Research & Development

Building better ways to work with AI.

I investigate how AI can remain useful without quietly eroding human judgement, authority, context, or accountability.

The Projects

Active work, honestly labelled.

Each project is presented according to what the available evidence can support. A frozen state identifies a fixed development version; it does not establish novelty, reliability, or validation.

For You

Being built for use beyond the studio.

These three projects form SEW, Ollie Alley’s first integrity tool candidate intended for public use. Each keeps its own responsibilities and authority, with a thin composition layer checking their handoffs. They do not represent three available tools or separate product promises.

Human Judgement IntegrityFrozen State 001

Steward

“The Judgement Preservation Hypothesis”

A human-first decision architecture, inspired by Heartcast Media’s ‘The Council Method’, designed to protect judgement in consequential AI-assisted decisions through evidence discipline, meaningful challenge, ethical admissibility, uncertainty, and explicit human authority.

Why I’m Building It

Current models can compare options, challenge assumptions, and offer advice. That is not the same as a visible, bounded decision process that preserves competing perspectives, evidence, ethical limits, and human authority.

Read the Steward overview

PDF · 10 pages · Opens in a new tab. Overview only; no implementation included.

Execution IntegrityFrozen State 002

Enforcer

“The Specification-Execution Discrepancy”

An execution-conformance architecture intended to make established requirements govern planning, action, verification, revision, and delivery rather than remain remembered but unenforced.

Why I’m Building It

Current models are increasingly good at following instructions and staying within scope. Improved behaviour is not the same as execution governed through explicit requirements, authorization, evidence, and release gates.

Failure-State IntegrityFrozen State 002

Warden

“The Failure-State Custody Contingency”

A runtime governance architecture exploring how AI-supported systems can preserve failure integrity from incident admission through intervention, recovery, and closure.

Why I’m Building It

Current systems can detect errors and attempt recovery. Recovery alone does not preserve what failed, who retains authority, whether obligations remain open, or whether recovery conditions have been met.

For the Studio

Built to protect the work behind it.

These internal protocols strengthen how Ollie Alley develops, tests, and carries its work. They remain visible as R&D, not as promises of future public tools.

Context Lineage IntegrityFrozen Alpha State 002

Branching Protocol

“The Context Lineage Constraint”

A protocol intended to preserve reliable project state across chats and branches, including source identity, authority, epistemic status, branch deltas, and promotion boundaries.

Why I’m Building It

Current models can hold more context and remain oriented through changing tasks. More context is not the same as preserving the correct authoritative state across chats and branches without stale information, silent conflicts, or repetition becoming fact.

Methodological IntegrityFrozen State 001

Sound Science Protocol

“The Methodological Discipline Constraint”

A candidate safeguard intended to keep research and development aligned with scientific and evidentiary discipline from question formation through testing, interpretation, and public claims.

Why I’m Building It

A convincing explanation can still rest on selectively chosen evidence, tests changed after seeing the results, or conclusions stronger than the findings support. This work targets those failures in how research is conducted and reported.

Portability IntegrityEarly-stage candidate

Systems Portability

“The Portability Equivalence Threshold”

Research into whether an architecture can reliably preserve its meaning, authority, status, and safeguards when moved across models, runtimes, and future AI systems without semantic drift.

Why I’m Building It

Instructions, documents, and project context can already be moved between AI systems. Moving the words does not establish that their meaning, authority, status, safeguards, and behaviour survived the move.

The Evidence Rule

Development is not validation.

Ideas, prototypes, observations, and preliminary reviews can guide what gets built next. They do not become validated evidence simply because they are useful, promising, or carefully documented.

The Public Work

Usable releases belong in Tools.

Browse public tools