0 cumulative citations
View corpus contextA deterministic GPT assistant cuts the drudgery from regulatory cost verification by reliably checking codes, units and prices across messy spreadsheets. Tested on 20 real estimates (16,100 rows) and controlled error cases, it preserves full auditability while remaining limited to single-province price rules and strict exact-match logic.
Citation observations
Cumulative provider counts captured on specific dates; providers are never combined.
1 cumulative citations
View corpus contextCost code verification in state-funded construction projects remains a labor-intensive and error-prone task, particularly given the structural heterogeneity of project estimates and the prevalence of malformed codes, inconsistent units of measurement (UoMs), and locally modified price components. This study evaluates a deterministic GPT-based assistant designed to automate Vietnam’s regulatory verification. The assistant was developed and iteratively refined across four Action Research cycles. Also, the system enforces strict rule sequencing and dataset grounding via Python-governed computations. Rather than relying on probabilistic or semantic reasoning, the system performs strictly deterministic checks on code validity, UoM alignment, and unit price conformity in material (MTR), labor (LBR), and machinery (MCR), given the provincial unit price books (UPBs). Deterministic equality is evaluated either on raw numerical values or on values transformed through explicitly declared, rule-governed operations, preserving auditability without introducing tolerance-based or inferential reasoning. A dedicated exact-match mechanism, which is activated only when a code is invalid, enables the recovery of typographical errors only when a project item’s full price vector well matches a normative entry. Using twenty real construction estimates (16,100 rows) and twelve controlled error-injection cases, the study demonstrates that the assistant executes verification steps with high reliability across diverse spreadsheet structures, avoiding ambiguity and maintaining full auditability. Deterministic extraction and normalization routines facilitate robust handling of displaced headers, merged cells, and non-standard labeling, while structured reporting provides line-by-line traceability aligned with professional verification workflows. Practitioner feedback confirms that the system reduces manual tracing effort, improves evaluation consistency, and supports documentation compliance with human judgment. This research contributes a framework for large language model (LLM)-orchestrated verification, demonstrating how Action Research can align AI tools with domain expectations. Furthermore, it establishes a methodology for deploying LLMs in safety-critical and regulation-driven environments. Limitations—including narrow diagnostic scope, unlisted quotation exclusion, single-province UPB compliance, and sensitivity to extreme spreadsheet irregularities—define directions for future deterministic extensions. Overall, the findings illustrate how tightly constrained LLM configurations can augment, rather than replace, professional cost verification practices in public-sector construction.
Summary
Main Finding
A deterministic, GPT-orchestrated assistant—tightly constrained by rule sequencing and Python-governed computations—can reliably automate much of Vietnam’s regulatory cost-code verification for state-funded construction projects. By performing only deterministic checks (exact or rule-governed equality) against provincial unit price books (UPBs), the assistant preserves full auditability, reduces manual tracing effort, and improves evaluation consistency without relying on probabilistic or inferential reasoning.
Key Points
- System design
- LLM used as an orchestrator within four Action Research cycles for iterative refinement.
- Strict rule sequencing and dataset grounding: Python executes deterministic computations; the LLM controls flow and reporting but not fuzzy inference.
- Deterministic checks include: code validity, unit-of-measure (UoM) alignment, and unit-price conformity for material (MTR), labor (LBR), and machinery (MCR).
- Deterministic equality assessed either on raw numbers or on values transformed by explicitly declared, rule-governed operations—no tolerance bands or inferential matching.
- Error recovery
- Exact-match mechanism activated only when a code is invalid; it recovers typographical errors only if a project item’s full price vector closely and exactly matches a normative UPB entry.
- Robust extraction & normalization
- Deterministic routines handle structural spreadsheet heterogeneity: displaced headers, merged cells, non-standard labels, and locally modified price components.
- Reporting & auditability
- Structured, line-by-line reports align with professional verification workflows and provide full traceability for auditing and human review.
- Evaluation & outcomes
- Tested on 20 real construction estimates (16,100 rows) and 12 controlled error-injection cases.
- Demonstrated high reliability across diverse spreadsheet structures and error types while avoiding ambiguity.
- Practitioner feedback: reduced manual tracing, improved consistency, supports documentation and human judgment.
- Limitations
- Narrow diagnostic scope (focused checks, not holistic estimate validation).
- Excludes unlisted quotations (items not in UPB).
- Compliance tested against a single-province UPB.
- Sensitive to extreme spreadsheet irregularities beyond the normalization heuristics.
Data & Methods
- Data
- Twenty real project estimates totaling ~16,100 spreadsheet rows.
- Twelve controlled error-injection scenarios to probe failure modes and recovery behavior.
- Provincial Unit Price Books (UPBs) as the normative dataset for verification.
- Methods
- Action Research methodology with four iterative development cycles involving practitioners.
- LLM (GPT-based) orchestrates verification steps; deterministic computations and data lookups executed in Python to enforce reproducibility.
- Extraction pipeline: deterministic parsing and normalization routines to convert heterogeneous spreadsheets into canonical item vectors (code, UoM, MTR/LBR/MCR prices, quantities).
- Verification logic: rule-governed functions check code validity, UoM consistency, and unit-price conformity; exact-match recovery is strictly conditional on full vector alignment.
- Output: structured, auditable reports with per-line decision logic and traceable computation steps.
Implications for AI Economics
- Operational efficiency and public procurement
- Automating deterministic verification reduces auditor labor and time-to-approval, potentially lowering administrative costs in public construction procurement.
- Improved consistency and documented traceability can accelerate compliance checks and reduce disputes over routine coding/price errors.
- Measurement quality and economic analysis
- Reducing measurement error in construction cost data strengthens the reliability of cost indices, budget forecasting, and empirical research on public investment productivity.
- Deterministic, auditable adjustments make downstream economic estimates (e.g., cost overruns, unit-cost benchmarking) more defensible.
- Governance, transparency, and corruption risk
- Line-by-line traceability and rule-based checks increase transparency in price-setting and validation, which can reduce opportunities for opportunistic misreporting—but only for items covered by the UPB.
- Adoption & scaling considerations
- Benefits depend on availability and maintenance of high-quality normative datasets (UPBs) and standardized digital inputs; scaling across provinces or countries requires harmonizing or mapping local UPBs.
- The deterministic approach is well suited to regulation-driven contexts where auditability and predictability are prioritized over recall/recall-flexibility.
- Labor market and organizational effects
- Likely to shift verifier roles from manual tracing to exception handling, interpretation, and policy decisions—implying upskilling rather than wholesale replacement.
- Research directions for AI economics
- Quantify cost/time savings and impacts on approval throughput at scale.
- Extend deterministic frameworks to multi-province UPBs and to formally handle unlisted quotations while preserving auditability.
- Evaluate hybrid designs that combine deterministic guarantees with constrained probabilistic matching (with auditable confidence metrics) for broader coverage.
- Study how tighter digital verification affects bidding behavior, contractor pricing strategies, and market competition in public construction.
Assessment
Claims (11)
| Claim | Direction | Outcome | Confidence & Evidence | Details |
|---|---|---|---|---|
| The assistant executes verification steps with high reliability across diverse spreadsheet structures, avoiding ambiguity and maintaining full auditability. Output Quality | positive | verification reliability / correctness of automated checks |
Reading fidelity
high
Study strength
medium
|
n=16100
|
| Deterministic extraction and normalization routines facilitate robust handling of displaced headers, merged cells, and non-standard labeling. Organizational Efficiency | positive | ability to parse and normalize heterogeneous spreadsheet structures |
Reading fidelity
high
Study strength
medium
|
n=20
|
| A dedicated exact-match mechanism recovers typographical errors only when a project item’s full price vector well matches a normative entry. Error Rate | positive | ability to auto-correct typographical errors in codes/entries |
Reading fidelity
high
Study strength
medium
|
n=12
|
| The system performs strictly deterministic checks on code validity, unit-of-measure (UoM) alignment, and unit price conformity for material (MTR), labor (LBR), and machinery (MCR) against provincial unit price books (UPBs). Regulatory Compliance | positive | conformance to regulatory pricing and code rules |
Reading fidelity
high
Study strength
medium
|
not reported
|
| Deterministic equality is evaluated either on raw numerical values or on values transformed through explicitly declared, rule-governed operations, preserving auditability without introducing tolerance-based or inferential reasoning. Ai Safety And Ethics | positive | auditability and determinism of verification logic |
Reading fidelity
high
Study strength
medium
|
not reported
|
| Structured reporting provides line-by-line traceability aligned with professional verification workflows. Regulatory Compliance | positive | traceability / audit trail quality |
Reading fidelity
high
Study strength
medium
|
not reported
|
| Practitioner feedback confirms that the system reduces manual tracing effort, improves evaluation consistency, and supports documentation compliance with human judgment. Developer Productivity | positive | manual tracing effort and evaluation consistency |
Reading fidelity
high
Study strength
low
|
not reported
|
| The assistant was developed and iteratively refined across four Action Research cycles. Other | neutral | development process (iterative refinement) |
Reading fidelity
high
Study strength
high
|
not reported
|
| Limitations include narrow diagnostic scope, exclusion of unlisted quotations, compliance only with a single-province UPB, and sensitivity to extreme spreadsheet irregularities. Other | negative | system limitations and failure modes |
Reading fidelity
high
Study strength
low
|
not reported
|
| The study demonstrates a framework for LLM-orchestrated verification and establishes a methodology for deploying LLMs in safety-critical and regulation-driven environments. Governance And Regulation | positive | methodological contribution to LLM deployment practices |
Reading fidelity
medium
Study strength
speculative
|
not reported
|
| Tightly constrained LLM configurations can augment, rather than replace, professional cost verification practices in public-sector construction. Task Allocation | positive | role of LLMs in professional verification (augmentation vs. replacement) |
Reading fidelity
high
Study strength
medium
|
n=20
|