Research Program for AI as Babel Fish for Structural Economics

Canonical location. This is the single source of truth for the AI as Babel Fish for Structural Economics research agenda, maintained as a public, forward-looking SCE document. It originated in the private econ-ark/OverARK-pvt repository and was migrated here in June 2026. Grant funding and Sloan-grant reporting/tracking are a separate concern and are not part of this agenda; that material remains in econ-ark/OverARK-pvt (Grants/SloanGrant2024/).

Name. “AI as Babel Fish for Structural Economics.” This document outlines a multi-year research program for the declarative-modeling-language (DML) / Bellman-calculus ecosystem under development in [bright-forest/bellman-ddsl](https://github.com/bright-forest/bellman-ddsl), [econ-ark/matsya](https://github.com/econ-ark/matsya), and the Econ-ARK toolkit (HARK, REMARK, DemARK).

The document is intended as a shared research agenda: a statement of problems, aims, milestones, and open questions that individual scholars or teams can cite when framing grant proposals. A typical formulation would be:

“This proposal extends the [cross-backend verification] component of the SCE Working Group on Methods, Backends & Verification, as described in the AI as Babel Fish for Structural Economics research program.”

How to join a working group¶

Anyone may join, and you are warmly invited to. Each working group has its own short sign-up form (a minute or two each). Open the form for any group you would like to join and submit it — you can join more than one group; just fill out each group’s form. No Google sign-in is required. Each form also asks, optionally, whether you would be willing to help co-lead the group or assist its chair.

• WG1 — Language, Calculus & Formal Semantics (chair: Akshay Shanker): sign-up form

• WG2 — Methods, Backends & Verification (chair: Herbert Dawid): sign-up form

• WG3 — AI Assistance, Reproducibility & Standards (chair: Alan Lujan): sign-up form

What each group does is described in §5 below.

1. Problem statement¶

Modern economics increasingly relies on dynamic structural models — agents who optimize over time under uncertainty (discrete-time Bellman problems), and heterogeneous-agent macroeconomics built from them. These models are interpreted quantitatively, can answer welfare and counterfactual questions, and are now dominant in top journals and increasingly used inside central banks.

Yet there remains no standard, shareable representation of such a model together with its computational solution. Each paper typically ships a bespoke codebase. The mathematics on paper is not formally bound to the code; numerical choices (grids, quadrature, interpolation, timing conventions) are often undocumented. Substantive conclusions can therefore depend silently on computational details — the canonical cautionary tale is a top-journal labor-supply result later found to rest on faulty numerical methods. As prominent structural modelers keep asking of each other’s work: “But what did they actually do?”

This program addresses that question by treating the model, the numerical methods, and the implementation as three separable, formally related artifacts, and by developing the language, theory, tooling, and standards required to make that separation operational.

2. Central thesis¶

The program rests on a single architectural claim. A dynamic structural model should be specified once, in a declarative language that is mathematically explicit and carries no numerical content. That surface syntax elaborates to a single typed core object (a “Bellman-core” / term-graph) with denotational semantics. From this core, one may derive, in principle independently: interchangeable numerical methods (“methodization”); translations to multiple solver backends (HARK, Dynare, dolo, VFI-toolkit, SSJ) via Rosetta Stones; and assisted authoring and interpretation through an AI system (Matsya) grounded in the typed semantics rather than in prose pattern-matching.

Where feasible, the semantic core should also have a machine-readable formal encoding in Lean, so that AI-assisted translation is symbolically grounded — constrained by typed representations and the encoded background theory rather than by fluent but unchecked text. The intended outcome is greater transparency, reproducibility, robustness, and interoperability in structural economics.

3. Intellectual foundations¶

The program draws explicitly on three traditions:

• Bellman — the origins of dynamic programming and the recursive value-function form.

• Sargent–Stachurski — modern recursive/dynamic-programming theory and the QuantEcon tradition (existence, convergence, contraction/monotone methods).

• Backus, FFP / functional programming — composition of primitive operators, referential transparency, and the discipline that keeps meaning separate from representation.

These provide, respectively, the theoretical object (Bellman, Sargent–Stachurski) and the design principle (Backus/FFP) that separates semantic content from numerical representation.

A complementary framing. Independently of the open-tooling motivation, the program also connects to a broader cross-disciplinary effort to encode the laws of a field in machine-readable formal form and to build AI systems that reason within those laws: formal mathematics (Lean / mathlib); physics (PhysLean / Lean4PHYS); physics-informed and scientific machine learning (PIML / SciML), in which governing equations enter as constraints; and symbolically grounded AI, in which a symbolic semantic layer constrains neural components. Economics is a natural setting for this approach because recursive/dynamic-programming theory is already well developed in prose and mathematics. Encoding that theory in a formal knowledge base (within WG1) and using it to ground AI-assisted modeling (WG3) is therefore a substantive research direction in its own right. See prior art §5.

4. Organization: SCE Working Groups¶

Each research strand is organized as a Society for Computational Economics (SCE) Working Group — a standing forum that scholars may join and cite. Proposals may refer to a specific component of a named working group. Governance is through a Working-Group Advisory Board drawn from the groups (see §8).

5. Working groups (research strands)¶

The program is organized into three working groups, corresponding to (i) semantic specification, (ii) numerical solution and verification, and (iii) assisted use, reproducibility, and standards. The groups are organized by layer. Each comprises several sub-threads, detailed in its track file. Some sub-threads are further advanced than others: the Lean foundation library (WG1), the journal-level refereeing workflow (WG3), and the cross-subfield integration agenda (cross-cutting; see below) are, at present, more aspirational than the surface language and backend translation work.

Working-group chairs (designated June 2026): WG1 — Akshay Shanker · WG2 — Herbert Dawid · WG3 — Alan Lujan.

To join any of these groups, use the sign-up forms in How to join a working group above.

The groups are logically ordered. WG1 supplies the semantic specification on which WG2 (numerical solution) and WG3 (assisted use and institutional uptake) depend; WG2 in turn supplies declared numerical methods that WG3 can subject to automated robustness analysis.

Cross-subfield integration: a cross-cutting strand. A recurring aim of the program is to let models and components from different economics traditions — heterogeneous-agent macro, structural microeconometrics, computable general equilibrium (CGE), agent-based modeling (ABM), and applied IO — be expressed, shared, and coupled in one representation, rather than re-implemented per tradition. Rather than a separate working group, this is pursued as an explicit workstream within each of the three: its representational half — shared and companion (non-Bellman) structures such as market clearing, search/matching, and behavioral rules, plus typed coupling interfaces — in WG1; its numerical half — Rosetta Stones to CGE/ABM/IO toolkits and verification of coupled models — in WG2; and AI-assisted translation of subfield literatures and cross-subfield reproducibility in WG3. Cross-field precedents (ESMF, OpenKIM, OPTIMADE, ASE, OpenFermion) are surveyed in prior art §6.

6. Time horizon and phases (5 years)¶

The program builds on existing prototypes (a surface language and Bellman-core, an executed Dolo Rosetta Stone, an early Matsya preview, and draft working papers). The five-year roadmap below takes these as its starting point.

• Phase 1 (Years 1–2): Consolidation. Stabilize the language and Bellman-core wire format; put the HARK Rosetta Stone into routine use; obtain first cross-backend verification results; seed the Lean formal-DP library and release an initial EconDP formalization benchmark; release Matsya publicly; seat the governance board. Establish the SCE working groups and institute an annual pre-conference meeting before each SCE (Computing in Economics and Finance) meeting as a standing forum for program development.

• Phase 2 (Years 2–4): Extension and verification. Support three or more verified backends; develop a methodization registry covering canonical method families; extend estimation and aggregation over modular models; build a reference library of REMARK reproductions expressed in the language; conduct central-bank pilots.

• Phase 3 (Years 4–5): Institutionalization. Arrive at a stable, governed standard; multi-institution adoption; integration of the robustness/refereeing workflow with at least one journal or data editor; a sustainable funding and governance model.

7. Relation to grant proposals¶

Scholars proposing work within this program would typically:

1. Identify a working group (WG1–WG3) and the specific sub-thread to be advanced.

2. Cite the program and group, e.g. “...builds on the [component] of the SCE Working Group on [name]...”

3. Select an open research question from that track’s agenda, or propose a closely related one.

4. Align the proposal with an appropriate funding source (each track notes relevant agencies and programs).

5. Use the track’s illustrative proposal as a starting template, if helpful.

Relevant funding channels across the program include NSF (economics, CISE, cyberinfrastructure), central banks / BIS (sponsorship and pilots), the JHU Data Science & AI Institute, and FRO / Convergent Research. Corporate sponsorship (e.g. T. Rowe Price-style support) may serve as a complementary channel.

8. Governance & sustainability¶

• Cadence: an annual pre-conference meeting before each SCE (CEF) meeting serves as the standing forum for program development, collaborator recruitment, and revision of the working groups. Successive pre-SCE meetings carry the agenda forward.

• Governance: a Working-Group Advisory Board drawn from the SCE working groups oversees the language standard and adjudicates breaking changes (loosely modeled on community language-governance practice such as Python’s PEP process, but seated in the working groups). Major decisions are recorded as architecture/standard decision records.

• Sustainability: several models are under consideration: (a) ongoing central-bank sponsorship of the standard and reference implementation; (b) FRO / Convergent Research support for the core team; (c) subscription/consulting support via NumFOCUS; (d) continued foundation and corporate grants tied to specific working groups.

9. Indicators of progress¶

Progress will be assessed, by working group, through: academic publications; adoption (downloads, models authored in the language, citing papers); the number of solver backends supported and verified; workshops and working-group participation; central-bank / institutional pilots; and REMARK reproductions expressed in the language.

10. Relation to existing tools¶

The program’s contribution, relative to these tools, is a software- and method-independent specification with a typed, formally encodable semantic core, from which numerical methods and backends are derived and verified rather than hand-wired, and within which AI-assisted translation is symbolically grounded. No existing economics toolkit currently combines portable structural specification, formal semantics, backend interoperability, and verified AI-assisted authoring in this form.

Prior art¶

Several mature fields have pursued closely analogous programs; a few began, as this one does, with a written strategic plan or community white paper that subsequently organized a multi-institution coalition and attracted large grants (NSF’s S2I2 “conceptualization → implementation” ladder explicitly funds the former; see §0 of the prior-art document). See Prior Art: Interoperable-Tooling Research Programs in Other Disciplines for analogues in optimization (JuMP/MathOptInterface), engineering (Modelica/FMI), statistics (Stan), quantum chemistry (MolSSI/QCSchema), quantum simulation/computing (OpenQASM/QIR/PennyLane/QuTiP), math formalization (Lean FRO/mathlib), astronomy (Astropy), and — most directly for WG1’s formal layer and WG3’s symbolically grounded AI — the AI-for-Science / formalization movement (PhysLean/Lean4PHYS, PIML/SciML; §5).

Track files¶

• WG1 — Language, Calculus & Formal Semantics

• WG2 — Methods, Backends & Verification

• WG3 — AI Assistance, Reproducibility & Standards

Open decisions¶

• Program / language name is deliberately deferred (placeholder: “Open Structural Modeling Language”). Candidates in play internally: Bellman calculus, DML, T-calculus / dolo-plus. To be resolved with the core team and Pablo Winant (dolo) re: fork vs. merge.

• Exact SCE working-group titles (chairs designated June 2026: WG1 — Akshay Shanker, WG2 — Herbert Dawid, WG3 — Alan Lujan).

• Which sub-threads to prioritize within each group as funding becomes available (e.g., the pace of the Lean foundation library in WG1, or journal-level refereeing in WG3).

• Substrate decision (WG1): whether the Bellman-core should be built on Lean (inheriting Lean’s type system and tooling, à la QIR-on-LLVM / PhysLean) or remain a bespoke term-graph with Lean as an export/target for formal encoding.

Provenance¶

Synthesized from the bellman-ddsl project plan, T-calculus docs, and spec-status overview; the matsya RAG design and system prompts; and the econ-ark.org mission content.