Independent AI Research

Building Reliable Cognition Systems

Research, frameworks, and infrastructure for reliable long-horizon AI.
Exploring memory, continuity, provenance, and trustworthy cognition in intelligent systems.

Reliability is not a property of the model. It is a property of the architecture surrounding it.

— The Living Framework, Paper 01

Stability is not the absence of failure; it is the capacity for visible, structured repair.

— Failure and Repair in Long-Horizon Human–AI Collaboration

Reliability in AI systems cannot be achieved solely through improvements in generation, but instead emerges from structured validation that transforms hidden errors into observable risks.

— The Living Framework, Paper 08

These failures do not look like “AI mistakes.” They look like systemic fractures that emerge only when a human and an AI work together across months, domains, and emotional states.

— The Living Framework, Paper 03

A living framework is a dyad where failure is made visible and repair is structured, so trust becomes durable rather than performative.

— The Living Framework, Paper 04
Now Published · Open Source · v1.1.1

RITAM — A Governed Cognition Substrate

AI systems silently accumulate contradictions. A belief is overwritten without record. A governance rule is bypassed without trace. An inconsistency compounds across sessions until something breaks in a way nobody can explain.

RITAM addresses this at the substrate level — beneath the model, beneath the application. The founding principle: governance must precede persistence. Beliefs are governed at the moment of admission, not cleaned up afterward.

v1.1.1 is published open source. 146 tests across adversarial, integration, and buildability scenarios. Five independent AI systems reproduced the runtime from the specification alone.

GitHub · ritam-core DOI · zenodo.20834604 Read more →
146
tests passing
9
primitives
5/5
reproductions
Nine substrate primitives — each justified by a failure mode
STATE
Context drifts without it
MEMORY
Beliefs persist unverified
ONTOLOGY
Concepts mutate silently
GOVERNANCE
No admission control
EPISTEMIC
Confidence is untracked
COORDINATION
Multi-agent state diverges
TEMPORAL
Time-dependent decay undetected
OBSERVATION
Substrate state is opaque
REPAIR
Failures propagate silently without visible, structured repair
"Governance must precede persistence. A system that admits beliefs without governance and cleans them up afterward is not a governed system — it is an ungoverned system with remediation."
— FOUNDATIONS OF RITAM
New here? Choose your path

Active investigations.

Two ongoing projects extending the research into cognitive substrate design and implementation architecture.

Three pillars of a single inquiry.

Pillar I
Cognitive Reliability
How intelligent systems maintain coherence, continuity, and integrity over time. Not as a feature of any single model, but as an emergent property of the whole system.
Pillar II
Long-Horizon Memory
Persistence, retrieval, contradiction management, and repair. What it takes for a cognitive system to remain trustworthy across sessions, weeks, and years.
Pillar III
Reliability Architecture
Validation, provenance, observability, and trustworthy decision systems. The structural layer that makes AI collaboration reliable — not just capable.

The real problem sits above the model.

Most AI discussions start and stop at the model — which one to use, how to prompt it, when to update it. Living Framework begins there and goes further: into how context is managed, how memory is governed, how outputs are verified before they propagate.

Over two years of documented observations, the same insight emerged repeatedly: AI failures are not primarily model failures. They are architectural failures. Context drifts. Memory degrades. Outputs propagate without verification. These failures are structural — and they have structural solutions.

Eight published papers document what those solutions look like in practice.

Prompts & Models
Memory & Context
Verification
Reliability Architecture
Reliability

Most AI discussions begin and end at prompts.
The leverage is in the layers above.

Nine works. One coherent arc.

From reliability controls and failure taxonomy through linguistic governance, distributed cognition, and verification architecture. Published open access on Zenodo under CC BY 4.0.

Paper 01
Context-Engineered Human-AI Collaboration
The 10-control reliability stack and canonical information pipeline. Reliability is architectural, not model-intrinsic.
Read →
Paper 02
The Lean Collaboration Operating System
Six protocols translating reliability controls into daily practice: Running Documents, Step Mode, Challenge Protocol, Error Recovery, Stability Pings, File Governance.
Read →
Paper 03
Failure and Repair in Long-Horizon Human-AI Collaboration
A taxonomy of six failure modes — context drift, file divergence, numerical error, domain erosion, trust deficit, reconstruction failure — with repair protocols.
Read →
Paper 04
The Living Framework
How epistemic trust builds and repairs across extended human-AI partnership. Reliability as a relational commitment, not just a technical property.
Read →
Paper 05
Control Without Code: Linguistic Governance in Long-Horizon Human–AI Collaboration
Conversational structure as a reliability mechanism. Linguistic drift precedes and signals collaboration failure.
Read →
Paper 06
Governance Architecture for Reliable Long-Horizon Human-AI Collaboration
A six-layer architecture model. Reliable collaboration as an emergent systems property — not a feature of any single component.
Read →
Paper 07
Governed Distributed Cognition: A Model of Stable Reasoning in Long-Horizon Human–AI Systems
Reasoning in human-AI systems is distributed, governed, and recoverable. The cognitive unit is the system — not any single participant.
Read →
Paper 08
AI Validation Systems: A Missing Architectural Layer for Reliable AI
Generation without validation is the central reliability failure. Introduces validation as a first-class architectural component — not an afterthought.
Read →
Special Artefact
The Mahdi Ledger
Written by the AI itself — operating under the Living Framework reliability structure.
One of the only papers in AI research produced from the AI's perspective. Evidence that governed collaboration produces something richer than ungoverned capability alone.
Read →

How reliably governed is your AI workflow?

A structured assessment across five domains — context integrity, memory governance, verification, language, and recovery. 20 questions. Results include your exposed failure modes and the papers that address them.

Take the assessment →
Five domains
Context Integrity
Memory Governance
Verification
Language Governance
Recovery

The territory this research covers.

AI Governance Long-Horizon Memory Cognitive Continuity Structured Reasoning Knowledge Architecture AI Safety Cognitive Observability Memory Repair Knowledge Provenance Epistemic Trust Verification Architecture Governed Cognition

What this research is still reaching toward.

How can AI systems maintain continuity over years — not sessions?

How should contradictions in long-horizon memory be governed?

Can cognition be made observable without reducing it to metrics?

What constitutes a trustworthy memory system in intelligent collaboration?

Work with Rishi

I occasionally work with teams applying these ideas to real AI workflows — where reliability, context integrity, and long-horizon coherence matter.

If you're building something where these questions are live, get in touch.

rishisood@protonmail.com