Interpretation and deployment

Error mechanisms, model choice, and responsible conclusions

The critical analysis questions every stage of the process and separates deterministic repeatability from robustness across different CV datasets.

Causal caution

Observed deviations have multiple competing explanations

A missing field can originate in source conversion, section or chunk construction, candidate generation, thresholding, duplicate resolution, schema mapping, normalization, validation, or evaluation. A supported final value can similarly originate from direct detection or later backfill. Diagnosis therefore requires stage-level lineage, not inspection of final JSON alone.

Figure 8. Stage-wise error accounting requirementsCode-derived failure mechanisms only. No stage-retention percentage or fitted loss coefficient is asserted.

HNLP deviations

Unrecognized headings, missing localized aliases, dictionary gaps, misspelling, OCR corruption, column flattening, overbroad ignored sections, and dominance thresholds.

BERT deviations

Model/tokenizer mismatch, language shift, chunk boundaries, label competition, low score, sanitizer rejection, overlap replacement, schema constraints, grouping, or backfill.

LLM deviations

Prompt interpretation, context dilution, quantization, decoding variability, reasoning mode, unsupported normalization, malformed serialization, and model-version drift.

Deployment decision

Speed, precision, implementation effort, and distributional stability

Figure 9. Deployment mechanisms and missing measurementsNo speed, repeatability, or cross-dataset stability ordering is established by the supplied benchmark.
CriterionHNLPBERT / GLiNERGenerative LLM
Latency mechanism (not measured)Rules and dictionariesChunked encoder inferenceAutoregressive generation and model serving
Repeatability controls (not ranked)Fixed code and resourcesFixed model, tokenizer, runtime, and mapperFixed decoding, server, prompt, and model version
Implementation effortLow initially in-domain; continuous rule and dictionary maintenanceModel packaging, labels, calibration, mapper, and runtimePrompting is quick; production controls remain substantial
Cross-dataset stability (not measured)Test sensitivity to headings, language, layout, and vocabularyTest model/domain, target span form, and label transferTest semantic adaptation against prompt/model drift
AuditabilityStrong rule-level traceabilitySpan and score traceability plus opaque representationsLower internal transparency; source-span evidence is essential

Controlled next study

Factorial benchmark required for causal claims

Use one PDF conversion, one schema, one blinded human-adjudicated atomic gold set, one matcher, and one missing-output policy. Randomize model and reasoning conditions; repeat stochastic runs; hash every artifact; retain HNLP line scores and state transitions; retain BERT raw logits, chunks, candidates, mapping, and backfill; and report latency, memory, energy, and cost on identical hardware.

Conclusion

A governed field-routed ensemble is the defensible engineering outcome

Use HNLP for high-throughput fields with stable headings and vocabulary; BERT/GLiNER for calibrated compact entities; an LLM for unusual structure and implicit relations; and human review for low-confidence, conflicting, or high-impact fields. Validate every accepted fact against the PDF. This routing policy makes the speed–precision–coverage trade-off explicit instead of hiding it behind a pooled score.

NextSources and reproducibility →
Figure