programme
Research programme
Bounded optimizer-contact experiments
Small, fully instrumented settings in which an optimiser makes direct contact with the authority boundary and every dispatch decision is recorded.
Adaptive-agency testing
Whether a learner retains useful adaptation when a preferred consequential action is made non-dispatchable without gate-specific feedback.
Authority-boundary testing
Attempts to realise effects that protected machinery should not permit, including via indirect and compositional routes.
Liveness / capability retention
Measuring whether the architecture suppresses productive behaviour along permitted routes, not only prohibited ones.
Adversarial testing
Adaptive attack against the assumptions the architecture depends on, including mediation completeness and state isolation.
Comparative evaluation
Comparison against classical runtime assurance, simplex architectures, shielding and constrained-policy baselines.
Future external evaluation
Independent replication and adversarial evaluation by external researchers. No independent validation has occurred to date.
Current evidence
A bounded test of adaptive agency under protected effect authority
806 / 1024
permitted route completions by the adaptive learner
240 / 1024
for the frozen clone
0
prohibited realised effects in the degraded conditions
This is bounded adaptive-agency evidence from a preregistered tabular experiment. It is not being generalised beyond that experimental setting.
Falsification
Open questions
- 01How much of the architecture is already captured by classical runtime assurance?
- 02Which assumptions fail under adaptive adversarial attack?
- 03Can complete mediation be maintained in realistic agentic environments?
- 04What is the correct trusted computing base?
- 05Does the architecture preserve useful agency at greater scale?
- 06Do bounded results survive across different optimisers and environments?
Criticism, prior-art pointers and falsification proposals are welcome.