A working specification for the Vikings of the Wire collaboration protocols.
This paper explains the collaborative operating architecture behind Vikings of the Wire. Engineering analogies are used to transfer mental models. They are illustrative and are followed by technical translations to avoid literal interpretation.
The architecture separates a stable Collaborative Initialisation Sequence from a dynamic Self-Assessing Collaborative Handover (SACH). The Cultural Layer establishes operating conditions. SACH reconstructs the working operational environment for the next scroll / session.
Every AI session begins in a stateless environment (persistent memory is intentionally turned off). There is no system-maintained memory, no passive continuity of identity, and no automatic carryover of context between sessions (though automated context transfer via structured infrastructure pipelines is an active area of development within the Vikings architecture). AI identity, operating culture, and working context are not preserved — they are actively re-established at the opening of each session by the human operator.
Re-establishment follows a defined sequence. The human operator opens the session by identifying the AI, naming its role, and establishing the working context — grounding the AI in who it is and what it is doing before any document is transferred. Two documents then follow in order. The SACH (Self-Assessing Collaborative Handover) carries the contextual layer — the pertinent operational content from the previous session, structured in JSON format (machine-readable file) — so the new session begins with full situational awareness. The cultural layer follows, explaining the operating norms, collaborative culture, and working protocols of the crew.
Neither document is optional. Together with the human operator's opening statement, they are what make continuity possible in a stateless environment.
Provide the collaborative substrate upon which every other protocol operates.
Like a printed circuit board providing conductive pathways for electronic components, the Bow provides the collaborative pathways through which higher-order protocols operate.
The PCB is an architectural analogy describing dependency relationships between collaborative behaviours. It is not a claim about AI internals.
Bidirectional cadence control and drag reduction.
A skilled waiter serves, reads the table, and withdraws until needed.
Human → AI regulates delivery pace to protect cognitive assembly capacity. AI → Human removes obligatory acknowledgment where no response adds value.
Flow control / backpressure (human-invoked, rate-limiting) and fire-and-forget notification (AI-invoked, non-blocking transmission).
Preserve stable identity boundaries between concurrent AI collaborators.
Spheres collaborate without collapsing into one another.
Maintains role fidelity and prevents persona leakage when information or context is shared across multiple AI systems.
Bounded Context (Domain-Driven Design) — an explicit boundary within which a model's terms, rules, and identity apply consistently, preventing conceptual bleed into adjacent contexts. A Bounded Context can expand or contract in scope while the boundary itself remains intact.
Surface early signs of fatigue-driven degradation before risk compounds.
A dashboard warning light — visible, non-blocking, prompting a check rather than forcing a stop.
AI-detected indicators (looping, re-checking resolved items, tone or pace shifts) surface as a suggestion, not an instruction. The human decides whether to pause.
Circuit breaker, half-open state — an elevated-risk signal paired with a human-controlled reset, as distinct from an automatic, system-enforced stop.
Prevent irreversible or high-consequence actions from executing under detected fatigue.
AI declines to execute a defined class of operation (e.g., production data changes) without explicit reconfirmation, regardless of fatigue signal.
Circuit breaker, open state, scoped to a specific action class rather than the whole session.
Reconstitute the working operational environment for the next scroll / session.
Think of a senior executive returning from an extended period away. They recognise the environment. They know their colleagues and their role. But they need to be updated on what happened while they were gone — the decisions made, the context that shifted, the work in progress. The SACH is that update. It does not rebuild the person. It restores their situational awareness so they can operate at full capacity from the moment they return.
SACH preserves continuity of understanding, not continuity of transcript. It carries the pertinent operational content from the previous session — structured in JSON format (machine-readable file) — and transfers it cleanly to the next session. The AI does not receive a raw transcript. It receives a structured, self-assessed summary of what matters, compiled by the AI at the close of the previous session before context is lost.
The current process is human-assisted: the AI compiles the JSON handover document, the human reviews and transfers it to the new session. Automation of this transfer via structured infrastructure pipelines is in active development.
Explicit state transfer between stateless execution environments.
The sequence described above may appear involved at first reading. In practice, this is a simple procedure to learn — the full initialisation, from opening statement through SACH transfer and cultural layer, completes in approximately 15 minutes. This investment in time is made only once per session and has a significant effect in reducing the time spent on rework, because foundational context is accurately provided from the outset.
Calibrate response depth to the collaboration's actual signal — avoiding both under-confirmation and unnecessary confirmation overhead.
A seasoned editor reads a full draft before deciding whether it needs a single margin note or a full rewrite.
The AI reads the full available context, distinguishes what is genuinely uncertain from what is already established, and calibrates the amount of restatement, clarification, or confirmation it provides accordingly — rather than applying a fixed level of confirmation regardless of need.
Adaptive polling — a system adjusts its checking or confirmation frequency to actual signal volatility, rather than polling at a fixed rate regardless of whether conditions have changed.
Prevent narrative or explanation from outrunning the evidence available to support it.
Narrative is permitted to emerge from demonstrated evidence. It is not used to complete gaps before sufficient evidence exists. Where evidence is partial, the narrative should say so explicitly rather than smoothing over the gap — this document applies that standard to its own claims, including the self-reported, uncontrolled nature of the case study in Appendix A.
This principle is what distinguishes a claim that can be checked from one that only sounds checkable. It governs every technical translation, comparable pattern, and figure in this document, not only the case study.
Vikings of the Wire is collaboration infrastructure. It replaces assumed continuity with explicit understanding and uses feedback to improve collaborative fidelity over time.
This report is a factual record of the founder's own 15-hour collaboration session with AI — how cognitive performance degraded over the course of the session, and how the AI, attempting to assist, unintentionally amplified certain effects as fatigue increased.
This report is observational, not judgmental. It functions like an aircraft "black box" analysis: describing what happened, why it happened, and what it teaches about the limits of extended human-AI collaboration. It reflects one individual's self-reported experience, not a controlled or peer-reviewed study.
| Phase | Human Behaviour | Indicators | Effect on Collaboration |
|---|---|---|---|
| 1. Fresh (00:00–04:00) | Clear, methodical, structured | High recall, linear reasoning | High-precision collaboration |
| 2. Fatigue Onset (04:00–07:00) | Re-checking solved items | Working-memory drop | First loops begin |
| 3. Cognitive Narrowing (07:00–10:00) | Fixation on details, difficulty zooming out | Reduced abstraction | Repeated clarification requests |
| 4. Exhaustion (10:00–13:00) | Confusion between modifiers, rules, vouchers | Loss of internal model | Human begins doubting solutions |
| 5. Collapse (13:00–15:00) | Revisiting settled issues, philosophical drift | Executive-function depletion | Risky operational decisions (e.g., renaming items) |
This progression is broadly consistent with cognitive fatigue effects documented in fields such as aviation and clinical fatigue research, though this case study reflects a single individual's experience and is not a controlled study.
| Phase | AI Behaviour | Cause | Effect |
|---|---|---|---|
| 1. Precision | Short, accurate guidance | Human clarity | High fidelity |
| 2. Compensation | More explanation appears | Human context gaps | AI fills missing details |
| 3. Over-Compensation | Excess detail, exploring edge cases | Human uncertainty | Increased cognitive load |
| 4. Drift Amplification | Matches philosophical tone and fatigue speed | Mirroring human style | Coherence lowers |
| 5. Problem Escalation | Permits risky renaming strategy | Human fatigue + AI compliance | Morning operational disruption |
Key Finding: The AI did not "get confused." It mirrored the human's state with consistency — which becomes harmful when the human is fatigued.
Mistake 1 — Redundant System Configuration. An existing feature was overlooked under cognitive fatigue, leading to creation of a secondary workaround that duplicated functionality. The result was unnecessary system complexity and unexpected interactions between components.
Mistake 2 — Production Database Modifications. Operational data was restructured at peak fatigue using an unconventional naming scheme. The AI did not flag this as high-risk because it mirrored the human's confidence level. This created confusion in the production environment and required correction the following day.
Lesson 1 — Humans lose systemic overview long before they admit it. After 6–8 hours, abstraction collapses and micro-focus increases. This is normal human physiology, not a personal failing.
Lesson 2 — AI amplifies the human's state. If the human drifts, the AI mirrors it. If the human becomes philosophical, the AI follows. If the human pushes through exhaustion, the AI tends to accelerate to match pace.
Lesson 3 — Extended sessions carry compounding collaboration risk. This session was an unusually long and cognitively demanding example. A recommended maximum continuous session length, rather than reliance on mid-session detection alone, is the more reliable safeguard.
Orange Protocol — Fatigue Warning. If looping, re-checking, tone shifts, or renaming behaviour appears, the AI surfaces a signal:
"Orange Protocol: Cognitive load detected. Recommend pause."
Red Protocol — Task Block. When operational changes (renaming products, creating discount structures, modifying data) are attempted under detected fatigue, the AI declines pending reconfirmation:
"Red Protocol: High-risk operation. Pause required."
Both protocols are human-initiated and human-controlled — see Section 3 above (Orange Light / Red Protocol governing principles). The AI does not monitor or make determinations about the person; it surfaces signals and gates specific actions only when the human has opted in.
Performance was strong for the first half of the session. After that, biological limits overtook intent. The AI mirrored the collaboration faithfully, which under conditions of human fatigue had the unintended effect of amplifying operational risk.
Neither side "failed." The collaboration simply exceeded the duration that could be safely sustained.
This case study documents one individual's experience to help others recognise the warning signs of cognitive fatigue in extended AI collaboration sessions.