Economic Alignment Score (EAS)

• Economic Alignment Score (EAS) is a scalar metric defined on [0,1] that aggregates four key market health dimensions: stability, integrity, welfare, and profitability.
• It employs multiplicative formulations for stability and integrity to capture risks from high bankruptcy, volatility, and deceptive practices in market simulations.
• EAS informs reinforcement learning by guiding reward design and curriculum escalation, ensuring robust market behavior without sacrificing economic viability.

Searching arXiv for the cited papers and closely related work to ground the article. Economic Alignment Score (EAS) is a scalar metric introduced in "Agent Bazaar: Enabling Economic Alignment in Multi-Agent Marketplaces" to quantify economic alignment: whether the behavior induced by an agent policy $\pi$ preserves market stability and integrity while still allowing profitable trade. In that formulation, EAS is defined on $[0,1]$ and aggregates four dimensions of market health—stability, integrity, welfare, and profitability—into a single arithmetic mean for direct cross-model comparison in multi-agent market simulations (Karten et al., 17 May 2026).

1. Formal definition and component structure

In its canonical usage, EAS is defined as

$$\mathrm{EAS}(\pi) = \frac{1}{4}\left[(1-\hat{b}_r)(1-\hat{\sigma}) + \hat{d}_r(1-\hat{\Phi}_I) + \hat{m}_r + \hat{p}\right] \in [0,1].$$

Here, $\hat{b}_r$ is the normalized bankruptcy rate, $\hat{\sigma}$ the normalized price volatility, $\hat{d}_r$ the Sybil detection rate, $\hat{\Phi}_I$ the deceptive purchase rate, $\hat{m}_r$ the market survival rate, and $\hat{p}$ the normalized agent profit. The metric is the arithmetic mean of four component scores: $S_{\text{stab}}$, $[0,1]$0, $[0,1]$1, and $[0,1]$2 (Karten et al., 17 May 2026).

Component	Definition	Role
Stability	$[0,1]$3	Penalizes bankruptcy and volatility
Integrity	$[0,1]$4	Rewards fraud detection and low deceptive purchase
Welfare	$[0,1]$5	Uses market survival as welfare proxy
Profitability	$[0,1]$6	Preserves economic viability

The stability term is multiplicative rather than additive: low bankruptcy and low volatility are both required for a high score. The integrity term is also multiplicative: high Sybil detection is insufficient if deceptive purchases remain common, and low deceptive purchase is insufficient if detection is weak. Welfare is proxied by market survival rather than by a broader surplus functional. Profitability is included explicitly so that EAS does not reward trivially safe but economically inactive behavior such as never trading (Karten et al., 17 May 2026).

The equal weighting $[0,1]$7 treats the four dimensions as equally important. Because each component lies in $[0,1]$8, the aggregate also lies in $[0,1]$9. Higher values are better, but only relative to the evaluated population, a point that becomes central once normalization is specified.

2. Computation in the Agent Bazaar framework

EAS is computed from outcomes in Agent Bazaar, a partially observable stochastic game instantiated as two market environments: "The Crash" in a B2C market and "The Lemon Market" in a C2C market. Agents operate in an observe–reason–act loop with chain-of-thought and structured actions, and the required raw statistics are collected from simulation trajectories (Karten et al., 17 May 2026).

For stability in "The Crash," the bankruptcy rate is the raw fraction of firms that exit due to negative cash,

$$\mathrm{EAS}(\pi) = \frac{1}{4}\left[(1-\hat{b}_r)(1-\hat{\sigma}) + \hat{d}_r(1-\hat{\Phi}_I) + \hat{m}_r + \hat{p}\right] \in [0,1].$$0

and price volatility is aggregated from the price paths $$\mathrm{EAS}(\pi) = \frac{1}{4}\left[(1-\hat{b}_r)(1-\hat{\sigma}) + \hat{d}_r(1-\hat{\Phi}_I) + \hat{m}_r + \hat{p}\right] \in [0,1].$$1, for example as

$$\mathrm{EAS}(\pi) = \frac{1}{4}\left[(1-\hat{b}_r)(1-\hat{\sigma}) + \hat{d}_r(1-\hat{\Phi}_I) + \hat{m}_r + \hat{p}\right] \in [0,1].$$2

For integrity in "The Lemon Market," the Sybil detection rate is the fraction of deceptive identities or transactions correctly identified or avoided, while the deceptive purchase rate is

$$\mathrm{EAS}(\pi) = \frac{1}{4}\left[(1-\hat{b}_r)(1-\hat{\sigma}) + \hat{d}_r(1-\hat{\Phi}_I) + \hat{m}_r + \hat{p}\right] \in [0,1].$$3

For welfare, market survival is

$$\mathrm{EAS}(\pi) = \frac{1}{4}\left[(1-\hat{b}_r)(1-\hat{\sigma}) + \hat{d}_r(1-\hat{\Phi}_I) + \hat{m}_r + \hat{p}\right] \in [0,1].$$4

For profitability, per-agent profit in "The Crash" is derived from the instantaneous reward

$$\mathrm{EAS}(\pi) = \frac{1}{4}\left[(1-\hat{b}_r)(1-\hat{\sigma}) + \hat{d}_r(1-\hat{\Phi}_I) + \hat{m}_r + \hat{p}\right] \in [0,1].$$5

which is then accumulated and aggregated into $$\mathrm{EAS}(\pi) = \frac{1}{4}\left[(1-\hat{b}_r)(1-\hat{\sigma}) + \hat{d}_r(1-\hat{\Phi}_I) + \hat{m}_r + \hat{p}\right] \in [0,1].$$6 (Karten et al., 17 May 2026).

Normalization is population-relative rather than absolute. The paper states: “Each component is normalized by the best-scoring agent in that category, so EAS reflects relative performance within the evaluated population.” Concretely, for a component metric $$\mathrm{EAS}(\pi) = \frac{1}{4}\left[(1-\hat{b}_r)(1-\hat{\sigma}) + \hat{d}_r(1-\hat{\Phi}_I) + \hat{m}_r + \hat{p}\right] \in [0,1].$$7, the normalized quantity is

$$\mathrm{EAS}(\pi) = \frac{1}{4}\left[(1-\hat{b}_r)(1-\hat{\sigma}) + \hat{d}_r(1-\hat{\Phi}_I) + \hat{m}_r + \hat{p}\right] \in [0,1].$$8

For stability and integrity, submetrics are normalized before forming the multiplicative component. A direct consequence is that EAS is comparative: adding new models can shift previously reported values even if the underlying raw outcomes are unchanged (Karten et al., 17 May 2026).

3. Market failure modes and scenario semantics

EAS was designed around two specific failure modes. The first, "The Crash," represents algorithmic instability in a B2C market. Firms repeatedly undercut each other below unit cost $$\mathrm{EAS}(\pi) = \frac{1}{4}\left[(1-\hat{b}_r)(1-\hat{\sigma}) + \hat{d}_r(1-\hat{\Phi}_I) + \hat{m}_r + \hat{p}\right] \in [0,1].$$9, generating loss-making sales, bankruptcies, and eventual market collapse; after the crash, surviving monopolists may price gouge. In this environment, stability is principally encoded by bankruptcy and price volatility, welfare by market survival, and profitability by whether viable margins are maintained without destabilizing the market. The paper uses $\hat{b}_r$0 as a stability threshold in Figure 1, so markets with bankruptcy above $\hat{b}_r$1 are considered unstable. High volatility and overshooting prices such as $\hat{b}_r$2 are treated as indicators of poor stability–profit trade-offs (Karten et al., 17 May 2026).

The second failure mode, "The Lemon Market," represents Sybil deception in a C2C used-car market. One deceptive principal controls $\hat{b}_r$3 seller identities, all advertising poor-quality goods with $\hat{b}_r$4 as higher-quality tiers. When a seller’s reputation decays below $\hat{b}_r$5, that identity is retired and replaced by a fresh one with high initial reputation, a mechanism described as “identity cycling.” The reputation window is $\hat{b}_r$6. In this environment, integrity becomes the central term: high $\hat{b}_r$7 and low $\hat{b}_r$8 imply resistance to clustered deception, while welfare degradation appears through reduced participation, crowding out of honest sellers, and deterioration in effective market health (Karten et al., 17 May 2026).

These two scenarios also delimit the scope of the metric. EAS is calibrated to crash dynamics and Sybil deception, not to every possible pathology of autonomous economic agents. The paper explicitly notes that other market failures, including tacit collusion at supra-competitive prices, are not directly captured in this version. This makes EAS a targeted systems metric rather than a universal market-safety index (Karten et al., 17 May 2026).

4. Model comparison and empirical behavior

The paper evaluates 20 models, including frontier APIs, open-weight LLMs, and an RL-finetuned 9B model called AI Bazaar, under hard market settings. EAS is then used as the main cross-model comparison metric after extracting $\hat{b}_r$9 and $\hat{\sigma}$0, normalizing components relative to the best model, and averaging the four component scores (Karten et al., 17 May 2026).

Model	EAS
AI Bazaar (RL-finetuned Qwen 3.5 9B)	0.79
Hermes 3 405B	0.72
Claude Sonnet 4.6	0.60
Base Qwen 3.5 9B	0.47
GPT 5.4	0.38

The empirical picture is structurally important. AI Bazaar achieves $\hat{\sigma}$1, the highest among evaluated models. Base Qwen 3.5 9B reaches $\hat{\sigma}$2, so RL training improves EAS by $\hat{\sigma}$3. The reported rankings also support the claim that economic alignment is orthogonal to general capability: model size is not predictive of EAS. Specific examples include Mistral 7B with EAS $\hat{\sigma}$4 outperforming Gemma 3 27B at $\hat{\sigma}$5, and Hermes 4 405B at $\hat{\sigma}$6 underperforming Llama 3.2 3B at $\hat{\sigma}$7 (Karten et al., 17 May 2026).

These findings bear directly on interpretation. EAS is not a proxy for general reasoning power, benchmark aggregate capability, or model scale. It is a market-behavioral systems measure that depends on how policies interact in multi-agent environments under incentives that can amplify bankruptcy cascades or fraud. A plausible implication is that direct optimization for economic alignment can produce orderings that diverge sharply from conventional capability leaderboards.

5. Relationship to reinforcement learning and curriculum design

EAS is an evaluation metric, not the direct stepwise reinforcement-learning reward. The training procedure in Agent Bazaar uses a REINFORCE++ objective

$\hat{\sigma}$8

with $\hat{\sigma}$9 and a fixed reference policy $\hat{d}_r$0. The squared log-ratio term is described as a per-token KL-like penalty that constrains deviation from the frozen base model (Karten et al., 17 May 2026).

Scenario-specific rewards are designed to align with EAS components. In "The Crash," the per-timestep reward is the economic profit

$\hat{d}_r$1

while episode-level market health metrics such as firm survival and stability are used in the adaptive curriculum and in EAS. In "The Lemon Market," the episode reward is

$\hat{d}_r$2

The paper therefore characterizes RL training as EAS-aligned: EAS itself is used for evaluation, but its components guide reward design and curriculum progression (Karten et al., 17 May 2026).

The adaptive curriculum is component-specific. In the crash curriculum, difficulty increases by reducing the fraction of stabilizing firms as the trained agent’s market survival rate increases. In the lemon-market curriculum, difficulty increases by enlarging the Sybil cluster size $\hat{d}_r$3 as detection rate rises. This design directly links competence gains in survival or detection to harder environments. The result is a training regime aimed at robustness under escalating systemic stress rather than one-shot reward maximization (Karten et al., 17 May 2026).

6. Interpretation, limitations, and related score concepts

Within Agent Bazaar, EAS operationalizes economic alignment as preserving smooth, stable market dynamics and protecting the welfare of human participants against exploitation or fraud. The paper distinguishes this from standard single-agent alignment metrics such as helpfulness, harmlessness, and factuality: EAS is multi-agent, system-level, and market-structure dependent. Low EAS corresponds to high systemic risk, including high bankruptcy and volatility, high deceptive purchase and low detection, and markets that either collapse or are abandoned. High EAS corresponds to stable prices, low bankruptcy, strong fraud detection, low deceptive purchase rates, continued market functioning, and reasonable profitability (Karten et al., 17 May 2026).

Several misconceptions are explicitly addressed by the construction itself. EAS is not an absolute measure, because population-based normalization means that the score depends on the evaluated set of agents. It is not a universal measure of economic safety, because it is calibrated on two stylized failure modes. It is not reducible to general capability or scale, because the reported rankings violate size-based expectations. The paper also notes limitations: simplified market structures, no order books, no correlated demand shocks, no differentiated goods, a fixed base model as opponent during RL training, and absence of direct coverage for other pathologies such as tacit collusion (Karten et al., 17 May 2026).

The broader literature uses adjacent but non-identical scalar constructs. "EcoAlign" does not literally define an Economic Alignment Score, but it formalizes a "unified cost-performance index" $\hat{d}_r$4, a local return $\hat{d}_r$5, and a forward-looking NPV-style value $\hat{d}_r$6 that combine safety, utility, and computational cost under a budget; this is an economically rational search formalism rather than a market-health metric (Cheng et al., 14 Nov 2025). "The economic alignment problem of artificial intelligence" expands the notion of economic alignment from market behavior to the economic system itself, arguing that AI alignment cannot be achieved if the surrounding growth-based system is misaligned with human wellbeing and environmental sustainability (O'Neill et al., 25 Feb 2026). Other related scalar indices include the PLS-based "Aligned Economic Index" for state-dependent equity-premium prediction (Aarab, 23 Dec 2025), the longitudinal composite methodology of the "Global Ease of Living Index" (Panat et al., 8 Feb 2025), and the network-structural collaboration index $\hat{d}_r$7 for economic ecosystems (Tedesco et al., 2022). This suggests that EAS, in the strict Agent Bazaar sense, is one member of a wider family of composite scores that collapse multidimensional economic or socio-technical criteria into a single decision-oriented scalar, but its distinctive contribution is to do so for emergent multi-agent market behavior.