Andrii Bidochko

• Updated: August 13, 2026
• 7 min read

Evidence-Grounded Constraint Checking in Construction Documents

Direct Answer

The paper introduces an evidence‑grounded pipeline that automatically extracts, normalizes, and checks constraints in construction‑project PDFs, while preserving the original source spans for auditability. This matters because it moves document review from a manual, error‑prone task toward a deterministic, traceable AI‑assisted process that can scale across large, revision‑heavy construction contracts.

Background: Why This Problem Is Hard

Construction contracts, design drawings, and change‑order logs are typically stored as multi‑page PDFs that intertwine free‑form text, geometric annotations, and version histories. A single compliance decision often depends on:

• Cross‑referencing a clause on page 12 with a dimension on a drawing on page 45.
• Understanding how a revision note on page 3 supersedes an earlier specification.
• Interpreting spatial relationships (e.g., “the pipe must stay 300 mm away from the support beam”).

Traditional rule‑based systems excel when the evidence is cleanly structured (e.g., JSON or database rows) but stumble when facts are scattered across heterogeneous visual and textual layers. Human reviewers compensate by scrolling, zooming, and manually annotating, which is time‑consuming, costly, and inconsistent. Existing AI approaches—such as generic OCR pipelines or large‑language‑model (LLM) “question‑answer” bots—either lose the geometric context or provide answers without citing the exact PDF span, making them unsuitable for legal or safety‑critical decisions.

What the Researchers Propose

The authors present a four‑stage, evidence‑grounded framework that treats constraint checking as a deterministic rule‑execution problem:

1. Fact Extraction & Normalization: An OCR‑plus‑vision module parses text, tables, and geometric symbols, converting them into a unified fact graph (e.g., {entity: "Pipe‑A", attribute: "clearance", value: 300mm}).
2. Four‑State Rule Engine: Each compliance rule can be in one of four states—Pass, Fail, Unresolved, or Escalated. The engine evaluates rules deterministically using the normalized facts.
3. Evidence Allocation: For every rule decision, the system records the exact PDF spans (page numbers, bounding boxes) that contributed to the outcome, enabling traceability.
4. Escalation Layer: When the rule engine cannot reach a definitive Pass/Fail (e.g., missing geometry), the case is flagged for human expert review.

Key agents in the pipeline are the Extractor, the Normalizer, the Rule Executor, and the Escalation Manager. Their responsibilities are cleanly separated, allowing each to be swapped or upgraded without breaking the overall workflow.

How It Works in Practice

The end‑to‑end workflow can be visualized as a linear chain with feedback loops:

1. Document Ingestion: A PDF is uploaded to the system. The pipeline first creates a low‑resolution page overview (thumbnail) and then generates overlapping high‑resolution tiles for regions that contain dense annotations.
2. Evidence Allocation Strategy: The authors experiment with two allocation budgets:
   • Page‑Breadth Mode: One overview image per page, covering the entire document but with limited detail.
   • Region‑RAG Mode: One overview plus three overlapping tiles that focus on high‑information zones (e.g., drawing tables).
3. Fact Extraction: OCR extracts raw text; a vision model detects geometric symbols (lines, arrows, dimensions). The outputs are merged into a structured fact store.
4. Normalization: Ambiguous units (feet vs. meters), synonyms (“clearance” vs. “spacing”), and revision markers are harmonized into a canonical schema.
5. Rule Execution: The deterministic engine walks through each compliance rule, consulting the fact store. If all required facts are present, the rule resolves to Pass or Fail; otherwise it lands in Unresolved.
6. Evidence Retention: For every decision, the system logs the exact page numbers and bounding‑box coordinates that satisfied the rule, producing a machine‑readable audit trail.
7. Escalation: Unresolved cases trigger a notification to a human reviewer, who can inspect the retained evidence, add missing facts, or override the decision.

What distinguishes this pipeline from prior work is the explicit coupling of evidence allocation (how many image tiles to request) with rule‑state outcomes, and the guarantee that every automated decision is backed by a verifiable PDF span.

Evaluation & Results

The authors evaluated the pipeline on a real‑world dataset comprising 160 reference‑based tasks drawn from 29 construction projects. Each task required the system to answer a yes/no compliance question using the PDF evidence.

Experimental Design

• Repeated Four‑System Test: Four identical systems were run on the same task set, each with a different evidence‑allocation budget (one overview + three tiles vs. four overviews, etc.).
• Disjoint Two‑System Breadth Extension: Two systems with mutually exclusive evidence‑allocation strategies were compared to assess the impact of broader page coverage versus deeper region focus.

Key Findings

• Switching from a four‑overview budget to a mixed budget (one overview + three overlapping tiles) improved standardized decision accuracy by **10.6 percentage points** (95 % bootstrap CI: 4.3 – 18.0; exact p = 0.031).
• The same region‑focused strategy did **not** generalize when applied across a broader block of tasks; accuracy dropped by **4.1 points** (95 % CI: –10.2 to 1.9; p = 0.209).
• Overall, exact finding‑set recovery remained low, and false passes (incorrectly marked as compliant) were still common, indicating that the pipeline is better at flagging problems than at confirming correctness.
• Repeated runs showed poor calibration: the same document could receive different decisions across system instances, highlighting sensitivity to evidence allocation.

These results suggest a **resolution‑breadth trade‑off**: allocating more visual detail to a few regions can boost accuracy for certain tasks, but a broader page‑level view is safer when the evidence distribution is unknown.

Why This Matters for AI Systems and Agents

For AI practitioners building autonomous agents that interact with complex enterprise documents, the paper offers three actionable takeaways:

1. Rule‑Aware Evidence Routing: Instead of feeding an LLM a monolithic PDF, agents should first classify which rule(s) are relevant and then request evidence at the appropriate granularity. This reduces hallucination risk and improves decision confidence.
2. Auditability as a First‑Class Feature: By preserving source spans, downstream systems can generate compliance reports that satisfy legal and safety auditors without additional manual tracing.
3. Hybrid Human‑AI Loops: The escalation mechanism demonstrates a practical pattern for integrating expert review into fully automated pipelines, a design that can be reused for finance, healthcare, or any domain where false positives carry high cost.

Enterprises looking to embed such capabilities into their workflow can leverage existing UBOS tools. For example, the Workflow automation studio can orchestrate the evidence‑allocation, rule execution, and escalation steps as a single, repeatable process.

What Comes Next

While the pipeline marks a significant step forward, several limitations remain:

• Low Finding‑Set Recovery: The system often fails to retrieve the exact clause needed for a rule, suggesting that more sophisticated retrieval (e.g., semantic search over the fact graph) is required.
• Scalability of Tile Generation: Overlapping tiles increase computational load; adaptive tiling based on preliminary OCR confidence could mitigate this.
• Generalization Across Domains: The current evaluation is limited to construction documents. Extending the framework to other regulated industries (e.g., aerospace, pharmaceuticals) will test its robustness.

Future research directions include:

1. Integrating Chroma DB integration for vector‑based retrieval of normalized facts, enabling faster rule matching.
2. Exploring multimodal LLMs that can directly ingest image tiles alongside text, reducing the need for separate OCR pipelines.
3. Developing a feedback loop where human escalations automatically refine the rule base and improve the extractor’s precision over time.

Practitioners interested in prototyping these ideas can start with the Enterprise AI platform by UBOS, which offers pre‑built connectors for PDF processing, rule engines, and human‑in‑the‑loop workflows.

References

Evidence‑Grounded Constraint Checking in Construction Documents

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.