Dysmemic Pressure: Selection Dynamics in Organizational Information Environments

Executive Summary

Organizations staffed by intelligent people converge on collective delusion with predictable regularity. This paper identifies a structural mechanism, dysmemic pressure: the compound selection force that emerges when organizations compress information to coordinate at scale. Three dynamics drive it: strategic communication degradation (Crawford and Sobel, 1982), adverse selection in idea markets (Akerlof, 1970), and transmission bias (Boyd and Richerson, 1985).

The paper formalizes the first dynamic through a hierarchical Crawford-Sobel model and validates it with agent-based Monte Carlo simulation. In the single-channel case, as preference divergence b approaches 1/4 the equilibrium partition count falls to one and mutual information goes to zero: the channel degenerates to babbling. In the hierarchical case, a sweep over 126 configurations shows that without noise the partition structure is preserved, but with noise mutual information degrades monotonically, as the data processing inequality predicts.

The Dysmemic Pressure Index DP(k) gives a measurable index of information destruction at each hierarchical layer. The analysis reframes organizational dysfunction from moral failure to selection outcome, explaining why standard interventions fail and what structural countermeasures might succeed.

Crawford-Sobel Derivation

Crawford and Sobel (1982) established that when a sender with private information communicates to a receiver whose preferences differ by a bias parameter b, equilibrium communication takes the form of partition signaling: the state space is divided into discrete intervals, and the sender reports only which interval contains the true state. Proposition 3.1 gives the maximum equilibrium partition count N*(b) for uniform priors and quadratic loss; as bias increases, the partitions become fewer and coarser. Proposition 3.2 shows that as b approaches 1/4, N*(b) falls to 1 and the sender's message becomes statistically independent of the true state.

Applied to organizations, the bias parameter corresponds to incentive misalignment between hierarchical levels. Distinctions within a partition are lost; the loss is structured rather than random. The implementation is validated against theory: at b = 0.1 the most-informative equilibrium has three partitions and a theoretical mutual information of 0.8813 bits, and the Monte Carlo simulation recovers 0.8778 bits.

Hierarchical Information Loss Model

The hierarchical model chains Crawford-Sobel channels: each layer observes only the previous layer's action (not the ground truth) and communicates through its own biased channel, with optional Gaussian noise. By the data processing inequality, each layer of strategic communication can only destroy information about ground-level reality, never create it.

The parameter sweep covers 7 hierarchy depths (3 to 10 layers), 6 per-layer bias levels (0.01 to 0.24), and 3 noise levels, for 126 configurations at 50 Monte Carlo iterations each. Without noise, partition structure is preserved through all depths. With noise, mutual information decays monotonically with depth. A phase diagram separates an informative regime (low bias, shallow hierarchies) from a blind regime (high bias, or deep hierarchies with noise). Near the babbling boundary (b = 0.24), even the first layer transmits only about 0.14 bits, and a 10-layer hierarchy leaves the top with virtually no information about ground truth.

The Dysmemic Pressure Index

The Dysmemic Pressure Index is defined as DP(k) = 1 - I(report_k; theta) / I(report_0; theta), where theta is the ground truth, report_k is the information available at layer k, and report_0 is direct observation. DP(k) = 0 means no information loss relative to direct observation (a transparent hierarchy); DP(k) = 1 means the report at layer k is independent of ground truth (a blind hierarchy); intermediate values measure the fractional loss.

Case Studies

Three case studies illustrate the mechanism across industries; the paper treats them as illustrations, not controlled tests:

• Boeing 737 MAX: The Lion Air and Ethiopian Airlines crashes killed 346 people. Engineers had identified MCAS as a single-point-of-failure risk, and test pilots had experienced uncommanded nose-down events in simulation. Mapped to the model, the information was compressed through the hierarchy (engineer to lead, program manager, division head, CEO, board), with each layer's bias oriented toward schedule and certification; the report that reached the top was "certified safe," a dysmemic pressure index near 1.
• Theranos: Elizabeth Holmes created an information environment with extreme preference divergence. Engineers who raised concerns about test accuracy were fired or threatened, so the remaining internal signals were optimized for internal fitness rather than correspondence with reality. Only external intervention, beginning with Carreyrou's investigative journalism, penetrated the information environment.
• Wirecard: The company collapsed in June 2020 when auditors could not confirm EUR 1.9 billion in cash balances. EY relied on confirmations supplied by Wirecard rather than independent verification, and BaFin filed a criminal complaint against the Financial Times journalists and short sellers rather than investigating their evidence. The case shows dysmemic pressure cascading beyond a single firm into its institutional ecosystem.

Engagement with Forecasting Literature

Tetlock (2005) found that "foxes," who draw on multiple frameworks and update frequently, significantly outperformed "hedgehogs," who organize their thinking around a single grand theory. In the dysmemic pressure framework, hedgehog organizations exhibit stronger pressure: a single strategic narrative becomes the selection criterion for internal signals, and the partition structure coarsens around the narrative boundary. Fox organizations resist through structural pluralism.

The superforecasting teams documented by Tetlock and Gardner (2015) map onto anti-dysmemic architecture: independent assessment before discussion, aggregation without compression, proper scoring rules that push the bias parameter toward zero, frequent updating, and cognitive diversity. The paper's proposed countermeasures follow the same logic: evaluation decoupling, external validation, prediction tracking with proper scoring rules, and protected channels for dissent.

Key References