skip to content
← featured work
Active research validation Research · NLP / retrieval Phase 1

NormGuard

A proposed domain-aware quality gate for text normalisation in Search, NLP and RAG systems.

GitHub repository ↗
OVERVIEW

What it is

NormGuard is investigating a cross-industry evidence layer around existing normalisation algorithms: compare candidate pipelines, measure retrieval utility and terminology safety, and recommend only when declared thresholds pass.

ℹSTATUS BOUNDARY
Proposed capabilities and measurement plans are separated from verified implementation.
PROBLEM

Why it exists

Stemming and lemmatisation are easy to enable but their downstream effects are rarely measured. A transformation may improve recall, do nothing for dense retrieval, or corrupt an exact identifier or domain term.

ARCHITECTURE

System flow

Dataset queries + relevance Candidates none · stem · lemma Matched runs retrieval modes Measurement utility + safety Decision pass · review · block
Figure 1 · Proposed matched-evaluation loop for normalisation utility, safety and operational cost.
CAPABILITIES

What is implemented

▸ Research charter with hypotheses and exit criteria
▸ Landscape review of closest tools and overlap
▸ Evaluation plan for retrieval, safety and efficiency
▸ Proposed no-normalisation, stemming and lemma baselines
▸ Cross-industry protected-term examples
▸ Architecture and trust-boundary documentation
▸ Visual evidence standards for uncertainty and traceability
▸ Evidence-focused contribution workflow
TECHNOLOGY STACK
Python research planNLPLexical retrievalDense retrievalHybrid searchRAG evaluationStatistical analysisCI quality gates
ENGINEERING

Key decisions

ADR-01

Quality layer, not another stemmer

accepted

The contribution evaluates existing transformations rather than claiming a new algorithm.

ADR-02

Safety beside utility

accepted

Protected-term violations are measured alongside retrieval metrics.

ADR-03

Inconclusive is valid

accepted

Uncertain evidence should trigger review rather than a forced recommendation.

TRUST

Security, privacy and quality controls

  • ◆Freeze benchmark and statistics before implementation claims
  • ◆Assess dataset licensing and provenance
  • ◆Treat protected terminology as a safety constraint
  • ◆Require labelled, traceable charts with uncertainty
  • ◆Measure token savings only when meaningful
EVIDENCE

Verified results

  • ✓Research charter, landscape, evaluation plan, architecture and visual system documented
  • ✓Repository supports reviewed, evidence-first collaboration
  • ✓No retrieval improvement, cost saving or novel algorithm is currently claimed
CURRENT LIMITATIONS

What this does not claim

▲ACCURACY BOUNDARY
  • • NormGuard is not yet a released library.
  • • The proposed pipeline and CI gate are not completed features.
  • • No statistically supported retrieval improvement exists yet.
NEXT

Roadmap

  • →Complete the sourced literature evidence matrix
  • →Select licensed datasets and freeze the benchmark
  • →Define safety taxonomy and statistical thresholds
  • →Make a documented go, narrow, reposition or stop decision