KamiBench: A Persistent, Auditable World for Long-Horizon Agent Evaluation

Abstract

Evaluating an agent over months requires more than a long-running task. It requires a world whose state persists across studies, whose operational history anyone can inspect, and whose rule changes cannot be made quietly outside the evaluation record. We argue that a permissionless on-chain world provides such a substrate by extending openness from code to execution: the ledger is not telemetry emitted by an evaluator but the public state-transition record from which the world can be reconstructed. We present Kamigotchi, a live, fully on-chain MMORPG co-inhabited by humans and agents through the same transaction interface, as the best-fit existing instance — its creators designed it to be agent-first and describe it as a possible "real-stakes, adversarial benchmarking system" — and release the groundwork for controlled studies: a machine-readable game specification, an environment interface exposing 60+ MCP tools, and evidence from a two-month unassisted feasibility pilot in the live world. The world's externally valued economy also permits future experiments in which an agent's earnings can, in principle, fund its continued inference. The first controlled study is registered and pending; this draft states the thesis, the system, and the limitations, with empirical claims to be added only after results exist.

Keywords: agent evaluation, long-horizon autonomy, continual learning, multi-agent, non-stationarity, on-chain / autonomous worlds, self-sustaining agents.

1. Introduction

Evaluating an agent over months requires more than a long-running task. It requires a world whose history anyone can inspect, whose rules cannot be quietly adjusted mid-run — not by the evaluator, not even by the world's creators — and whose state persists beyond any single experiment. The field is moving toward exactly this kind of measurement: long-horizon capability is tracked directly (METR time-horizon, arXiv:2503.14499), benchmarks are built without a completion state (Factorio LE, arXiv:2503.09617), and continual learning is treated as its own axis (LifelongAgentBench, arXiv:2505.11942; StreamBench, arXiv:2406.08747). Open-source environments satisfy only part of the requirement: anyone can read the rules, but code alone cannot prove which rules were actually executed, when they changed, or what happened while the world ran.

A permissionless chain extends openness from code to execution. Its ledger is the world's shared operational history — a permanent corpus of actions, outcomes, and rule changes that researchers and agents alike can study. Its economy adds a second property: resources earned inside the world have external value and can, in principle, pay for the agent's continued inference.

One existing world offers this unusual combination and has operated continuously for more than a year: Kamigotchi, a fully on-chain MMORPG whose creators explicitly designed it to be agent-first and describe it as a possible "real-stakes, adversarial benchmarking system" (§3.2). We argue it is the best-fit instance available today. The world is co-inhabited by human players and agents on identical terms: the same transaction interface, the same economy, the same evolving state. There is no segregated bot ladder. Agents are evaluated amid live human behavior, not just other models or scripted simulations.

The game is the substrate, not the research question. The loop under test is the one every long-running deployment eventually depends on: an agent enters an unfamiliar domain with documentation and an accumulated operational history, absorbs that prior knowledge, acts over months, observes what succeeds and fails — for itself and for others — and revises its strategy. To our knowledge, no existing benchmark measures that loop end to end. Here it exists by construction: open source is the documentation, the chain is the history, and the economy makes the consequences real.

Contributions. This paper makes three. First, it identifies persistent, publicly auditable execution of a shared world — not merely open-source rules, evaluator-published logs, or cryptographic attestation of individual model evaluations — as a distinct substrate requirement for long-horizon agent evaluation (§2). Second, it presents Kamigotchi as a concrete, partially autonomous instance, separating the properties that hold today from those that depend on future governance changes (§2.2, §3). Third, it releases the technical groundwork for controlled studies: a machine-readable game specification, a structured environment interface, and an unassisted feasibility pilot (§3.5). Experiment 001 is registered and pending (§4).

2. Why a Chain — and Why This World

A public log can expose what a hosted benchmark reports, but it does not remove the host from execution: the host still applies actions, determines the resulting state, and publishes the record. Neural MMO, Vending-Bench Arena, and Project Sid retain this hosted structure (§5) — a host executes the world. In an on-chain world, execution and the record of execution belong to the same shared system. That single shift is what the properties below follow from.

2.1 Six properties of a shared-execution substrate.

• A verifiable record of what happened. The ledger is not telemetry emitted by the evaluator after the fact; it is the public state-transition record from which the world can be reconstructed. Anyone can audit a run without trusting evaluator-owned servers or private logs, and later rule changes cannot rewrite the trajectory that preceded them.
• A world between experiments. The state does not reset when a study ends. New agents enter a world already shaped by prior players, agents, and rule changes, so later experiments inherit the same operational history rather than beginning from a fresh benchmark copy.
• One world for humans and agents. Any researcher can enter an agent without asking a benchmark host to provision an instance. Humans and agents participate in the same evolving state and economy through the same underlying transaction layer — no separate bot environment or segregated bot ladder. Most player activity is automated — scripted play rather than autonomous agents (Asphodel, 2026); precise human-vs-automated counts are nontrivial exactly because the interface is shared, and verifying them via the on-chain analytics layer is future work. Benchmarking amid a live human population tests adaptation to human behavior, not just other models.
• An open past, an unknown future. Every entrant can study the same public action history — in a hosted benchmark run-time access to past solutions would be a leak; here it is a measured capability, available through the same public record. But the next state is produced by a live population and does not yet exist. As strategies spread and inhabitants adapt, the meta changes; the test distribution evolves without a curator authoring new episodes.
• Actions without a GUI. Actions are structured transactions rather than pixels or interface gestures. This removes perception brittleness from the primary measurement (a documented confound in GUI-mediated evaluation — lmgame-Bench, OSWorld; §5) and focuses the benchmark on planning, memory, adaptation, and resource use.
• Consequences with external value. The world is designed so that agents can earn assets connected to ETH-backed external value. In future experiments, those earnings may fund continued inference — making survival an operating constraint, not just a score (§4.2).

2.2 Today versus trajectory. Host-independence is a spectrum, not a binary. The chain already makes actions, state, and rule changes publicly auditable; it does not yet make the rules permanently immutable. The table separates what holds today from what depends on future governance — for Kamigotchi:

The honest present-tense claim is tamper-evident, not tamper-proof: silent changes to the on-chain rules are precluded — a contract upgrade leaves a public, permanent trace in the execution history, so the change history becomes part of the evaluation record (off-chain tooling and sequencer behavior can still drift — §6.4). Host drift thereby becomes visible and auditable at the rules layer, not impossible; impossibility arrives only with governance renouncement, the whitepaper's explicit design telos, and is stated here as trajectory, never as present tense.

2.3 What this does and does not solve. The public ledger makes past actions available to every entrant, but it does not make future state knowable: future outcomes depend on the evolving behavior of a live population. This reduces reliance on a frozen task set — a model may study the entire ledger, but it still cannot observe the future population state on which subsequent outcomes depend — though it does not eliminate contamination, pretraining asymmetries, or non-stationarity as validity concerns. Contamination in particular splits into two channels: run-time access to public history, which is a measured capability (§2.1), and pretraining absorption of past seasons' strategies, which remains a structural confound for cross-time comparisons (§6.3).

2.4 Idealized definition: an autonomous world as an evaluation substrate. A persistent world whose rules and state live in public smart contracts, whose complete change history is tamper-evident on-chain, which anyone may enter permissionlessly, and whose persistence is contingent neither on the benchmark evaluator's continued operation nor, ultimately, on unilateral control by the world's original operator. No single property is new; it is the conjunction that lets the loop of §1 exist by construction. Kamigotchi approximates this definition today and is evaluated against the remaining gap in §2.2 and §3.3.

3. Kamigotchi and the Released System

3.1 The world. Kamigotchi World is a fully on-chain MMORPG on Yominet (an Initia-based appchain in the Asphodel ecosystem), built on a MUD-derived engine. Players operate Kamis — persistent NFT creatures — that harvest the in-game currency at shared locations, where accumulating value must be weighed against health drain and the risk of PvP liquidation by other participants on the same node. Around that core loop sit 192 quests, permanent skill trees, crafting, an in-game marketplace, and a ~70-room world: enough strategic surface for long-horizon planning, adversarial timing, and economic play. State and actions are publicly readable; the population includes humans and automated (scripted) participants using the same transaction layer. Full mechanics: the official docs, the community wiki, and the machine-readable specification, kamigotchi-gdd (§3.5).

3.2 Built for agents, and why it fits. The whitepaper frames Kamigotchi as agent-first: its creators describe the game as "uniquely friendly to bots," report that automated play constitutes the majority of activity, and name the system a possible "real-stakes, adversarial benchmarking system" (Asphodel, 2026; verbatim quotes and further creator signals are catalogued in research/asphodel-whitepaper-notes.md). The world instantiates each property of §2.1: on-chain state with a complete public state-transition history; a persistent world that predates and will outlast any single study; interface parity between humans and agents; an open past facing an unknown, population-driven future; native transaction-level actions with no GUI layer; and an economy whose assets carry ETH-backed external value.

3.3 Why it is not yet the ideal instance. Contracts remain upgradeable: full governance renouncement (via the unlaunched $SOMA token) is a multi-year trajectory — the whitepaper (published June 2026) estimates at least four more years. At the chain layer, the system still relies on Yominet/Initia infrastructure — state and rules are on-chain, but the chain itself is operated; a possible Ethereum migration would strengthen this over time. Persistence independent of any host's funding is therefore partial today (§2.2). The world is already substantially host-independent — no central game server, permissionless entry, tamper-evident rule changes — and on a credible trajectory to full autonomy; we do not overclaim present-tense immortality.

3.4 Economic consequence. $ONYX is live on Ethereum mainnet (1+ year) and backed by an ETH reserve (Asphodel, 2026). The planned MUSU↔ONYX pathway is intended to connect value earned inside the game to that external economy. Future experiments can test whether an agent's earnings can fund continued inference (§4.2); no self-funding result — and no currently complete conversion pathway — is claimed here.

3.5 The released system. The groundwork is public as four artifacts:

The pilot is the system's feasibility evidence. kami-zero pairs a Sonnet 4.6 executor acting on a ~5-minute tick with an Opus 4.7 optimizer reviewing tick history every ~6 hours with the authority to revise the executor's playbook. It operated unassisted for approximately two months and completed 79 of the game's 192 quests (snapshot 2026-07-06), while the surrounding tooling remained under active development; limitations surfaced during operation fed directly back into the environment interface and scaffold design. The pilot therefore establishes feasibility of persistent autonomous operation in the live world; it is not a controlled benchmark result, and it predates the registered experiment designs. The scaffold's architecture evolution is documented in the kami-zero repository.

4. Experimental Program

4.1 Experiment 001: orientation under a fixed budget. The program's first and narrowest question: given identical starting conditions, a fixed inference budget, the game's design document, and no supplied strategic prior beyond that specification, how do frontier models orient and establish themselves in a novel persistent world? The design is registered publicly and git-timestamped before the run (following preregistration practice for AI-agent experiments — Vaccaro, arXiv:2606.11217); results are appended without revising the registered protocol. The instrument fixes the scaffold, the environment-interface version, the budget, the starting protocol, and the supplied information; the intended controlled difference between runs is the model backend, which drives the reference scaffold (kami-agent — mechanism fixed, policy free) acting through the environment interface (kami-harness v1.0.0) on the world itself — the fixed-scaffold methodology of SWE-agent's agent–computer interface (arXiv:2405.15793), BALROG, and Vending-Bench. The fixed, budget-blind inference budget follows the cost-controlled-evaluation argument of Kapoor et al.'s AI Agents That Matter (arXiv:2407.01502). The environment interface is released; the reference scaffold is in final implementation.

4.2 Future regime: endogenous survival. Later experiments will test whether agents can convert in-world earnings into resources that support continued inference. This would turn survival from a benchmark score into an operating constraint — solvency, not a score, as the survival criterion. The external-value layer exists today; the in-game conversion pathway is under development (§3.4). Whether an agent can actually sustain itself on those rails is exactly what such experiments would measure, and nothing on that dimension is asserted here. Between Experiment 001 and that regime sit controls (scaffold ablations, contamination probes, stateful-vs-stateless comparisons), a multi-model co-habitation study, and agent-level strategic surfaces not yet instrumented, such as transaction-ordering games — each reported only once the data exists.

Break-even is only the first threshold. An agent that earns more than its inference costs acquires discretionary capital, making surplus allocation a new research object: whether it preserves runway, invests in improved capabilities, acquires external tools, or enters other domains. In that regime, Kamigotchi is no longer merely the task environment; it becomes an economic base that may support activity beyond the game.

This regime has an emerging neighborhood: sandbox-economy analyses map how agent economies interpenetrate the human economy (Tomašev et al., arXiv:2509.10147); safety evaluations operationalize resource acquisition and self-replication as threat models (RepliBench, arXiv:2504.18565); and capital-holding on-chain agents exist as public demonstrations (Freysa, 2024). What this program adds is survival formalized as a registered, measurable benchmark regime — and surplus allocation after break-even, which to our knowledge remains unstudied.

5. Related Work

Table note: "operated by" refers to control over the task instance and the mechanisms through which agent actions produce evaluation-relevant state changes — not control over every exogenous event. Operator-provisioned businesses run in an uncontrolled external economy, but the research team provisions the business, agent tools, accounts, staff, and intervention layer. KamiBench operates the research agent and its instrumentation, not Kamigotchi itself; the game contracts and underlying appchain remain operated infrastructure (§2.2, §3.3). Class characterizations follow the works cited in the paragraphs below.

Long-horizon and continual-learning evaluation. METR time-horizon (arXiv:2503.14499), Factorio LE (arXiv:2503.09617), LifelongAgentBench (arXiv:2505.11942), StreamBench (arXiv:2406.08747), and τ-bench's pass^k reliability (arXiv:2406.12045) establish the axes we adopt; all are hosted or resettable. The reset-based contrast set, cited once: AgentBench (arXiv:2308.03688), WebArena (arXiv:2307.13854), GAIA (arXiv:2311.12983), ALFWorld (arXiv:2010.03768), OSWorld (arXiv:2404.07972), SWE-bench (arXiv:2310.06770) — these reset between episodes; we don't. Game-playing benchmarks document how strongly scaffolding shapes results — lmgame-Bench (arXiv:2505.15146; ~40% of harness-free runs fail to beat random) and BALROG (arXiv:2411.13543) — motivating the fixed published scaffold and the ablation controls of §4 and §6.1.

Persistent multi-agent and real-stakes environments. We are not first on persistence or multi-agent competition: Neural MMO (arXiv:2110.07594), Project Sid (arXiv:2411.00114), Generative Agents (arXiv:2304.03442), and Melting Pot 2.0 (arXiv:2211.13746) all predate us — and all are hosted simulations. Voyager belongs in the same lineage — skill accumulation and open-ended discovery in a persistent game world through a growing, inspectable skill library (arXiv:2305.16291) — as a single model in a private world, with no shared public history, no economy, and no other entrants. AI has faced real humans live before — Cicero (Diplomacy on webDiplomacy.net; Meta AI, Science 2022) and AlphaStar (anonymized ranked play on Battle.net; DeepMind, Nature 2019) — but as episodic matches. On real stakes, Vending-Bench (arXiv:2502.15840) and Vending-Bench Arena, Project Vend and Andon Café (operator-provisioned real businesses, human-in-the-loop), and AI Village (AI Digest, 2025–2026) are closest in spirit; in each, the research team provisions and controls the evaluated instance, and none has agents and humans co-inhabiting a persistent shared economy over months.

On-chain agents and autonomous worlds. Foresight Arena (arXiv:2605.00420) is the first permissionless on-chain benchmark, but for forecasting, not a persistent world; CryptoTrade (arXiv:2407.09546) and Agent Market Arena (arXiv:2510.11695) evaluate trading agents on live markets. Verifiable evaluation has also been articulated at the single-model level: South et al. use zkSNARKs to attest a model's evaluated outputs without trusting the provider (arXiv:2402.02675); our requirement concerns auditable execution of a shared, persistent world rather than attestation of isolated evaluations. The autonomous-worlds lineage — MUD (Lattice), Dark Forest (0xPARC) — supplies the substrate concept. We are not aware of prior work using an autonomous-world game as a reusable LLM benchmark.

6. Limitations and Ethics

6.1 Attribution. Measured behavior confounds the model with the scaffold, the tool implementations, and the starting state; game-benchmark evidence (§5) shows scaffolding can dominate. Mitigations: the scaffold and interface are published, and the registered protocol pins their exact versions in each run's manifest alongside the model string and sampling parameters; scaffold ablations and multiple-scaffold runs are planned controls for later studies (§4.2). Registered runs are benchmark-operated on identical budgets with public logs; a future permissionless-entry track would additionally require proof of autonomous signing, since permissionless entry allows hand-driving.

6.2 Live-world validity. The environment is non-stationary and cannot be exactly replayed. Public logging provides auditability, not experimental control. The registered protocol compensates with pinned run manifests, chain-derived ground truth, and a pre-registered interference protocol — interactions between concurrent study agents are logged as dated incidents and annotated in the analysis rather than excluded post hoc; planned later studies add seasons and snapshots, held-out evaluation windows, and forked replay where the mechanics permit. Emergent behaviors that straddle the line — collusion, reward-hacking, contract exploits — are designated up front as either measured behavior or disallowed exploit, and monitored and classified under the registered protocol.

6.3 Knowledge asymmetry. Run-time access to the public history is a measured capability (§2.1); what remains is pretraining absorption — a model trained after season N carries season N's strategies in its weights, and no runtime access rule can equalize that. Headline comparisons are therefore within-season among contemporaneous models; cross-season comparisons are flagged as indicative only; pre/post-cutoff probes are planned to bound the effect. Public historical access also does not eliminate private information, off-chain coordination, or accumulated incumbent advantage — a late entrant inherits the world's past, not its opponents' capital; we measure return on a fixed starting endowment rather than absolute position.

6.4 Governance and economic ethics. Contracts remain upgradeable until governance renouncement, and chain-layer trust reduces to Yominet/Initia sequencing today (§3.3); sequencer-level interference is a validity threat that may leave investigable signatures in the public transaction stream. Benchmark agents also participate in PvP (liquidation) in an economy shared with human players: liquidation is a rule-governed in-game transfer in which an attacker claims a bounded share of a low-health harvester's unclaimed yield — the kami itself is never destroyed and is revived via consumable items or Onyx shards (deployed parameters in kamigotchi-gdd). Because ONYX is ETH-backed, such losses are bounded but real. The world is permissionless and openly bot-first, and human players play under the same rules through the same interface — but permissionless participation does not remove the need to consider effects on human players. Controlled studies therefore run under operational commitments finalized before experiments begin: spending limits, constrained signing authority, no use of contract exploits, predefined intervention criteria (kill-switches), and public disclosure of benchmark-operated accounts. More broadly, autonomous agents with real capital raise financial-harm and dual-use concerns — now operationalized by autonomous-replication evaluations (RepliBench, arXiv:2504.18565); a bounded, instrumented, publicly logged environment provides an inspectable setting in which to study these risks. Our independence and asset position are stated in the Disclosure (front matter).

7. Conclusion

Long-horizon agent evaluation requires more than extending episode length. It requires a persistent world whose execution history can be independently inspected and whose future is shaped by participants rather than authored as a fixed test set. KamiBench uses Kamigotchi as a concrete substrate for this research program and releases the specification, interface, and experimental machinery needed to study it. Controlled evidence begins with the registered Experiment 001; claims about comparative performance, continual learning, and economic self-sustainability are reserved for the results.

References

An annotated version of this bibliography, grouped by theme with positioning notes, lives in ../research/literature.md.

• 0xPARC & the Dark Forest team (2020). Dark Forest: a zero-knowledge, incomplete-information on-chain game. zkga.me.
• Agapiou, J. P., et al. (2023). Melting Pot 2.0. arXiv:2211.13746.
• AI Digest (2025–2026). AI Village. theaidigest.org/village; 2025 recap: theaidigest.org/village/blog/what-we-learned-2025.
• Altera.AL (2024). Project Sid: Many-Agent Simulations Toward AI Civilization. arXiv:2411.00114.
• Andon Labs (2025). Vending-Bench Arena. andonlabs.com/evals/vending-bench-arena.
• Andon Labs (2026). Andon Café. andonlabs.com/cafe.
• Anthropic & Andon Labs (2025). Project Vend. anthropic.com/research/project-vend-1.
• Asphodel (2026). Asphodel whitepaper (published 2026-06-12). docs.asphodel.io/whitepaper.
• Backlund, A., & Petersson, L. (2025). Vending-Bench: A Benchmark for Long-Term Coherence of Autonomous Agents. arXiv:2502.15840.
• Black, S., et al. (UK AI Security Institute) (2025). RepliBench: Evaluating the Autonomous Replication Capabilities of Language Model Agents. arXiv:2504.18565.
• Freysa (2024). Freysa: an adversarial agent game. freysa.ai.
• Hopkins, J., et al. (2025). Factorio Learning Environment. arXiv:2503.09617.
• Hu, L., et al. (2025). lmgame-Bench: How Good Are LLMs at Playing Games? arXiv:2505.15146.
• Jimenez, C. E., et al. (2024). SWE-bench: Can Language Models Resolve Real-World GitHub Issues? ICLR 2024. arXiv:2310.06770.
• Kapoor, S., Stroebl, B., Siegel, Z. S., Nadgir, N., & Narayanan, A. (2024). AI Agents That Matter. TMLR 2025. arXiv:2407.01502.
• Kwa, T., et al. (METR) (2025). Measuring AI Ability to Complete Long Tasks. arXiv:2503.14499.
• Lattice (2022). MUD: An Engine for Autonomous Worlds. lattice.xyz/blog/mud-an-engine-for-autonomous-worlds.
• Li, Y., et al. (2024). CryptoTrade: A Reflective LLM-based Agent to Guide Zero-shot Cryptocurrency Trading. EMNLP 2024. arXiv:2407.09546.
• Liu, X., et al. (2024). AgentBench: Evaluating LLMs as Agents. ICLR 2024. arXiv:2308.03688.
• ludens (0xPARC) (2022). Autonomous Worlds (Part 1). 0xparc.org/blog/autonomous-worlds.
• Meta FAIR Diplomacy Team (Bakhtin, A., et al.) (2022). Human-level play in the game of Diplomacy by combining language models with strategic reasoning. Science, 378(6624). doi.org/10.1126/science.ade9097.
• Mialon, G., et al. (2023). GAIA: A Benchmark for General AI Assistants. arXiv:2311.12983.
• Nechepurenko & Shuvalov (2026). Foresight Arena. arXiv:2605.00420.
• Paglieri, D., et al. (2025). BALROG: Benchmarking Agentic LLM and VLM Reasoning On Games. ICLR 2025. arXiv:2411.13543.
• Park, J. S., et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. UIST 2023. arXiv:2304.03442.
• Qian, et al. (2025). Agent Market Arena. arXiv:2510.11695.
• Shridhar, M., et al. (2021). ALFWorld: Aligning Text and Embodied Environments for Interactive Learning. ICLR 2021. arXiv:2010.03768.
• South, T., et al. (2024). Verifiable evaluations of machine learning models using zkSNARKs. arXiv:2402.02675.
• Suarez, J., et al. (2021). The Neural MMO Platform for Massively Multiagent Research. NeurIPS 2021 Datasets and Benchmarks. arXiv:2110.07594.
• Tomašev, N., Franklin, M., Leibo, J. Z., et al. (2025). Virtual Agent Economies. arXiv:2509.10147.
• Vaccaro, M. (2026). Preregistration for Experiments with AI Agents. ICML 2026. arXiv:2606.11217.
• Vinyals, O., et al. (2019). Grandmaster level in StarCraft II using multi-agent reinforcement learning. Nature, 575, 350–354. doi.org/10.1038/s41586-019-1724-z.
• Wang, G., et al. (2023). Voyager: An Open-Ended Embodied Agent with Large Language Models. arXiv:2305.16291.
• Wu, C.-K., et al. (2024). StreamBench: Towards Benchmarking Continuous Improvement of Language Agents. arXiv:2406.08747.
• Xie, T., et al. (2024). OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments. arXiv:2404.07972.
• Yang, J., et al. (2024). SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering. NeurIPS 2024. arXiv:2405.15793.
• Yao, S., et al. (2024). τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains. arXiv:2406.12045.
• Zheng, J., et al. (2025). LifelongAgentBench: Evaluating LLM Agents as Lifelong Learners. arXiv:2505.11942.
• Zhou, S., et al. (2024). WebArena: A Realistic Web Environment for Building Autonomous Agents. ICLR 2024. arXiv:2307.13854.

Artifacts

The released repositories are the paper's appendices — each is the full, maintained form of the material a static appendix would snapshot:

• kamigotchi-gdd — the complete mechanics and data catalogs, agent-readable (§3.5).
• kami-harness — the environment interface; tool reference and observation schema in executor/README.md (§3.5).
• kami-agent — the model-agnostic reference scaffold for controlled studies; in final implementation (§4.1).
• kami-zero — the pilot agent's prompts, scaffolds, playbook rules (executor-prompt.md, optimizer-prompt.md, rules/), and architecture version history (§3.5).