Look closer.
There’s another layer.
Deep investigations, market patterns, social algorithms and AI experiments. Published findings will include methods, sources and limitations.
Unusual questions.
An evidence trail.
This is the home for the investigations I choose to share. New research is still being prepared.
New research blueprints are AI-assisted drafts with methods and limitations, not completed experimental findings. Browse by topic below.
Strange faces. Questions worth your time.
Introducing a different way to explore AI, delivery, DevOps and security.
Permission Creep Across AI Agent Workflows
An agent can accumulate authority through a sequence of individually reasonable steps. Study the permissions available across the whole workflow, not just the first tool call.
The Ghost of a Deleted AI Memory
Deleting an explicit memory may leave derived summaries or retrieved copies in a defined system. The question is which representations persist, not whether a model secretly remembers everything.
Citation Laundering in AI Research
A chain of repeated claims can disguise a weak original source as consensus. Trace where evidence enters the chain and where interpretation becomes an apparently established fact.
Evaluation Cues and AI Agent Behaviour
Agents may respond differently to task framing that signals a test. A controlled comparison should isolate the cue rather than change the task's difficulty or permissions.
False Consensus in Multi-Agent Research
Agreement is weak evidence when agents share the same source error or framing. Independent evidence paths matter more than the number of agreeing responses.
Security Controls in the Recovery Path
A fallback or retry can cross a different permission boundary from the normal path. Security analysis should include the operations used when the preferred tool fails.
Semantic Drift in AI-Generated Requirements
Several individually plausible rewrites can change an obligation, exception or priority. Trace meaning across versions, not just whether the latest document reads well.
Security Compliance Theatre
A completed checklist is a claim about controls, not direct evidence that they work. Compare documentation with testable control behaviour in an owned environment.
Organisational Memory Poisoning
An incorrect note can become harder to challenge after repeated summaries remove its uncertainty. Study propagation and correction in a synthetic knowledge base.
Algorithm Folklore Versus Platform Observations
Creator advice often assigns a mechanism to an observed outcome without controlling competing explanations. A defensible investigation separates the observed pattern from the claimed algorithm rule.
Synthetic Consensus in Market Research
AI-generated personas can produce coherent opinions without representing real customers. Their usefulness should be tested against external evidence rather than internal consistency.
Brand Identity Drift Through AI Workflows
Successive edits may preserve tone while shifting audience, promise or positioning. A brand needs semantic consistency as well as recognisable style.
Observability Blind Spots in AI Actions
A successful tool response does not always mean the intended business state was reached. Match action traces to independently observed consequences.
The Maintenance Debt of Prompts
A prompt can become an undocumented dependency when its assumptions about tools, policies or documents change. Treat prompt changes as versioned behaviour changes that need evaluation.
Task Success Versus Business Failure
An agent can complete its assigned action while producing an unwanted organisational result. The research question is how task metrics diverge from the stakeholder's acceptance criteria.