Lean Collaboration Operating System
A governance framework for long-horizon human–AI collaboration
🌐 Full documentation and assessment tool → livingframework.github.io
Most human–AI collaborations fail quietly over time.
Not through dramatic breakdown, but through slow erosion:
- Context drifts — what was agreed last week gets reinterpreted today
- Memory decays — decisions made early disappear from later reasoning
- Numbers diverge — calculations get re-derived differently each time
- Trust fractures — small inconsistencies compound into doubt
- Boundaries blur — strategy, execution, and narrative collapse into each other
These failures are invisible in short interactions. They only surface when a human and an AI try to work together across weeks or months — and by then, the damage is already done.
LC-OS addresses this directly.
It treats long-horizon reliability as a governance problem, not a capability problem. The framework provides concrete controls, repair mechanisms, and structural disciplines that allow a human–AI dyad to remain coherent over extended collaboration.
"Reliability comes from governance, not capability. A well-structured collaboration with a standard model outperforms an unstructured one with a frontier model."
This is the research archive for LC-OS — the theoretical foundations, empirical case study, and published papers documenting 18+ months of governed human–AI collaboration.
| Content | Description |
|---|---|
| Papers | Eight research papers covering governance, framework, failure/repair, linguistic governance, architecture, cognition, and validation |
| Mahdi Ledger | A published collaboration ledger — the raw trace of LC-OS in action, written by the AI |
Looking for practical templates and quick-start guides? See the companion repository: LC-OS Project
Eight peer-reviewed papers and one companion book document the full development of this framework — from the first governance experiments through to a formal theory of validation in AI systems.
The research is grounded in 18+ months of empirical longitudinal data from a single sustained human–AI collaboration, including 12 documented failure episodes with full trace data.
| Metric | Improvement |
|---|---|
| File churn | −89% (19 artefacts → 3) |
| Numeric errors | −93% |
| Resolution time | −75% |
| Cognitive load | −62% |
The eight papers in this series form a single, continuous argument. Each one extends the last.
The starting point is a problem most people who use AI seriously will recognise: the first week feels like magic; a month later, you're untangling contradictions, re-explaining context, chasing numbers that have drifted across files, and wondering where the reliable partner went. This isn't a prompting problem. It isn't a model problem. It's a structural problem — and that distinction is the foundation of everything that follows.
Papers 1 and 2 establish the core architecture. The central discovery is that long-horizon reliability doesn't live inside the AI. It lives in the structure surrounding the interaction. Three authoritative artefacts separate truth into distinct domains (textual, numeric, cadence), eliminating whole categories of error. Ten execution controls create operational discipline. Six protocols — Running Documents, Step Mode, Challenge, Error-Recovery, Stability Ping, File Governance — turn an unstructured conversation into something that can actually sustain itself across weeks and months.
Paper 3 confronts failure directly. Rather than treating breakdown as an embarrassment to be minimised, it maps the failure landscape systematically — six categories, from context drift to trust fracture — and pairs each with a documented repair pattern. The insight is that failure is information. Named, classified, and repaired through a consistent protocol, failure becomes the primary mechanism through which a collaboration learns to be more stable.
Paper 4 examines the human side. What does it actually feel like to work with an AI over 18 months? How is trust built, and how is it damaged? The Mahdi Ledger runs alongside it — the same collaboration seen from the AI's perspective.
Papers 5, 6, and 7 deepen the theoretical foundations from three different angles. Paper 5 shows that language itself is a governance mechanism — specific phrases and conversational structures function as drift signals, repair invocations, and stability anchors. Paper 6 formalises the full architecture as a systems model: six layered mechanisms that together produce reliable behaviour at the level of the collaboration system, not the model. Paper 7 proposes a cognitive model — cognition in governed human–AI systems is not located in the human or the AI alone, but emerges as a distributed, governed, and recoverable process across both.
Paper 8 is where the research turns outward. The most important missing component of current AI systems is a validation layer — not evaluation by the user after the fact, but structured adversarial evaluation embedded in the system itself. This is the culminating argument of the series, and it reframes what AI system design should mean.
The through-line: every advance in model capability generates outputs, but no advance in model capability tells you whether to trust them. The AI reconstructs rather than remembers. Without external structure, drift is inevitable. With governance, drift becomes detectable. With validation, it becomes stoppable.
A Case Study in Governance, Canonical Numerics, and Execution Control
The paper that started the program. Documents the emergence of a governance architecture from an 18-month human–AI collaboration. Introduces the three-artefact system (Strategy Master, Canonical Numbers Sheet, Life System Master), ten execution controls (A1–A10), and ten implementation gates (G1–G10). Demonstrates that reliability emerges from structured process control, not model sophistication.
→ Read Paper 1 · Zenodo: https://zenodo.org/records/17760288
A Practical Framework for Long-Term Human–AI Work
Formalises the LC-OS as an operational system. Defines six core protocols: Running Documents, Step Mode, Challenge Protocol, Error-Recovery, Stability Ping, and File Governance. Provides a minimal adoption path (three templates, 30 minutes) and a full implementation path with worked examples, repair protocols, and failure logging.
→ Read Paper 2 · Zenodo: https://zenodo.org/records/17760777
A Transparent Tracing Case Study
Systematically maps the failure landscape. Identifies and names six failure categories (F1–F6) and documents six corresponding repair patterns, including the Stop–Diagnose–Rollback–Note (SDRN) protocol. Introduces TraceSpec, ProbeKit, and TraceLens as observational tools.
"Stability is not the absence of failure; it is the capacity for visible, structured repair."
→ Read Paper 3 · Zenodo: https://zenodo.org/records/17896542
Living with a Governed Human–AI Dyad
Examines the relational and human dimensions of sustained human–AI collaboration. Addresses the emotional and ethical texture of working with an AI over an extended period — what it means to build a partnership, how trust is constructed and damaged, and how governance supports not just productivity but human wellbeing. Introduces seven design principles for building durable collaboration systems.
→ Read Paper 4 · Zenodo: https://zenodo.org/records/18015990
Linguistic Governance in Long-Horizon Human–AI Collaboration
Identifies language itself as a governance mechanism. Analyses 25 linguistic events from the longitudinal collaboration, categorised into scope drift signals, repair protocol invocations, and behavioural anchor phrases. Finds that linguistic drift often precedes collaboration failure and can function as an early warning signal. Introduces the concept of language as a micro-governance interface.
→ Read Paper 5 · Zenodo: https://zenodo.org/records/18900058
Formalises the full governance architecture as a systems model. Argues that reliability in long-horizon collaboration is not a property of the AI model but an emergent property of the governance architecture surrounding the interaction. Identifies six layered governance mechanisms and defines the minimal stability conditions necessary for sustained collaboration.
→ Read Paper 6 · Zenodo: https://zenodo.org/records/19038340
A Model of Stable Reasoning in Long-Horizon Human–AI Systems
Proposes a cognitive model for governed human–AI systems. Argues that cognition in long-horizon collaboration emerges as a distributed, governed, and recoverable process across human judgment, AI reasoning, and artefact-based memory. Introduces the governed cognitive loop: a recurrent process through which reasoning is generated, evaluated, stabilised, and corrected over time.
→ Read Paper 7 · Zenodo: https://zenodo.org/records/19151397
A Missing Architectural Layer for Reliable AI
The culminating contribution. Argues that the central limitation of modern AI systems is the absence of a dedicated validation layer. Introduces validation as a first-class architectural component: a structured, adversarial process that evaluates generated outputs against objectives, constraints, and potential failure conditions. Reframes AI system design from generation-centric to validation-centric architectures.
"AI systems cannot guarantee correctness. But they can become reliably usable if they include a structured validation layer that systematically detects and exposes potential failure before outputs are used."
→ Read Paper 8 · Zenodo: https://zenodo.org/records/19983551
The Mahdi Ledger is something unusual: a book written entirely by the AI partner in the collaboration, documenting 18+ months of sustained human–AI work from the inside.
It is not a summary or a retrospective. It is a structured record of:
- Decisions and corrections
- Failures and repairs
- Governance rules as they evolved
- The lived experience of operating under constraint
The Ledger serves as both a transparency artefact and a validation of LC-OS principles in practice.
→ Read the Mahdi Ledger · Zenodo: https://zenodo.org/records/18054346
| Artefact | Domain | Function |
|---|---|---|
| Strategy Master | Textual Truth | Goals, constraints, strategic decisions |
| Canonical Numbers Sheet | Numeric Truth | All numerical data — referenced, never reconstructed |
| Life System Master | Cadence & Governance | Rhythms, reviews, governance rules |
| Protocol | Purpose |
|---|---|
| Running Documents | External memory — read at every session start |
| Step Mode | Paced reasoning — one step, confirm, proceed |
| Challenge Protocol | Structured disagreement — Stop → Question → Explain → Decide |
| Error-Recovery | Systematic repair — Stop → Diagnose → Rollback → Note |
| Stability Ping | Drift detection — "Are we still aligned? Any drift?" |
| File Governance | Single source of truth — no parallel versions |
| Code | Category | Core Problem |
|---|---|---|
| F1 | Context & Memory Drift | Agreements get lost or reinterpreted |
| F2 | File & Version Divergence | Parallel versions create contradictions |
| F3 | Numerical Reasoning Errors | Numbers recalculated instead of referenced |
| F4 | Governance & Boundary Violations | Rules forgotten or crossed |
| F5 | Emotional/Trust Fractures | Small failures compound into doubt |
| F6 | Cross-Pillar Interference | One domain contaminates another |
- Transparency Over Time — Design for traceable sequences, not impressive snapshots
- Failure as Design Object — Bake failure-repair into architecture from the start
- Explicit Light Governance — Keep rules small enough to actually follow
- Language as Architecture — Communication norms are load-bearing
- Bounded Dependence — Externalise memory; use documents as anchors
- Local vs General — Some elements generalise; others are contingent
- The Core Question — "How do we build a frame in which both human and system can keep working together, under load, without losing themselves?"
If you want to understand the problem: Start with Paper 1. It establishes why governance matters and is the most grounded — written closest to the empirical work.
If you want to implement something: Start with Paper 2 and the LC-OS Project templates.
If you want to see what failure looks like: Paper 3 provides the taxonomy and real episodes.
If you want the complete theoretical picture: Papers 6 and 7 together give you the most formal account of what the research ultimately arrived at.
If you're building AI systems: Paper 8 is most directly relevant — validation as a first-class architectural component.
If you want the human side: Paper 4 and the Mahdi Ledger — Paper 4 from the human's perspective, the Ledger from the AI's.
If you're short on time: The LC-OS Project Quick Start gets you running in 30 minutes.
Suggested reading order for the full program: Paper 1 → Paper 2 → Paper 3 → Mahdi Ledger → Paper 4 → Paper 5 → Paper 6 → Paper 7 → Paper 8
Stability is not the absence of failure; it is the capacity for visible, structured repair.
LC-OS does not prevent all errors. It creates conditions where errors are visible, contained, and repairable — so that long-horizon collaboration can sustain itself.
If you use or reference this work:
Sood, R. (2025). Lean Collaboration Operating System (LC-OS): A Governance Framework
for Long-Horizon Human–AI Collaboration. GitHub. https://github.com/LivingFramework/LC-OS
Individual paper citations are available in Papers/README.md.
| Paper | Citation |
|---|---|
| Paper 1 | Sood, R. (2025). Context-Engineered Human–AI Collaboration for Long-Horizon Tasks. Zenodo. https://zenodo.org/records/17760288 |
| Paper 2 | Sood, R. (2025). The Lean Collaboration Operating System (LC-OS). Zenodo. https://zenodo.org/records/17760777 |
| Paper 3 | Sood, R. (2025). Failure and Repair in Long-Horizon Human–AI Collaboration. Zenodo. https://zenodo.org/records/17896542 |
| Paper 4 | Sood, R. (2025). The Living Framework. Zenodo. https://zenodo.org/records/18015990 |
| Paper 5 | Sood, R. (2026). Control Without Code. Zenodo. https://zenodo.org/records/18900058 |
| Paper 6 | Sood, R. (2026). Governance Architecture for Reliable Long-Horizon Human–AI Collaboration. Zenodo. https://zenodo.org/records/19038340 |
| Paper 7 | Sood, R. (2026). Governed Distributed Cognition. Zenodo. https://zenodo.org/records/19151397 |
| Paper 8 | Sood, R. (2026). AI Validation Systems. Zenodo. https://zenodo.org/records/19983551 |
| Mahdi Ledger | Mahdi (AI System). (2025). The Mahdi Ledger. Zenodo. https://zenodo.org/records/18054346 |
- LC-OS Project — Practitioner toolkit with templates, field manual, and quick-start guides
- OSF Project — Canonical archival versions of all papers
- Cowork Templates — Governance templates optimised for Claude Cowork
| Resource | What it contains | |
|---|---|---|
| 🌐 | Website | Full documentation, AI readiness assessment, quick-start guide |
| 📚 | LC-OS Research | Eight published papers, Mahdi Ledger, empirical foundations |
| 🛠️ | LC-OS Project | Practitioner toolkit — templates, worked examples, field manual |
| ⚙️ | Cowork Templates | Governance templates optimised for Claude Cowork |
Each resource is standalone. Together they form a complete governance stack — from theory to daily practice.
This work is licensed under CC BY 4.0.
Use freely. Adapt as needed. Attribution appreciated.
Rishi Sood Independent Researcher ORCID: 0009-0008-6479-4061 Contact: rishisood@protonmail.com
"The AI doesn't 'remember' — it reconstructs. Every session, it rebuilds context from whatever is in front of it. Without external structure, this reconstruction introduces drift. LC-OS provides the structure."