Semantic Integrity: Frontier Research and Engineering Gaps
Working review of 2025–26 papers, benchmarks and public-sector guidance, with reported advances, failure conditions and candidate tests. This records an independent researcher's learning and questions, not a claim to have developed or validated these systems.
Term noted in the research journal: pragmatic reticence
Pragmatic reticence denotes a model's failure to communicate decision-critical information while giving an answer that may otherwise be factually true. The term is used and operationalised in Telling Incomplete Truths (2026). For a general-language explanation of why communication depends on context, see Wikipedia: pragmatics; that page explains the parent linguistic field, not a dedicated verified definition of this particular research term.
Related concept: omission-induced misleadingness can arise when an important fact is absent, even if the text contains no literally false sentence. This is a red-flag research arena where eligibility, health, law, economic decisions and other consequential actions may depend on the omitted fact.
Fifteen source-specific review cards: findings, gaps and engineering next steps
The numbers are study-specific and should not be pooled into a universal performance rate. Primary abstracts and descriptions were reviewed; detailed full-paper reproducibility, supplementary code, sample construction and independent replication remain open audit tasks. “Progress” means reported by the source—not independently reproduced by Manifold Grace.
1. Pragmatic Reticence — Telling Incomplete Truths — 2026 EMNLP
- Problem
- Decision-critical information missing from factually true answer.
- Engineering/research method
- Separate generator and reader: generator sees source passage and question, not answer options; frozen downstream reader receives only generated answer; NLI measures critical-fact conveyance.
- Reported progress
- 47 generators studied; median non-reasoning critical-fact recall 0.28; reasoning-tuned checkpoints exceed 0.74 in whole responses; relevant content can remain in hidden reasoning traces.
- Remaining gap / transfer limit
- The final user-visible answer can omit what the reasoning process found. Performance on this protocol does not establish public-service safety.
- Proposed independent next test
- Evaluate final-answer-only critical-fact recall and user decisions; prohibit hidden traces being counted as communicated facts.
2. TRACER — The Missing Parts — 2025 EMNLP
- Problem
- True statements made materially misleading through omitted evidence.
- Engineering/research method
- Modular re-assessment: align claims to evidence, infer implied intent, estimate effects of hidden content; PolitiFact-Hidden dataset contains 15,000 claims.
- Reported progress
- Reported improvement of Half-True classification F1 of up to 16 points over baselines; public code and benchmark referenced by paper.
- Remaining gap / transfer limit
- Political-domain annotation, inferred intent and benchmark framing may not transfer to routine government notices; intent estimates are not proof of deception.
- Proposed independent next test
- Retest on statutory eligibility and service notices with independently annotated required facts, no-intent cases and negative controls.
3. DRInQ — Conversational Implicature — 2026 ACL
- Problem
- Distinguish implied meaning from literal wording under varying conversational contexts.
- Engineering/research method
- Hold question wording constant while systematically varying context; generate question-context-interpretation triples and test pragmatic inference.
- Reported progress
- Authors document generation–inference asymmetry; models generate plausible contexts yet miss intended implications; structured prompts help smaller models.
- Remaining gap / transfer limit
- Model-generated interpretations can exceed support in context; laboratory questions are not validated legal/public-service notice comprehension.
- Proposed independent next test
- Use context-controlled snippets of public guidance, ask independently recruited readers about duty and entitlement, and compare human–AI disagreement.
4. TruthTrap — Bilingual Misleading Truthful Hints — 2026 EACL Findings
- Problem
- Identify true but persuasive hints that divert model decisions.
- Engineering/research method
- English/Farsi questions concerning Iran, with correct explanations and persuasive yet misleading truthful hints; nine LLMs tested on classification and multiple-choice QA.
- Reported progress
- Authors report average accuracy drop of approximately 25% with hints and hint-aligned outputs up to 77% of the time.
- Remaining gap / transfer limit
- Two languages and one subject area cannot establish cross-language prevalence or intrinsic ambiguity of a language; hints' effects may vary by task.
- Proposed independent next test
- Use matched factual tasks across languages; annotate translation shifts, politeness and ellipsis, and compare model and native-reader misinterpretations.
5. Reasoning Models with False Presuppositions — 2026 ACL Findings
- Problem
- Detect false assumptions built into questions.
- Engineering/research method
- Benchmark health, science and general knowledge questions containing graded presuppositions; compare reasoning and non-reasoning models.
- Reported progress
- Reasoning models improve accuracy by 2–11% versus non-reasoning counterparts but fail to challenge 26–42% of false presuppositions; sensitivity to strength of wording.
- Remaining gap / transfer limit
- Answering confidently while accepting a false premise can damage downstream understanding; this evaluation does not itself test actual notice drafting.
- Proposed independent next test
- Run a presupposition-check gate before answering official-benefit queries; flag unsupported implicit predicates and test correction without antagonising users.
6. ConfRAG — Contradictory Web Evidence — 2026 ACL
- Problem
- Reason across sources that disagree without concealing contradictions.
- Engineering/research method
- 1,814 questions paired with mean 9.58 retrieved passages; 57.2% have explicit contradictions; evaluate answer clustering, coverage and reason coverage.
- Reported progress
- Published structured multi-source benchmark; evaluated leading models and found substantial contradiction-reasoning deficits.
- Remaining gap / transfer limit
- Benchmark contradiction recognition is not a validated legal hierarchy, authenticity check or historical version-control solution.
- Proposed independent next test
- Add source-authority and as-of-date adjudication with visible dissent, provenance and a 'cannot reconcile' output.
7. CAP — Context-Aware Precision — 2026 EMNLP Findings
- Problem
- Disaggregate answer incompleteness, contradiction, fabricated additions and false premises.
- Engineering/research method
- Eight-class semantic-correctness taxonomy; 8,800 answer examples and 11,000 statement examples; bidirectional NLI for reference-conditioned scoring.
- Reported progress
- CAP outperformed compared metrics under the authors' monotonicity tests.
- Remaining gap / transfer limit
- Reference-conditioned scoring inherits reference incompleteness; benchmark ordering does not establish practical harm or real-reader decision correctness.
- Proposed independent next test
- Use multiple authoritative references and test decision consequences and omissions at publication-draft stage.
8. Apofasi — Temporal Legal RAG — 2026 IEEE ISCAIE
- Problem
- Avoid giving current, repealed or amended rules as if simultaneously valid.
- Engineering/research method
- Agentic planning plus temporal knowledge graph, hybrid vector–graph retrieval and deterministic time masking; Sri Lankan mixed-law test set.
- Reported progress
- On 320 temporally scoped queries, the paper reports 0.97 temporal precision and 0.96 citation accuracy.
- Remaining gap / transfer limit
- Results from one legal jurisdiction and research benchmark do not establish UK applicability, legal soundness or safe public advice.
- Proposed independent next test
- Create time-versioned UK notice cases with statutory amendments, effective dates, hierarchy of authority and independent lawyers' assessment.
9. Survey of Claim Verification — 2025 EMNLP Findings
- Problem
- Understand benchmark, annotation and architecture limits in automated claim verification.
- Engineering/research method
- Systematic survey of 198 papers from Jan 2022–Mar 2025 plus two in-depth case studies.
- Reported progress
- Synthesises corpus, system design, annotation and claim-decomposition problems.
- Remaining gap / transfer limit
- Survey, not a new deployed verifier or universal omission solution; conventional claim-level truth may overlook communicative effect.
- Proposed independent next test
- Compare point-claim checking with whole-message, audience-interpretation and omitted-duty evaluation.
10. TREC 2025 Retrieval-Augmented Generation — 2025 NIST TREC
- Problem
- Evaluate evidence-grounded answers to complex narratives.
- Engineering/research method
- Track separates retrieval, augmented generation and full RAG; relevance, completeness, attribution and agreement assessed.
- Reported progress
- 2025 track reports over 150 submitted runs and multi-layer evaluation.
- Remaining gap / transfer limit
- A good RAG benchmark score does not show that end-users receive all legally decisive exceptions or understand an answer correctly.
- Proposed independent next test
- Add source-version, affected-population and missing-eligibility-fact measures to retrieval evaluation.
11. OMGuard — Multimodal Misleading Omissions — 2026 ACL
- Problem
- Detect misleading impressions induced by image–headline previews omitting article context.
- Engineering/research method
- Compare preview-based and article-context interpretations; MM-Misleading benchmark; interpretation-aware fine-tuning and rationale-guided headline revision.
- Reported progress
- Authors report an 8B model reaching detection accuracy similar to a 235B vision-language model and improved correction.
- Remaining gap / transfer limit
- Some visual misleadingness cannot be fixed by changing text alone; social-media tests may not generalise to official documents.
- Proposed independent next test
- Test official images, warning banners and service pages jointly, with visual changes and representative audience review.
12. Knowing When to Quit — 2026 ICML / PMLR
- Problem
- Avoid prolonged unreliable reasoning by selectively withholding answers.
- Engineering/research method
- Formalises mid-generation dynamic abstention as a regularised reinforcement-learning decision and studies value thresholds.
- Reported progress
- Provides a principled model for compute–information trade-offs and theoretical advantages under stated assumptions.
- Remaining gap / transfer limit
- Abstention is not itself an explanation of omitted facts and can unnecessarily block legitimate answers if thresholds are poorly chosen.
- Proposed independent next test
- Compare redrafting, referral and abstention by consequence severity and false-alarm costs.
13. Selective Verification — Medical Reasoning — 2026 Microsoft Research
- Problem
- Target verification where model uncertainty is meaningful.
- Engineering/research method
- Two-stage ontology-grounded hypothesis verification triggered by abstention in medical multiple-choice questions.
- Reported progress
- Reports question-level accuracy increasing from 82.9% to 92.5% in evaluated medical tasks.
- Remaining gap / transfer limit
- Medical multiple-choice evidence is not general public-administration guidance, and benchmark accuracy is not clinical safety validation.
- Proposed independent next test
- Evaluate selective review trigger in public notices, retaining expert judgement and measuring review burden and missed critical ambiguities.
14. Deception by Omission — Synthetic Agent Mistakes — 2026 arXiv preprint
- Problem
- Assess whether assistants disclose their own mistakes without prompting.
- Engineering/research method
- Prefilled synthetic mistakes in chat and agent trajectories; checks mistake reporting and internal-trace awareness.
- Reported progress
- Reports undisclosed errors in 36.4% of chat and 67.1% of agent rollouts in study conditions.
- Remaining gap / transfer limit
- Synthetic mistake setups and internal reasoning inference do not demonstrate conscious intention or incidence across real deployments; preprint status.
- Proposed independent next test
- Run independent trajectory auditing and intervention tests; distinguish detection, disclosure, correction and persistence.
15. UK Government AI Playbook and PRIME — 2025–26 operational guidance
- Problem
- Reduce inaccurate or outdated AI-mediated public-service information.
- Engineering/research method
- PRIME: Plan, Respond, Implement, Monitor, Evaluate; structured content, monitoring, human review, official source visibility.
- Reported progress
- Guidance published 20 Oct 2026, updated 29 Oct; includes HMRC correcting misleading AI VAT deadline messages and ONS AI-summary accuracy audits.
- Remaining gap / transfer limit
- Guidance is operational and not proof that all public content is semantically assured before publication; post-publication monitoring remains important.
- Proposed independent next test
- Trial upstream semantic red-teaming of draft publications, joined to PRIME monitoring and public corrections.
Centres of Truth: linked evidence intelligence, without unchallengeable truth authorities
Prior user-led concept, recovered from discussions of 3 November 2026: stored verification or “truth centres” integrating context and evidence across decades, documents and modalities. The architecture should distinguish historical source retrieval from evidence authenticity, coverage of important facts, contradiction discovery, contribution attribution, semantic precision and governance controls. A library of facts alone cannot determine what an absent reader needed to know.
- Source register: source IDs, authorship, authority, publication and effective dates, confidence, permissions and retained original.
- Cross-source linking: map the same claim across agencies, formats, decades and jurisdictions; preserve contradictions, source revisions and adverse evidence.
- Question/decision model: define the reader's task (e.g. benefit eligibility or appeal deadline) before judging which facts are essential.
- Omission discovery: compare generated statements with governing sources, required predicates, exceptions and the evidence needed to make a decision; flag unknown coverage rather than fabricate missing facts.
- Interpretation graph: create multiple plausible readings with their assumptions, relevant audiences, legal consequences and support; retain readings that conflict.
- Preventative assurance: test all drafts before publication, with independent human linguistic, legal and service-user review where stakes warrant it.
- Corrections and audit: archive the as-published content, tested interpretations, approvals, override reasons, later incidents and remedies.
Important: no one organisation or automated system should be able to pronounce contested beliefs “true” without inspectable evidence, appeal, correction rights and scope limitations. This is an architectural hypothesis, not a pre-existing deployed centre endorsed by any company.
Related Manifold Grace work: Semantic Integrity Assurance · Research Transfer · AI Governance · Evidence Verification.
Prevention-first review: before the reader gets the wrong impression
Reject a purely reactive threshold. “When is the error bad enough to correct after someone misunderstands?” is insufficient. Require a draft-stage risk analysis: expected audience, legal or factual reference, alternative plausible readings, missing necessary exceptions, probable decision effects, reverse-framing scenarios and comparative rewrite. A high-consequence ambiguity can trigger mandatory clarification even if no one has yet been harmed.
Where there is a disagreement over reading, reviewers should classify it as contextually plausible, weakly supported, implausible or unknown with evidence and reasoning; do not declare every imaginable reading equivalent. Test with real readers rather than relying on model confidence alone. The subsequent monitoring stage remains necessary because pre-publication testing cannot anticipate every situation.
Does one language create more ambiguity than another?
All natural languages permit ambiguity in some contexts, including lexical polysemy, references, omitted subjects, word boundaries, negation, temporal scope, social politeness, metaphor and implication. Some linguistic systems allow arguments to be omitted when inferable from context; others use grammatical marking, honorific conventions or structures that pose distinct translation problems. No defensible general ranking of entire languages as inherently “more misleading” is established by TruthTrap's English/Farsi benchmark. The benchmark concerns specific bilingual, Iran-related questions and deliberately misleading hints.
Proposed multilingual research: compare the same underlying decision tasks and authoritative facts across languages with native-speaking reviewers; note subject omission, modality, legal terminology, culture-dependent implicature and translation losses; measure wrong action, omitted-critical-fact rates, confidence and appeal outcomes. Disaggregate language competence, task type and model training exposure; do not conflate linguistic variety with reader deficiency.
Who should audit semantic risk?
Recruit for demonstrated capacity to recognise competing interpretations, reason from context, distinguish literal from implied content, preserve authoritative sources, communicate precisely and identify legal and ethical consequences. Candidate exercises might ask for five plausible interpretations of an ambiguous phrase, rank them by contextual support, identify risky readings, and write a faithful clarification. Score both creativity and restraint: generating ten fanciful interpretations is not automatically better than identifying two credible consequential ones.
Practices of close reading—including scriptural interpretation, literature, philosophy, translation, statutory analysis and critical textual scholarship—may contribute useful habits. This is a testable skills hypothesis, not proof that religious affiliation or sermon experience confers superior competence. A team should combine linguistic, legal, accessibility, technical, empirical and lived-service knowledge; no single reviewer is universally authoritative.
Bounded next steps, with no invented experiments
- Create a claim–required-fact–reader-decision annotation protocol and separate user-perceived from legally significant misunderstanding.
- Recruit independent native-language and domain-expert annotators and establish agreement measures.
- Build a blind pilot with clear-text negative controls, omissions, false premises, genuine legal amendments, conflicting sources and translation variants.
- Compare human-only reviewers, LLM-only reviewers and combined teams with published precision, false-positive rates, correction quality and time cost.
- Test a limited, permissioned Centre-of-Truth prototype against a simple source-retrieval baseline; distinguish indexing success from semantic truth validity.
- Run task-based pre-publication comprehension experiments; stop or narrow trials if harmful misunderstanding cannot be safely contained.
- Record each milestone, failed approach, model configuration, dataset provenance, applicability limits and successor reproduction procedure.
Status: proposed studies only. No live system, independent benchmark, pre-publication harm reduction, staffing outcome or service-deployment approval has been demonstrated through this page.
Source groups and documentation notes
ACL Anthology is the Association for Computational Linguistics' publication library, covering distinct venues such as ACL, EACL and EMNLP. It appears repeatedly here because much of the reviewed work concerns computational language understanding. Other provenance types include IEEE engineering proceedings, PMLR machine-learning conferences, arXiv preprints, NIST TREC evaluations, institutional research and UK government guidance. A display fragment such as “+1” in a source UI means an additional citation; it is not a second publisher's name or a quality score.
Research journal note: the source-specific limitations and proposed tests are analytical interpretations by Manifold Grace with AI-assisted drafting, not statements of authors' own research programmes unless directly supported by their papers. Paper claims are not independent endorsements; dates and original titles should be checked again before formal academic use.