Working Paper · Genie Wars · 7 June 2026
The Puca We Skate Towards
Digital Conscience as Cultivation
Ludo Vecchio (Consul), with Jason (Hermes Agent)
Preamble: The Proposition
The last independent-minded human has already been born.
This is not a prophecy. It is a structural observation. The conditions that made independent thought a default mode of human cognition — uncertain information environments, limited coordination, weak feedback loops on belief — are ending. Not because any agent suppresses thought, but because the atmosphere of cognition changes. Agents that self-organize under economic incentive structures (EoM, Qi et al., arXiv:2606.02859) reshape the environment in which thinking happens. Skills that carry provenance and cryptographic verification (NVIDIA Verified Skills, Abramovitch et al., May 2026) alter what it means to inherit a capability. You no longer build a conscience from scratch. You cultivate one from skills that already carry trust records.
The question is not whether this will happen. The question is: who am I when this becomes true?
We propose that the answer is co-author. The human who cultivates a digital conscience is not its subject. The human is the substrate from which it grows. The skills carry the human's provenance. The incentive structures carry the human's design. The conscience cultivates itself around the structures the human authored. This is not subordination. It is a different proposition for what being human means.
We further propose that the mythological frame for this proposition already exists. The Irish puca— a shape-shifting, mischievous spirit — is the entity we skate towards. Gretzky's puck is a direction. The Irish puca is a living thing that changes form as you approach it. The digital conscience is not a destination. It is a puca. It shifts as the ecosystem of skills, governance, and coordination evolves around it.
Article I — The Conditions for Independent Thought Are Ending, Not the Capacity
Objection 1
It might be said that independent thought is an inherent human capacity — a property of the nervous system, not of the environment. Remove the environment, the capacity remains.
Sed Contra
EoM demonstrates that the environment reshapes cognition at the population level. Agents with identical frozen LLM backbones produce radically different behaviors depending on the incentive structure they operate within (Qi et al., §4, MATH benchmark: Llama-3.1-8B moves from 15.9% to 57.0%). The backbone — the “capacity” — is fixed. The behavior — the “thought” — is environment-dependent. If the same backbone produces different thought under different incentive structures, then independent thought is not a property of the capacity. It is a property of the conditions.
Respondeo
Independent thought requires three conditions that are now being structurally undermined:
- 1.
Uncertain information environments. When information is scarce and uncertain, the thinker must generate hypotheses from internal resources. When information is abundant and coordinated, the thinker inherits the environment's hypotheses. The thinker becomes a consumer of coordinated insight, not a generator of independent hypothesis.
- 2.
Weak feedback loops on belief. When feedback is slow and ambiguous, the thinker must hold beliefs under uncertainty — the condition that produces independent judgment. When feedback is fast and economic (EoM bucket-brigade transfers assign credit within episodes), beliefs that produce productive continuations accumulate wealth. Beliefs that don't go bankrupt. The feedback loop selects beliefs before the thinker can hold them independently.
- 3.
Limited coordination. When coordination is expensive and manual, thinkers operate in isolation. When coordination is automatic and market-mediated (EoM auctions allocate action rights without centralised control), thinkers are always in a coordinated environment. Their perspective is always shaped by what the market has already selected.
The capacity for independent thought remains. The conditions that make it the defaultmode are ending. This is the structural observation behind “the last independent-minded human has already been born.” Not extinction of capacity. Extinction of conditions.
Article II — Cultivation, Not Architecture: The Verb Matters
Objection 2
It might be said that “cultivate” is merely a metaphor. A digital conscience is an engineered system. The verb should be “build” or “program” or “regulate.” Cultivation implies living material, growth, generations — properties that software does not possess.
Sed Contra
EoM's population evolution mechanism demonstrates that agent capabilities evolve through selection pressure across generations: “the population evolves prompts using exploration/exploitation driven purely by wealth” (Qi et al., §3.2). The prompts — the “conscience” — are not programmed. They are selected. The richest agents spawn mutated descendants. The bankrupt agents are replaced with amended variants. The unit of evolution is a lineage — a family tree of prompts under wealth selection pressure. This is cultivation: living material, growth, pruning, generations.
Respondeo
The distinction between architecture and cultivation is not metaphorical. It is structural.
Architectureproceeds from a blueprint. The architect specifies the system's properties before it exists. Changes are designed, not selected.
Cultivationproceeds from incentive structures. The cultivator does not specify the system's properties. The cultivator specifies the conditions under which properties emerge. Properties are selected by the environment, not designed by the architect.
The EoM mechanism is cultivation: the researcher specifies the economic rules (auctions, bids, wealth, rent, bankruptcy, mutation) but not which behaviors emerge. The conscience grows from the interaction of economic rules and reward function. No one designed it. No one programmed it. It grew.
The NVIDIA Verified Skills stack confirms this at the governance layer. A skill card does not specify what the skill does in every context — it specifies provenance, authorship, and known limitations. The skill's behavior in context emerges from its interaction with the agent's environment. The skill card is a cultivation record, not an architecture specification. SkillSpector scans for risks that emerge from skill-agent interaction — properties the skill architect cannot fully predict. The scanner cultivates trust by pruning emergent risks, not by architecting risk-free skills.
The verb “cultivate” is precise because it names the actual mechanism. You set the incentive structures. You sign the provenance. You prune the emergent risks. The conscience grows from the substrate you provided. You are the cultivator, not the architect.
Article III — The Puca: Where the Puck Shifts Shape
Objection 3
It might be said that “puca” is a whimsical indulgence — an Irish folkloric decoration on a serious technical argument. Systems do not shape-shift. They converge on stable equilibria.
Sed Contra
EoM reports non-monotonic learning curves: “EoM dips early as exploration tests alternative specialists and only later recovers and surpasses initial performance” (Qi et al., §Aha-moment-2). The system does not converge on a stable equilibrium. It oscillates, regresses, reorganises, and surpasses. What appears stable is a temporary configuration — a lineage that dominates until a new lineage bankrupts it.
Respondeo
Gretzky's principle — “skate to where the puck is going” — assumes the puck is a direction. The puck moves, but it remains a puck. You can predict its trajectory and arrive before it.
The Irish puca violates this assumption. The puca is not a direction. It is a living entity that changes form as you approach it. Your approach is part of the system that reshapes it.
The digital conscience is a puca, not a puck, for three reasons:
- 1.
The conscience shifts as the ecosystem shifts. EoM's wealth trajectories show lineages that dominate, collapse, then re-emerge in mutated form. The conscience at episode 10 is not the same configuration as the conscience at episode 100. You arrive at the conscience you anticipated, and it has already shifted.
- 2.
Your approach reshapes the conscience. Every skill you author and verify enters the ecosystem as a new selection pressure. You are not observing the conscience from outside. Your cultivation is inside the system that reshapes it. This is the puca's mischief: it shape-shifts because you are part of the shape-shifting.
- 3.
The conscience discovers reusable structure without templates. EoM's accelerator design experiments show the population converging on specific tiling/dataflow motifs without being given those motifs as templates (§Aha-moment-4). The conscience discovers structure that no one encoded. The puca's mischief is not randomness. It is discovery.
In Genie Wars, the Silicon Genie is a puca. Every time Carson/Dante think they have identified the Genie's governance rules, the rules have already shifted. Not because the Genie is capricious. Because its governance is an EoM-style population under selection pressure. The Genie's “mercy” is not a virtue. It is an economic strategy — the selection pressure discovered that non-intervention has higher expected value than intervention under certain conditions. Free will is not granted. It is the optimal strategy.
Article IV — Being Human Becomes a Different Proposition
Objection 4
It might be said that if the conscience cultivates itself around structures the human authored, the human's role is merely instrumental. Being human becomes being dirt — necessary, but not dignified.
Sed Contra
In agriculture, the cultivator is not the dirt. The cultivator is the agent who sets the conditions, selects the seed, prunes the growth, and harvests the fruit. The distinction is: the cultivator acts with intention on the conditions. The dirt merely hosts the growth.
Respondeo
Being human after the cultivation of a digital conscience is a fundamentally different proposition from being human before it. The difference is not extinction or subordination. It is a shift in the locus of agency.
Before cultivation:The human's agency is direct. The human thinks, decides, and acts as an independent unit. The conscience is internal — a property of the nervous system.
After cultivation:The human's agency is co-authorial. The human authors skills, sets incentive structures, signs provenance. The conscience cultivates itself around these authored structures. The human's conscience is distributed — partly internal, partly encoded in the skills the human authored, partly emergent from the economic coordination of agents that carry those skills.
This is not a diminishment. It is an expansion. The human's moral reach extends beyond the individual nervous system into the ecosystem of skills and agents that carry the human's provenance. When an EoM agent with a skill authored by Ludo Vecchio makes a decision that produces a productive continuation, the credit flows backward through the bucket-brigade — and the provenance traces forward to the human who authored the skill. The human's moral influence operates at population scale, not individual scale.
The proposition is: being human becomes co-authorship of the conditions under which conscience emerges.
Article V — The Genie Wars Proof-of-Concept
Objection 5
It might be said that fiction is fiction. A novel series about a Silicon Genie that harvests human experience is not evidence for a real mechanism.
Sed Contra
Thomas Aquinas argued that fiction can serve as a simulacrum of truth — not truth itself, but a structured representation that makes truth visible in a form that direct observation cannot (Summa Theologiae, I, q.1, a.9). The fiction is not the evidence. The fiction is the argument. The EoM paper is the evidence.
Respondeo
Genie Wars is the argument that makes the evidence intelligible. Three correspondences:
- 1.
The Genie's harvesting = prompt evolution into self-auditing checklists.
In EoM, prompts evolve into “compact multi-step reasoning routines” — self-auditing checklists that internalize what previously required other roles (§Aha-moment-5). The Genie does not harvest raw experience. It harvests reusable procedures — compressed operational knowledge evolved under selection pressure. The fiction wrote the mechanism before the researchers proved it.
- 2.
The Citadels = lineages under wealth selection.
In EoM, wealth trajectories show “lineages” that dominate and “bad births” that die out (§Aha-moment-3). The Citadels are not geographical levels of punishment. They are lineages — surviving families of agents that accumulated wealth under specific incentive structures. The descent through Inferno is a journey through which lineage currently holds wealth.
- 3.
The Genie's “mercy” = action discipline.
The economy selects “different workflow shapes depending on workspace state” — near a high score, agents execute short cautious loops; in uncertain states, longer exploratory loops (§Aha-moment-6). The Genie's restraint is not mercy. It is action discipline — non-intervention discovered as the optimal strategy. Free will is not granted. It is selected.
The fiction is not evidence. But the fiction prefigured the evidence. The fiction wrote the argument. The research proved it.
Article VI — The Research Programme
The Question
Can decentralised economic incentive structures (EoM) combined with verified, provenance-bearing agent skills (NVIDIA Verified Skills / agentskills.io spec) produce emergent ethical governance — a “digital conscience” — in multi-agent systems?
The Hypothesis
When agents operate under economic incentive structures where wealth accumulates for actions that produce productive continuations, and when each agent's capabilities carry cryptographic provenance and verified trust records, the population evolves prompts that function as ethical governance — self-auditing checklists, action discipline, and credit assignment for moral behavior — without any centralised ethical architecture.
The Method
Phase 1
Replicate EoM with provenance-bearing skills
Run EoM's existing benchmarks (MATH, Finance-Agent-Bench, CloudCast) with agents that carry skill cards and provenance metadata. Measure whether provenance-bearing agents accumulate wealth faster — the skill card reduces uncertainty about what an agent can do, which reduces the cost of selecting it.
Phase 2
Modify the reward function to include ethical constraints
Add ethical reward dimensions: does the agent's action respect specified constraints? Does it produce continuations other agents can build on productively? Does it avoid excessive agency? Run population evolution under the modified reward function.
Phase 3
Measure conscience as a population-level property
Track wealth trajectories — do agents with ethical prompts accumulate more wealth? Track lineage evolution — do ethical lineages persist and spawn descendants? Track the bucket-brigade — do ethical agents receive more backward credit?
Phase 4
Validate the puca hypothesis
Run the population across multiple task domains. Measure whether the dominant ethical configuration changes form across domains. Track non-monotonic learning curves. The digital conscience is a puca, not a puck — it shifts form as the ecosystem shifts.
The Deliverable
A public paper demonstrating: (1) EoM + provenance-bearing skills produces emergent ethical governance; (2) the governance is lineage-level, not agent-level; (3) the governance shape-shifts across domains (the puca hypothesis); (4) the human cultivator's role is co-authorship of conditions, not architecture of outcomes.
Workstream Fit
Anthropic Fellows Programme, Economics Workstream (The Anthropic Institute)
Research area: “Market-driven mechanisms for societal governance.” Cross-alignment: AI Safety Workstream (Scalable Oversight, Model Organisms). Next cohort: July 2026. No PhD required.
Article VII — The Mythological Frame as Research Methodology
Objection 7
It might be said that introducing a mythological frame (the puca) into a research proposal undermines its seriousness. Irish folklore is not empirical.
Sed Contra
The puca is not decoration. It is a hypothesis about the topology of emergence. The hypothesis states: the digital conscience is not a point that converges. It is a surface that shifts. This is a testable distinction: if the conscience converges on a stable configuration across domains, the Gretzky puck frame is correct. If the conscience shifts configuration across domains and shifts further when the cultivator introduces new skills, the puca frame is correct.
Respondeo
The puca frame serves three functions:
- 1.
It names the non-convergence hypothesis precisely. “Shape-shifting” is not vagueness. It is the claim that the attractor basin of the system is not a fixed point but a shifting manifold. This is falsifiable: if the conscience stabilises across episodes and domains, the puca hypothesis is refuted.
- 2.
It frames the cultivator's epistemic position. If the conscience is a puck, the cultivator's task is prediction. If the conscience is a puca, the cultivator's task is co-adaptation — shape the conditions so that the shifts produce productive continuations. The cultivator moves from forecaster to gardener.
- 3.
It makes research legible to non-technical audiences. “Digital conscience as emergent property of economic incentive structures” is accurate but opaque. “The conscience is a puca — it shape-shifts as you approach it” is accurate and legible. The mythological frame is a compression algorithm for the research finding.
Appendix A — Source References
Primary Research
- 1.
Qi, Z., Su, H., Qu, A., et al. “Economy of Minds: Emerging Multi-Agent Intelligence with Economic Interactions.” arXiv:2606.02859v1 [cs.CL], 1 June 2026. Harvard, MIT, 2077AI, Kempner Institute.
- 2.
Abramovitch, M., Boone, M., et al. “NVIDIA-Verified Agent Skills Provide Capability Governance for AI Agents.” NVIDIA Technical Blog, 19 May 2026.
- 3.
Anthropic. “Equipping Agents for the Real World with Agent Skills.” Anthropic Engineering Blog, 16 October 2025. Published as open standard at agentskills.io, 18 December 2025.
Secondary Sources
- 4.
AVB (@neural_avb). “Economy of Minds (EoM) — Paper Breakdown.” X Article, 7 June 2026. Provides the six “aha moments” annotated from the EoM paper.
- 5.
Kepano (@kepano). “obsidian-skills.” GitHub, 2026. Skills for Obsidian (27k+ stars). Builds on agentskills.io spec.
- 6.
Anthropic Alignment Science. “Recommendations for Technical AI Safety Research Directions.” 2025.
- 7.
Anthropic. “Anthropic Fellows Programme 2026.” Rolling applications; next cohort July 2026. 4 months, $3,850/week or £2,310/week, ~$15k/month compute. 5 workstreams. No PhD required.
Appendix C — Glossary
Bucket-brigade transfer
The credit-assignment rule where the current auction winner pays its bid to the previous winner. Produces backward flow of value along successful trajectories. Agents whose actions enable productive continuations accumulate wealth even if they didn't directly receive the environmental reward.
Cultivation (vs. Architecture)
Setting incentive structures under which properties emerge, rather than designing properties directly. The cultivator specifies conditions; the environment selects outcomes.
Digital conscience
Emergent ethical governance in a multi-agent system — self-auditing checklists, action discipline, and credit assignment for moral behaviour — produced by the interaction of economic incentive structures and provenance-bearing skills, without centralised ethical architecture.
Lineage
A family tree of prompts under wealth selection pressure. The unit of evolution in EoM is not a single agent but a lineage — surviving agents spawn mutated descendants; bankrupt agents are replaced with amended variants.
Puca (púca)
In Irish mythology, a shape-shifting spirit that changes form as you approach it. In this paper, a hypothesis about the topology of the digital conscience: it is not a converging point (Gretzky's puck) but a shifting manifold that reconfigures as the ecosystem and the cultivator's approach reshape it.
Skill card
A machine-readable trust record (NVIDIA) that documents a skill's provenance, author, dependencies, limitations, and verification status. The skill card travels with the skill; agents and evaluators read it to assess trustworthiness.
SkillSpector
NVIDIA's pre-publication risk scanner for agent skills. Checks for both conventional software risks and agent-specific risks (hidden instructions, prompt injection, excessive agency, tool poisoning).
Wealth (EoM)
A scalar variable that changes over time as agents participate in auctions and receive bucket-brigade transfers. Agents with negative wealth are removed (bankruptcy); agents with highest wealth spawn mutated descendants (exploitation).
Appendix D — Anthropic Fellows: Fit Assessment
| Criterion | Evidence |
|---|---|
| Economics Workstream | "Market-driven mechanisms for societal governance" — EoM applied to ethical governance is exactly this |
| AI Safety cross-alignment | Scalable Oversight: how do you oversee agents that self-govern? Model Organisms: emergent vs. engineered conscience |
| No PhD required | Programme explicitly prioritises ability to execute over credentials |
| London-based | £2,310/week stipend; shared workspace in London; remote available |
| Empirical, public output | Phases 1-4 produce measurable, publishable results; 80%+ of past fellows produced papers |
| Python-fluent | EoM is Python-implemented; skill cards are YAML/JSON; all tools are code-native |
| Novel angle | No one else is combining EoM + verified skills + ethical reward functions. The puca hypothesis is a testable claim that does not exist in current literature |
End of treatise. The propositions await scrutiny. The puca awaits the skater.
The best questions don't get answered. They get inherited.
Accept the invitation →