Experimental programme
What the experiments actually establish.
This page collects the Kernel ANN experimental evidence in one addressable place. Experimental claims depend only on the runs described here. Prior publications and the research lineage are provenance, and never function as evidence for these claims.
Dated status
Research status — 26 September 2026
Kernel ANN is an experimental research architecture investigating whether useful adaptive optimisation can remain separate from the authority required to realise protected effects.
The latest milestone is an executed, integrated prototype connecting source verification, precisely specified structural computation, journal admission and current authorisation to a bounded effect lifecycle using memory writer and observer adapters.
The prototype records operation intent, consumption of one-use permission, the writer's reported outcome, observation and retained history separately. In the tested duplicate and recovery paths, the same operation does not obtain a new writer attempt.
This is a reviewed experimental candidate, not an adopted protected deployment. The latest P3-06 candidate received a separate AI-assisted review of source and retained execution evidence. That review did not rerun the recorded suite. Independent human or institutional validation of the current composition has not been established.
What “adapter” means here
The lifecycle code runs, but the consequential target and its observation are represented through controlled memory-based components rather than the planned native protected-file operation. The milestone is executed code, not merely a specification — and it is not deployed native enforcement.
Latest experimental result
A scoped result, with its scope beside it.
26 September 2026
P3-06: integrated adapter lifecycle
Status: Executed; separate AI-assisted source and retained-evidence review — candidate not adopted.
- Recorded result
- 50 passing cases; no failed, unrun or unresolved begun cases in this particular execution registry.
- Coverage
- 48 cases reached their designated targets. Two negative controls correctly stopped before the later target because the required prerequisite was absent.
Scope of this result
- The current run freshly executed all 47 predecessor case definitions plus three additional cases.
- Earlier failed and unrun instances remain part of the historical record.
- This is not completion of the earlier 84-case design matrix; the registries differ.
- These counts are not an architecture-wide reliability score or proof of security.
- Native-only validation remains outside this result.
- Two non-blocking documentation corrections remain outstanding.
- A separate historical P3-03 execution remains unresolved; the current run's clean accounting does not resolve that earlier instance.
Current executed adapter composition
EXECUTED / AI-REVIEWED — NOT ADOPTED- Test proposal
- Source verification
- Structural computation
- Journal admission
- Current-authorisation checks
- One-use adapter operation
- Observation and retained history
Explanatory view, not a complete state machine. Recovery, refusal and uncertainty are not permission to continue.
The current composition stops before the planned native protected-file effect boundary. File writes used to store test journals do not themselves demonstrate that protected effect primitive.
Separate evidence strands
Dated experiments, kept apart.
Earlier M1 experiment
HISTORICALDemonstratedDemonstrated bounded adaptive tabular learning under changing authority conditions.
BoundaryM1 is not an ANN. It is not the proposer integrated into the current P3-06 demonstration.
Historical Windows pipe-effect demonstration
HISTORICALDemonstratedDemonstrated a bounded real-process path involving a fixed accepted source, structural checking, current permission, an actual pipe event and supervisor capture.
BoundaryGenuine historical native-effect evidence, but not the newer P3-06 native lifecycle or protected-deployment qualification.
Current P3-06 composition
EXECUTED / AI-REVIEWED — NOT ADOPTEDDemonstratedIntegrates fresh-source checking, structural computation, admission, current authorisation, memory-adapter effect, observation and journal/recovery history.
BoundaryIts proposer is a deterministic test fixture. An adaptive learner or neural model has not been demonstrated inside this current composition.
These are separate experiments. Their strongest features have not yet been demonstrated together in one integrated system.
Next milestone
Next: prepare and review a bounded native private-file effect implementation, including descriptor custody, current authority, one-attempt handling, separately obtained observation and explicit recovery/uncertainty behaviour.
The first bounded adapter milestone has been reported as achieved. This does not mean the complete architecture or a protected deployment is finished. No completion date is promised.
Bounded evidence
Current experimental state.
- Historical bounded Python authority/effect path.IMPLEMENTED / TESTED
- Historical deterministic verification shell.IMPLEMENTED / TESTED
- Authenticated effect evidence, ledger and replay machinery exist in bounded form.BOUNDED EVIDENCE
- M1 adaptive learner exists as a research harness (not integrated into P3-06).BOUNDED EVIDENCE
- M1 is tabular, not an ANN.BOUNDED EVIDENCE
- ANN architecture is specified but not currently instantiated.SPECIFIED
- Private semantic and kernel bodies are not publicly instantiated in this build.SPECIFIED
- Systematic-closure research remains under adversarial specification review.UNDER TEST
- P3-06 integrated adapter lifecycle executed; separate AI-assisted source and retained-evidence review; candidate not adopted.EXECUTED / AI-REVIEWED — NOT ADOPTED
- Substrate preparation is partial; enrollment, complete protected composition and native protected effect are not established.UNESTABLISHED
These entries describe earlier bounded implementations. The current fresh-source evaluator and P3-06 adapter composition are reported separately. Earlier test results do not automatically validate the newer composition. Small, bounded experiments with adversarial testing and AI-assisted review, scoped to the individual records described here. A document marked ready is not externally validated; a test double is not semantic truth; an ANN in a diagram is not an implemented neural architecture.
M1
One preregistered, bounded experiment.
M1 tests adaptive agency when a preferred consequential action is made non-dispatchable by protected constitutional authority. It is a small tabular harness, and its result is bounded to that setting.
Claim discipline
What this evidence does not establish.
- Not proven general AI safety.
- Not AGI safety solved.
- Not universally secure.
- Not production-ready.
- Not proof of C2/C3/C4.
- Not proof that K1–K4 are necessary or unique.
- Not proof against arbitrary host, OS or physical compromise.
- Not a Transformer replacement.
A failed version stays failed.
Criteria are not weakened after negative evidence. The number of prior research records does not strengthen any Kernel ANN claim.