Gated human oversight makes LLM-driven economic theory work more auditable and generally more useful: in five matched tasks a gated pAI-Econ-claude workflow cut mean failure severity from 1.58 to 1.16 and raised usefulness from 2.60 to 3.10, with blinded evaluators preferring the gated system in four cases; but scaffolding can also over-constrain models and erase economically important mechanisms.

Main Finding

A gated, human-in-the-loop multi-agent workflow (pAI-Econ-claude) that stages economic-theory development, inserts targeted model-based gates, and concentrates irreversible human judgment improves the auditability and usefulness of LLM-assisted economic-theory outputs relative to an ungated baseline. In a five-task matched evaluation, two blinded evaluators preferred the gated workflow in four tasks (baseline preferred in one); mean failure severity dropped from 1.58 → 1.16 and mean usefulness rose from 2.60 → 3.10. The architecture increases inspectability and catches many practical failure modes, but it does not and cannot substitute for task-complete formal verification.

Key Points

• Failure taxonomy (F1–F5):
  • F1 Canonical mismatch (wrong model family)
  • F2 Trivial propositions (formal but empty results)
  • F3 Hidden proof gaps (skipped cases, unjustified signs)
  • F4 Interpretive overreach (unsupported welfare/institutional claims)
  • F5 Unreliable citations (fabricated/misattributed sources)
• Three design principles:
  • P1 Intercept failures while the relevant choices remain visible (staged persistent outputs).
  • P2 A checker without an oracle can diagnose but must not certify (gates emit PASS/REFRAME, not correctness certificates).
  • P3 Concentrate human judgment at points of highest irreversibility (e.g., fixing the equilibrium concept).
• System architecture:
  • Stage-based pipeline (Stage 0–10 with sub-stages 2a, 3b, optional 7b). Each stage writes an inspectable file to a shared workspace.
  • Specialized agents: stage agents produce outputs; gate agents evaluate and recommend loopbacks; adversarial agents (reality check, counterexample finder); a persona council (mechanism, formalization, positioning, skeptic) for model-lineage debate.
  • Gates target specific failure modes and return PASS/REFRAME verdicts with recommended loopbacks and severity; humans adjudicate and may override.
  • Human checkpoints (HiL-1..HiL-6; HiL-N1..N3 for numerical work) — HiL-4 (equilibrium choice) is an unconditional hard stop.
  • Coordination via a blackboard/shared workspace: persistent, inspectable state that supports audit trails and human edits.
• Evaluation:
  • Five matched tasks across economic domains: human capital, health economics, nutrition & macro-development, consumer information, agri-food system transformation.
  • Matched-pair design: same underlying model and user-level task; compare gated workflow vs ungated baseline.
  • Two blinded evaluators scored finished manuscripts and agreed on all five pairwise comparisons.
  • Outcome metrics: pairwise preference, failure-severity scores, overall usefulness.
• Empirical outcomes:
  • Gated architecture preferred in 4 of 5 tasks.
  • Mean failure severity reduced: 1.58 → 1.16.
  • Mean usefulness increased: 2.60 → 3.10.
  • Large win example: nutrition-label task — reality check rejected false monopoly premise; proof review led to revising a false welfare claim.
  • Negative case: scaffolding in one task compressed an economically important mechanism too aggressively, producing an inferior result.
• Artifact: workflow and code available at https://github.com/maxwell2732/pAI-Econ-claude.

Data & Methods

• System construction:
  • Implemented as an orchestration core that routes runs through the staged pipeline according to four entry modes (Model Extension; Phenomenon-to-Model; Model Critique; Full Pipeline).
  • Agents are role-specific LLM invocations with separate prompts and read/write access to the shared workspace.
  • Nine gates (novelty risk, reality fit, canonical fit, theory lineage, model coherence, non-triviality, proof integrity, economic meaning, final math review).
  • Gate outputs: PASS (no targeted failure detected) or REFRAME (diagnosis, severity, recommended loopback, override permission).
• Evaluation protocol:
  • Five matched-evaluation tasks spanning diverse subfields; each run produced a staged workspace and a final manuscript skeleton.
  • For each task, two runs: gated workflow vs ungated baseline (same underlying model).
  • Two human evaluators blinded to which run used the gated workflow.
  • Evaluators scored manuscripts on failure modes and overall usefulness; pairwise A/B ranking recorded.
  • Process traces (stage outputs, gate reports, adjudication logs) were inspected to explain outcome differences.
• Coding of F5 (citations): bibliography usability and whether cited sources support the claims assigned to them were checked.
• Limitations of methods:
  • Small-n evaluation (five tasks) with two evaluators — informative but not large-scale.
  • Gates themselves are LLMs, so their diagnostics are imperfect and not certification.
  • Results depend on the targeting quality of gates and on human adjudication decisions.

Implications for AI Economics

• Practical design lessons:
  • For tasks lacking task-complete verifiers (most economic-theory work), architecture matters more than pure agent autonomy: bounded autonomy + staged inspectability + targeted gates is a better reliability strategy than leaving agents to self-organize.
  • Placing human checkpoints at irreversible decision points (e.g., equilibrium concept) concentrates scarce human attention where it yields the largest downstream value.
  • Persistent intermediate artifacts (blackboard files, gate reports, adjudication logs) are essential for auditability and traceable correction of subtle failures (canonical mismatch, hidden proof gaps).
• What this enables:
  • Safer deployment of LLMs for theory development, drafting, and exploratory model-building while preserving researcher control over substantive judgments.
  • Systematic exposure of common LLM failure modes so researchers can audit, loop back, and iteratively refine models and claims.
• Cautions and open questions:
  • The workflow improves auditability but cannot replace formal verification; it reduces but does not eliminate the risk of confident but incorrect claims.
  • Over-scaffolding risks: gates or scaffolds can bias or oversimplify mechanisms if checks are not well-targeted or human adjudication is mistaken.
  • Generalizability needs more testing: larger task suites, more evaluators, diverse LLMs, varied researcher expertise, and alternative gate designs.
• Research agenda:
  • Empirically compare different gate designs, human-checkpoint placement strategies, and autonomy budgets.
  • Quantify trade-offs between developer/researcher time costs and improvements in reliability/usefulness.
  • Extend to other social sciences where institutional fit and interpretive judgment matter.
• Policy/practice takeaway:
  • When adopting LLM agents for economic research, prioritize workflows that expose intermediate decisions, require human sign-off on irreversible choices, and log gate diagnoses rather than presenting opaque machine certification.