Research Overview

One question. Eight papers. An ongoing investigation.

This page maps the full research programme — what was studied, what was found, what remains open, and where the work is going next.

The Research Question
Central investigation

How do humans and AI systems work together reliably over extended periods of time — and what structural conditions make that reliability possible?

This question is deceptively simple. It does not ask which model is best, or how to write better prompts. It asks what makes the entire collaboration system — human judgment, AI reasoning, shared memory, governance structures — remain coherent and trustworthy across weeks, months, and hundreds of hours of work. That question turns out to be both harder and more important than most AI research currently addresses.

Research Areas

Six themes of investigation

The research spans six interconnected areas. No theme stands alone — each illuminates a different aspect of the same underlying problem.

AI Reliability
Why AI systems fail at predictable structural points, and what architectural conditions prevent those failures. Reliability as a systems property, not a model property.
Context Engineering
How context degrades over extended collaboration and what structural mechanisms keep it stable. The design of context as a governed resource rather than an assumed constant.
Memory Systems
Persistent external memory as a reliability mechanism. How decisions, corrections, and shared knowledge are preserved across sessions without degradation.
Verification Systems
The missing architectural layer in most AI deployments. Generating outputs and verifying them are two distinct operations — both necessary, rarely treated as such.
Governance Architecture
Layered structural models for reliable long-horizon collaboration. How multiple governance mechanisms interact to produce reliability as an emergent property.
Human-AI Collaboration
The distributed cognitive system: human judgment, AI reasoning, and governed shared memory. How trust builds, fractures, and repairs over time.
Major Findings

What the research has established

Five findings that hold across the full body of work. Each is supported by multiple papers and grounded in documented observation, not theoretical inference.

F1
AI reliability is architectural, not intrinsic to the model.
The same model can produce reliable or unreliable collaboration depending on the structural conditions surrounding it. Improving the model does not resolve failures that are caused by the absence of memory governance, verification, or context management. These are architecture problems, not capability problems.
F2
Failure modes in AI collaboration are predictable, recurring, and structurally distinct.
Failures do not occur randomly. Six failure modes — Context Drift, File Divergence, Numerical Error Propagation, Domain Erosion, Trust Deficit, and Reconstruction Failure — each follows a recognisable trajectory with identifiable precursors. This predictability makes prevention and repair possible.
F3
Language is a governance mechanism — and linguistic drift is an early signal of collaboration failure.
How human and AI communicate within a session is not merely a style choice. Specific language patterns predict collaboration stability. When the language of the collaboration begins to shift — instructions become vague, corrections become implicit, shared vocabulary erodes — structural failure typically follows. The signal appears in language before it appears in outputs.
F4
Reliable collaboration is a distributed cognitive achievement, not a model capability.
The relevant unit of analysis is not the human or the AI individually, but the governed system they form together. Human judgment, AI reasoning, and shared external memory — operating under explicit governance controls — constitute a single distributed cognitive system. Reliability is a property of the system, not of any participant.
F5
Current AI systems lack a dedicated verification layer — and this is the central unresolved reliability problem.
Generation and validation are treated as one operation when they are two. AI systems produce outputs; they do not natively verify them. In the absence of a structured verification layer, errors propagate unchecked. This is not a model limitation. It is an architectural omission — and it is addressable.
Research Timeline

How this work developed

The research did not begin with a theory. It began with observations — and the papers followed.

2024 — Early work
First observations and foundational papers
Extended AI collaboration revealed recurring patterns — context loss, file divergence, trust erosion — that no existing framework explained. The first four papers formalised these observations: the reliability architecture (A1–A10), the operational protocols (LC-OS), the failure taxonomy, and the philosophy of sustained partnership.
Paper 01 Paper 02 Paper 03 Paper 04 Mahdi Ledger
2025 — Deeper mechanisms
Language, architecture, cognition, and verification
Papers 05 through 08 went deeper into specific mechanisms. Linguistic governance showed how language functions as a reliability layer. The six-layer architecture model showed how governance mechanisms interact. Distributed cognition theory provided the theoretical grounding. Paper 08 identified the missing verification layer as the central outstanding problem.
Paper 05 Paper 06 Paper 07 Paper 08
2026 — Active
Architecture development and continued investigation
RITAM is doing foundational substrate research into the nine architectural primitives a governed executable cognition system requires. NIYOM is live: a four-stage AI verification harness at v1.0.0 — Plan, Execute, Verify, Repair. Failure documentation continues. The convergence of industry terminology (context engineering, reliability, agent memory) with concepts documented in the papers confirms the research was tracking the right problem early.
Open Questions

What the research has not yet resolved

These are genuine open questions — not rhetorical. They define the next phase of investigation.

How does the reliability architecture scale to multi-agent systems?
The current framework addresses single-AI collaboration. In systems with multiple AI agents working in parallel or in sequence, the failure modes multiply and interact. Context drift in one agent may cause downstream failures in another. The architecture for this is not yet understood.
Can any part of the verification layer be automated without reintroducing the original reliability problem?
Paper 08 identifies the need for a verification layer. But if that layer is itself an AI system, it introduces a recursive reliability problem. Some verification requires human judgment. The boundary between what can be automated and what cannot is not yet clearly drawn.
What are the reliable early warning signals for Reconstruction Failure before it becomes terminal?
Reconstruction Failure is the most costly failure mode — it results in complete loss of collaborative progress. But the precursors are subtle. By the time the failure is visible, recovery is already expensive. Earlier, more reliable indicators are needed.
How do memory architectures degrade differently across short, medium, and long-horizon collaborations?
The research covers long-horizon collaboration (weeks and months). The failure modes may manifest differently, or at different rates, in shorter engagements. Whether the reliability controls are necessary or sufficient across all collaboration timescales is not yet established.
Research Roadmap

Completed, active, and planned

This research programme is ongoing. What follows is an honest account of where the work currently stands.

Completed
Papers 01–08 (published, open access)
Mahdi Ledger (published)
Failure taxonomy (6 documented modes)
10-control reliability stack (A1–A10)
Six-layer governance architecture
Distributed cognition model
Linguistic governance analysis
Active
RITAM — reliability architecture development
NIYOM — verification harness (v1.0.0, in development)
Continued failure episode documentation
Verification layer design (following Paper 08)
Multi-agent reliability investigation (early stage)
Under Investigation
Assessment diagnostic engine — AI reliability scoring
Failure mode extension — additional patterns
Terminology translator — framework-to-industry mapping
Research timeline documentation
Explore the Research

Where to go next

Start with the papers if you want the primary research. Start with the Failure Library if you want something immediately practical. The pages below extend the research in specific directions.

Read all eight papers → Browse the Failure Library → Terminology Translator → Early Observations → Research Roadmap → About this research → Research Advisory →