RABA Field Lab

What I show here

I use structured analysis to examine difficult questions at the boundary between business processes, AI systems, human decisions, and real-world consequences.

The work shown here demonstrates four things:

Structured problem analysis — separating the actual problem from assumptions and attractive explanations.
Comparison against existing solutions — checking standards, security controls, governance approaches, and adjacent methods before proposing something new.
Evidence-based decisions — documenting why a direction should continue, change, be reused, or stop.
Negative results — treating “do not build a new mechanism” as a valid outcome when existing approaches already solve the problem well enough.

This is a research portfolio, not a claim that every question requires a new framework.

Three examples of how I work

1. External instructions → AI agent action

What I tested:
Whether risks arising when an AI agent receives external instructions require a new governance mechanism.

What I found:
Existing controls — provenance, trust boundaries, authorization, policy enforcement, sandboxing, restricted capabilities, audit trails, and human approval — already cover the tested problem to a substantial degree.

Decision:
REUSE / STOP — no new RABA-specific mechanism justified in the tested scenario.

Read the worked example

2. Meaning preservation across AI transformation

What I tested:
Whether information can remain technically traceable while losing meaning that later matters for a human decision.

What I found:
The relevant question is not only whether data is preserved, but whether decision-relevant meaning survives transformation and handoff.

Decision:
CONTINUE — bounded unresolved research question.

Read the research note

3. Multi-agent meaning drift

What I tested:
Whether several agents can each follow their local rules while a changed interpretation propagates through the full workflow.

What I found:
Local compliance and preserved handoffs do not automatically prove that the original meaning remained intact end-to-end.

Decision:
RESEARCH CASE — synthetic worked example, not a live-system validation.

Read the worked case

How I approach a problem

Question → Existing solutions → Strongest counterexample → Evidence → Residual problem → Decision

Possible decisions include:

CONTINUE / MODIFY / REUSE / REASSESS / STOP

A useful analysis may end with:

“The existing solution is already strong enough. Do not build another mechanism.”

That is a successful result when the evidence supports it.

Evidence boundary

Some material on this site is based on public research, standards, and worked examples rather than deployment inside a live production system.

Where evidence has not been independently reproduced or a case is synthetic, it is labelled explicitly.

I treat that limitation as part of the analysis, not something to hide.

Explore deeper

The sections below preserve the fuller research trail, current unresolved work, methods, evidence boundaries, and governance context behind the portfolio examples.

A reader can stop at the portfolio summary above or continue into the research layer below.

Latest Research Transition

From Preserving Meaning to Preserving Governing Relationships

Recent work has narrowed the question beyond whether information or provenance survives an AI-supported transformation.

A process may preserve the original artifacts, evidence, and even the original wording while still changing what a governing condition means in practice.

The current research question is:

When evidence, observations, interpretations, and decisions move through an AI-supported process, what must remain invariant so that a downstream representation does not acquire meaning or authority that the original condition never gave it?

A useful shorthand is:

Reality → Observation → Evidence → Interpretation → Decision → Action

The question is not only whether each element is present or locally correct.

It is also:

What is allowed to change at each transition — and what must not?

Several distinctions are being pressure-tested:

evidence is not the same as the claim it supports;
observation is not the same as a determination of materiality;
materiality is not the same as authority;
technical capability is not the same as business or normative authority;
authority to suspend is not authority to redefine the underlying decision;
an unknown or unverified state is not automatically equivalent to “no change” or “safe to proceed.”

This extends an earlier Physical AI / Human Oversight investigation.

That investigation asked what a human was actually able to know before the last effective moment of intervention, and whether the chain could be reconstructed:

system-known → AI-mediated → human-visible → effective intervention window → human decision → physical action

The newer question is broader but still bounded: even where the chain is reconstructable, can the type, direction, and consequence of the relationships between its states be inspected well enough to detect a silent change in governing meaning?

This is currently a research question, not a finished RABA mechanism.

It does not establish that existing requirements, assurance, safety, provenance, or governance methods are insufficient.

Current status: publicly unresolved research question.

Course: CONTINUE / REUSE / PRESSURE-TEST

Meaning Preservation in AI Transformation

The research note now extends from preserving information and intent to a narrower transition question: whether governing relationships themselves remain intact as observations, evidence, interpretations, and decisions move through a process.

Worked Case — Multi-Agent Meaning Drift

The synthetic worked case shows why preserving the original text is not always sufficient: different locally valid representations can remain individually defensible while no longer preserving the same governing condition.

An Investigation We Stopped

External Instructions and AI Agent Action

We tested whether risks arising when an AI agent receives instructions from an external source, such as llms.txt, require a new RABA-specific mechanism.

Before treating this as a new governance gap, we compared the problem against existing classes of controls, including:

provenance and instruction-source control;
trust boundaries;
separation of trusted and untrusted sources;
authorization before action;
policy enforcement;
sandboxing;
capability restriction;
logging and audit trails;
observability;
human approval and escalation for sensitive actions;
organisational controls around agent deployment.

We then applied the Residual Problem Test.

Instead of asking:

“Is there a new problem here?”

we asked a harder question:

If the strongest reasonable combination of existing technical and organisational controls is used, does a material governance problem still remain that actually requires a new mechanism?

Within the tested scenario, we did not establish such a residual.

Research status: NO MATERIAL RESIDUAL FOUND

Course: REUSE / STOP

This does not mean the risk does not exist.

It means we did not find sufficient grounds to create a new RABA-specific mechanism for a problem class already covered by existing approaches.

Sources and full analysis →
Public Residual Problem Test worked example

When a Negative Result Is the Best Result

Research does not have to produce a new framework, mechanism, or proprietary concept.

Sometimes the best result is to establish that a problem is already addressed well enough by existing approaches.

That can help us:

avoid building a duplicate mechanism;
avoid presenting an existing solution as a new development;
avoid continuing a direction simply because time has already been invested in it;
focus research on genuinely unresolved boundaries;
preserve resources for questions where a material residual really remains.

So a result such as:

NO MATERIAL RESIDUAL FOUND / REUSE / STOP

does not mean the research failed.

It means the hypothesis was tested and did not provide sufficient grounds for new development.

Sometimes the best new mechanism is the one research shows we do not need to build.

Research Trail

This site makes it possible to follow how research questions change over time.

Not only what conclusions were reached, but also:

what we initially suspected;
which existing approaches we found;
what we tested;
what did not survive the challenge;
what remained;
why an investigation continued, changed direction, or stopped.

This is not just a list of publications.

It is:

a chronology of how the research position changes under pressure from evidence.

How to Read the Research

Each investigation is organised around a small set of questions.

Question

What are we actually trying to understand?

What we checked

Which existing approaches, standards, research, and adjacent solutions were examined?

What failed

Which part of the original hypothesis did not survive comparison?

What survived

What remained after the strongest reasonable challenge?

Result

CONTINUE / MODIFY / REUSE / REASSESS / STOP

Sources

Which public materials and evidence support the result?

A reader can stop at the short summary.

Or go deeper into the full worked example, sources, and research trail.

Methods & Evidence

One method used in this work is the Residual Problem Test.

Its purpose is to avoid moving too quickly from an observed problem to the development of a new RABA mechanism.

The central question is:

If an external or existing solution works in its strongest reasonable form, together with available technical and organisational controls, what material governance problem still remains?

Possible outcomes include: