Luyu Zhang · LangGenius, Inc. · March 2026

TL;DR

Modern knowledge work is full of friction — bureaucratic document requirements, endless status meetings, costly verification processes. We've been blaming bad tools and poor communication. The real cause is deeper: we lack a fundamental infrastructure layer for representing the state of real-world things.

We call this the State Infrastructure Gap (SIG).

This paper introduces a cross-disciplinary framework — drawing from information economics, AI knowledge representation, and digital twin engineering — to explain why this gap exists, why it has persisted for decades, and why Large Language Models are the first technology that makes closing it possible.

The Core Insight (in 60 seconds)

The argument unfolds in six steps:

1. White-collar work, globally, converges on three activities: writing documents, organizing data, coordinating communication.

2. These three activities serve one underlying purpose: letting another party know the current state of some entity.

3. CRUD-heavy jobs (accounting, HR, supply chain, customer service) are essentially "human state synchronizers" — biological middleware between reality and digital systems.

4. Tedious verification processes (visa applications, loan approvals, due diligence) exist because there is no trusted source to query an entity's state directly. So we reconstruct state from documents.

5. This problem exists in both structured state (fits in a database) and unstructured state (lives in human heads and conversations).

6. LLMs make machine-processing of unstructured state possible for the first time — but the infrastructure to harness this doesn't exist yet.

Key Concepts

State ≠ Data

Data is a recorded value — static, discrete, context-free. A database field reading status: approved is data. State is the actual condition of a real-world entity at a particular point in time. It has four properties data lacks:

What conventional databases actually store are compressed traces of state — the last accepted snapshot after uncertainty, ambiguity, and source conflict have been suppressed. This isn't a bug; it's an ontological commitment. Structured systems trade semantic richness for consistency and queryability. That trade-off is often useful. It becomes costly when the underlying state can't be cleanly reduced without material loss of meaning.

The consequence: organizations routinely confuse the existence of data with the existence of state. A CRM record may show a deal stage, but not the confidence behind it, the evidence supporting it, or the dependencies threatening it. A ticket may say "in progress" while the real state — blocked on legal, deprioritized by a manager, or effectively abandoned — lives elsewhere.

The Human State Synchronizer

A new concept introduced in this paper. Refers to workers whose primary function is to serve as a translation layer between reality and digital systems — detecting state changes in the real world and manually entering them into systems that cannot detect those changes autonomously.

This is analytically distinct from both Nonaka's "knowledge worker" (whose value lies in knowledge creation) and Tushman's "boundary spanner" (who brokers across organizational boundaries). The human state synchronizer is defined more narrowly: converting observed, inferred, or locally held conditions into records that other people or systems can act on.

This extends far beyond back-office roles into sales ops, project coordination, compliance, recruiting, customer success, and managerial reporting. These jobs exist not because the labor is intrinsically valuable, but because the infrastructure for automated state capture doesn't exist. Organizations have misidentified a large class of derived labor as intrinsic work. In many cases, people are hired to compensate for missing infrastructure.

State Reconstruction Cost

When you can't query an entity's state directly, you pay the cost of reconstructing it from indirect evidence:

Every visa application, every loan review, every audit, every status meeting — is a state reconstruction activity. In each case, artifacts function as evidence chains rather than ends in themselves: a bank statement stands in for a financial state, an employment letter stands in for occupational continuity, a burn-down chart stands in for project progress.

Three Domains, One Blind Spot

The SIG framework unifies three bodies of theory that have independently addressed fragments of this problem but never talked to each other:

1. Information Economics

Akerlof (1970), Spence (1973), Stiglitz & Weiss (1981)

explains why state verification is expensive. Visa documents are Akerlof signals. University degrees are Spence signals. Loan applications are the verification mechanisms Stiglitz analyzed. A signal is credible not because it is intrinsically informative, but because its cost structure makes imitation difficult. What it misses: it treats information cost as exogenous and doesn't ask what happens when technology radically reduces state verification cost.

2. Knowledge Representation (AI)

Davis et al. (1993), Minsky (1975), Hayes (1979), Nonaka (1994), Orlikowski (2002)

every knowledge representation is a set of ontological commitments determining what can be expressed. Relational databases trade semantic richness for queryability. Orlikowski went further: knowing in organizations is an ongoing accomplishment of practice, not a static stock to be captured — which helps explain why so much relevant state stays informal. What it misses: KR research focuses on reasoning systems, not on organizational state synchronization infrastructure.

3. Digital Twins (Engineering)

Grieves & Vickers (2017), Tao et al. (2019), Listl et al. (2024)

the closest existing paradigm, maintaining bidirectional real-time synchronization between physical entities and digital representations. They show what becomes possible when state synchronization is built into the substrate. What it misses: Digital twins rely on sensors. Knowledge-work states are inherently judgmental and can only be captured through natural language. The problem is not the absence of telemetry alone; it is the absence of semantic, confidence-aware telemetry.

When state descriptions are expressed in natural language — rich in context, nuance, and subjective judgment — how can they be captured, stored, synchronized, and queried at scale?

No field had answered this, because the technological prerequisite didn't exist — until LLMs.

The State Infrastructure Gap: Four Dimensions

Semantic

Storage systems can't preserve the semantic richness of real-world states. "Project mostly on track but auth module has unresolved design disagreement" collapses to status: yellow.

Confidence

All records treated as equally authoritative. A Jira ticket updated 30 min ago by the engineer ≡ one updated last week by a PM from an old email.

Synchronization

Updates are human-triggered. State changes in meetings, Slack, code reviews don't auto-propagate to tracking systems.

Aggregation

State is scattered across dozens of disconnected tools. The "true" customer state requires synthesizing CRM + email + support + billing — always by a human.

The Real Cost: Four Economic Consequences

State reconstruction cost isn't just a formula. It cascades through organizations in four distinct ways:

1. Derived Labor Demand

Where state can't be directly queried, organizations hire people to collect, translate, reconcile, and move it. This is the labor market footprint of human state synchronization — extending far beyond back-office roles.

2. Structural Delay

Query-based systems answer immediately. Reconstruction-based systems answer only after evidence has been assembled and interpreted. Time becomes part of the price of trust.

3. Degraded Org Intelligence

When state is fragmented across tools and human memory, decision quality depends on access to the right people at the right moment. Employee turnover breaks the institution's queryable memory.

4. Distorted Management

Because knowing what's actually happening is expensive, organizations substitute proxy metrics and reporting rituals. Goodhart's Law kicks in: teams optimize what can be entered rather than what is true.

Why the SIG Has Persisted

Three reinforcing factors:

1. Technological constraint. Before LLMs, natural-language state descriptions couldn't be machine-processed. Organizations faced a binary choice: sacrifice semantics for queryability (databases) or preserve semantics at the cost of machine access (documents). In practice they did both, producing the now-familiar landscape of partially structured systems surrounded by manual coordination.

2. Economic inertia. Local workarounds — a better dashboard, a stricter workflow, another integration — were often sufficient. SIG costs are real but diffuse, dispersed across applicants, analysts, managers, and partners. Because no single budget owner fully internalizes these costs, incremental software improvement dominates infrastructural redesign. Classic collective action problem.

3. Institutional resistance. State standards are not politically neutral. Whoever defines what counts as valid evidence, how states are categorized, and who has authority to update them wields real power. A system that makes state directly queryable reduces not only labor but also gatekeeping authority, information rents, and procedural leverage. This is why adoption of better state infrastructure is not merely a technical optimization — it also redistributes control.

LLM as Enabling Technology

LLMs dissolve the trade-off between semantic richness and machine processability:

Semantic processing of unstructured state. LLMs can process meeting transcripts, email threads, chat conversations, and extract structured state information without human intermediation. In pre-LLM architectures, natural language was where semantics went to die: stored for human reading, ignored by system logic. With LLMs, natural language becomes a first-class medium of operational state.

From event streams to state hypotheses. Much organizational state is never explicitly declared — it is implicit in event streams. Code commits, test results, calendar events, emails, approvals, chat messages all create traces of changing conditions. A project need not wait for a manager to write "at risk" if the system can observe repeated slip signals, unresolved dependencies, and stalled tickets. An account need not wait for a human to manually downgrade health if usage collapse and unanswered escalations are already present in the evidence. These aren't perfect judgments, but they are often better than no shared state at all — especially when the system surfaces the evidence chain rather than only the predicted label.

Confidence estimation. A useful state infrastructure can't merely produce a label; it must indicate how much the label should be trusted. LLMs can participate in a confidence pipeline by aggregating provenance cues: source type, diversity, recency, consistency, and corroboration. "Vendor compliant, verified yesterday from two audited systems" and "vendor compliant, inferred from a self-attested PDF last quarter" are not equivalent states. The system should make this difference visible.

What LLMs Cannot Solve Alone

The first-mile problem: LLMs can't verify that information entering the system is truthful. If the original observation is false, strategically manipulated, or never recorded, downstream inference can't restore ground truth. Analogous to the oracle problem in blockchain.

Hallucination: LLMs can confidently assert states that don't correspond to reality. State infrastructure must incorporate confidence calibration, evidence grounding, and human-in-the-loop verification.

Legitimacy: Many verification processes exist within regulatory frameworks requiring human judgment and accountability. The question is not only whether a model can infer a state, but whether relevant actors will accept that inference as a valid basis for decision.

Bottom line: LLMs are enabling technology, not a solution. Connecting an LLM to a database doesn't create state infrastructure, any more than connecting a sensor to a spreadsheet creates a digital twin. The most promising architectures are hybrid: formal systems for deterministic fields, permissions, and transactional integrity; LLMs where meaning must be interpreted across heterogeneous traces. Not replacement, but a layered design where semantic inference and formal record-keeping inform each other.

Requirements for New State Infrastructure

A viable state infrastructure must satisfy five properties simultaneously:

The Fundamental Representational Shift

The basic unit of this infrastructure is not the row, the document, or the task. It is the state claim — a dated, evidence-linked, provenance-aware representation of an object's condition that can be revised, contested, and queried. Some state claims are highly structured and deterministic. Others remain narrative and probabilistic. A useful infrastructure supports both without collapsing one into the other prematurely.

This matters because it acknowledges the epistemological reality: there is no absolute state, only state claims of varying confidence. A project can be "on track" for one stakeholder and "at risk" for another depending on which deadline or quality threshold matters. The infrastructure must accommodate provenance-rich, revisable, and sometimes plural state claims rather than a mythical frictionless source of absolute truth.

A Distributed Systems Analogy

This is not an improvement to existing tools. It is a new infrastructure layer beneath them.

Distributed systems research offers a useful parallel. In robust distributed systems, convergence and conflict handling are built into the substrate rather than delegated to end users. CRDT research (Shapiro et al., 2011) shows that replicated state can converge under explicit rules rather than constant human reconciliation. Kleppmann (2017) emphasizes that data-intensive systems succeed when consistency, lineage, and coordination semantics are handled at the infrastructural layer.

Knowledge work today asks humans to do the equivalent by hand: compare versions, infer updates, reconcile discrepancies, and decide what is current. A state infrastructure paradigm moves that burden downward into the system — the same architectural move that TCP/IP made for networking and that CRDTs made for distributed data.

Implications

For Work

• Human State Synchronizer labor will shrink — not because AI does the same job better, but because the job disappears when state capture is automated. Human effort shifts from routine synchronization to exception handling, audit, and judgment calls where discretion is genuinely needed.

• Organizational tacit knowledge becomes capturable for the first time — making coordination more resilient to turnover and scale.

• AI agents gain persistent, shared memory — the prerequisite for multi-agent collaboration, where multiple agents query the same state layer over persistent objects, histories, and confidence levels.

For Organizations

• Investing in "better tools" addresses symptoms. The root cause is architectural.

• The highest-value LLM application may not be content generation but state capture and synchronization.

• Watch for Goodhart effects: if your teams are optimizing what can be entered rather than what is true, you're paying the cost of SIG in decision quality.

For Policy

• The workers most vulnerable to displacement aren't those with the most skilled jobs, but those whose jobs exist because of infrastructure deficiency.

Limitations

• This is a theoretical framework, not empirically validated.

• The confidence dimension needs formal specification (building on Gärdenfors, Halpern, epistemic logic).

• Different organizational contexts face different SIG profiles.

• LLM state-inference accuracy hasn't been systematically measured in organizational contexts.

• Many states are contestable or perspective-dependent — the framework acknowledges this but does not yet provide a formal mechanism for adjudicating conflicting state claims.

Future Research

• Empirical measurement of state reconstruction costs.

• Systematic LLM state-inference evaluation.

• Institutional and political factors in adoption.

• Formal confidence calculus.

• Comparative case studies of automated vs. traditional state capture.

• Governance models for state infrastructure: who decides what counts as a valid state claim, and how disputes are resolved.

References

• Akerlof, G. A. (1970). The market for "lemons." QJE, 84(3), 488–500.

• Arrow, K. J. (1963). Uncertainty and medical care. AER, 53(5), 941–973.

• Brachman, R., & Levesque, H. (2004). Knowledge Representation and Reasoning. Morgan Kaufmann.

• Davenport, T., & Prusak, L. (1998). Working Knowledge. HBS Press.

• Davis, R., Shrobe, H., & Szolovits, P. (1993). What is a knowledge representation? AI Magazine, 14(1), 17–33.

• Drucker, P. F. (1959). Landmarks of Tomorrow. Harper & Row.

• Gärdenfors, P. (1988). Knowledge in Flux. MIT Press.

• Grieves, M., & Vickers, J. (2017). Digital twin. In Transdisciplinary Perspectives on Complex Systems (pp. 85–113).

• Halpern, J. Y. (2003). Reasoning About Uncertainty. MIT Press.

• Hayes, P. J. (1979). The naive physics manifesto. In Expert Systems in the Microelectronic Age.

• Ji, Z., et al. (2023). Survey of hallucination in NLG. ACM Computing Surveys, 55(12), 1–38.

• Kadavath, S., et al. (2022). Language models (mostly) know what they know. arXiv:2207.05221.

• Kleppmann, M. (2017). Designing Data-Intensive Applications. O'Reilly.

• Listl, F. G., et al. (2024). Knowledge graphs in the digital twin. IEEE Access.

• Minsky, M. (1975). A framework for representing knowledge. In The Psychology of Computer Vision (pp. 211–277).

• Nonaka, I. (1994). A dynamic theory of organizational knowledge creation. Organization Science, 5(1), 14–37.

• Orlikowski, W. (2002). Knowing in practice. Organization Science, 13(3), 249–273.

• Packer, C., et al. (2024). MemGPT: Towards LLMs as operating systems. arXiv:2310.08560.

• Park, J. S., et al. (2023). Generative agents. ACM UIST '23.

• Shapiro, M., et al. (2011). Conflict-free replicated data types. SSS (pp. 386–400).

• Spence, M. (1973). Job market signaling. QJE, 87(3), 355–374.

• Stiglitz, J. E., & Weiss, A. (1981). Credit rationing. AER, 71(3), 393–410.

• Tao, F., et al. (2019). Digital twin-driven product design. Int'l J. Adv. Manufacturing Tech., 94, 3563–3576.

• Tushman, M. L. (1977). Special boundary roles. ASQ, 22(4), 587–605.

• Williamson, O. E. (1975). Markets and Hierarchies. Free Press.

Luyu Zhang is the CEO of LangGenius, Inc.

This paper is released for open discussion. Comments and critiques welcome via Issues or Discussions.