Regime-Aware Context Validity for Trustworthy Cognitive AI

Semantic windows, admissible memory and safety-governed adaptation for Deep Reasoning, Abstraction and Planning under regime change.

Target call	EIC Pathfinder Challenge DeepRAP · HORIZON-EIC-2026-PATHFINDERCHALLENGES-01-03.
Deadline / instrument	28 October 2026, 17:00 Brussels local time; EIC Pathfinder Challenge, Research and Innovation Action; grants up to EUR 4M considered appropriate, 100% eligible-cost funding, according to the 2026 Work Programme.
Role sought	Bounded scientific work package in a strong cognitive-AI / DeepRAP consortium; not a standalone product proposal.
Maturity	TRL 1–4 scientific concept; proof-of-principle / validation protocol; field-proven operational lineage but no empirical claim for the generalized method.

Core proposition

Trustworthy cognitive AI requires more than stronger reasoning, abstraction and planning modules. It requires a way to determine whether the context used by those modules is still valid. Under regime change, a system may reason correctly over the wrong history.

We propose context validity as a cognitive primitive: a regime-aware layer that determines which parts of past experience remain admissible for the current task, which parts are obsolete or contaminating, and when the system should adapt, abstain, escalate or re-plan.

Why this fits DeepRAP

DeepRAP targets Deep Reasoning, Abstraction and Planning towards trustworthy Cognitive AI Systems. Regime-aware context validity addresses a prior condition for all three capabilities: before an AI system reasons, abstracts or plans, it must know whether the context grounding that operation still belongs to the current regime.

DeepRAP capability	Context-validity contribution
Reasoning	Prevent correct inference over contaminated history; expose why a context was accepted, rejected, down-weighted or abstained.
Abstraction	Separate stable, transferable context from obsolete context; constrain abstraction to regime-compatible experience.
Planning	Trigger re-planning, fallback or human escalation when plan assumptions come from invalid context.

Pathfinder-native breakthrough thesis	Implication
Missing primitive	Current cognitive AI architectures typically improve models or planners, but rarely formalise whether the context feeding those components remains valid after a regime shift.
Scientific leap	Move from detecting distribution change to reasoning about admissible memory: which past experience remains valid for the present cognitive task.
Expected proof-of-principle	A demonstrator in which the same host reasoner/planner behaves differently because the context-validity layer filters, rejects, down-weights or abstains from contaminated memory.

Tegrity.AI is the research and publication circle of The Integral Management Society, a Swiss non-profit association in Geneva. OÜ JUBAP / JubAp.EU is the European innovation and contracting entity; JubAp.Net is the original engineering lineage behind several historical operational-intelligence systems now being reviewed or formalised.

The ecosystem brings mission-critical operational intelligence, systems orchestration, process intelligence and enterprise architecture experience, including Swiss and Swiss-headquartered engagements with SGS, Nestlé and Richemont. These references are included as field provenance only; they are not presented as endorsement, certification or validation of the Pathfinder claim.

The applied history matters because DeepRAP demonstrators need realistic failure modes: cascading constraints, stale planning assumptions, brittle handovers, hidden dependencies, context drift, human escalation and safety-governed adaptation. Operational lineage informs demonstrator realism; it does not prove the generalized semantic-window formulation.

Lineage / environment	Capability evidence	Pattern brought into DeepRAP
Nokia R&D lineage	Advanced digital and telecommunications systems practice.	Large-scale engineering discipline and frontier systems thinking.
PEMEX / GEPLAN	Mission-critical logistics intelligence and enterprise-system architecture.	Planning-execution loops, propagation control and field-operational constraints.
AICM / airport orchestration	Gate-slot / resource orchestration in live critical infrastructure.	Constraint-driven planning under disruption and handover pressure.
xSeil / Experiencias Xcaret	Passenger transport orchestration with fixed commitments, hard capacity and continuous disruption.	Dynamic reallocation, context validity, cascading-effect control and explainable operational decisions.
Phylons lineage	Explainable, influence-structured operational and signal systems.	Regime-sensitive confidence, selective action, abstention and controlled adaptation.
Richemont / Swiss luxury-manufacturing context	Application portfolio, process-intelligence and transformation-governance work in demanding Swiss-linked environments.	Process telemetry, governance drift and high-value operational context control.

Scientific object

A semantic window is a selected historical context defined not by a fixed lookback period, but by the subset, boundary or weighting of past experience that remains semantically, statistically or operationally compatible with the current task and regime.

The central mechanism is boundary contamination: a chosen historical boundary can inject incompatible history into an estimator, planner, world model or reasoning trace. A boundary is useful only if it improves the validity, safety or recovery behaviour of the downstream cognitive operation, not merely because it marks a distributional change.

The updated claim package separates the programme into three validation objects. The ranges below are protocol-sizing hypotheses used to design evaluation sensitivity and sample requirements. They are not expected results, guarantees, sales claims or commitments. Conservative ranges are the only figures suitable for external quotation before validation.

Hypothesis	What is tested	Primary discipline	Protocol-sizing ranges
H1 · Safety Governor / context validity	Whether a causal safety governor reduces decisions authorised after historical context has ceased to be admissible, without excessive false re-anchoring.	Headline: WICE at matched FRR under a central action-cost schedule. Diagnostics: TTSP, ICB, tail loss, base half-life, false re-anchor cost and bounded-hindsight capture ratio.	Conservative: 10–30% invalid-context exposure reduction; 10–25% false re-anchor reduction; 5–20% tail-loss reduction. Base target: 30–70% WICE reduction; 25–60% false re-anchor reduction; 20–50% tail-loss reduction, when independent invalidation events exist.
H2 · Prediction support	Whether cleaner context anchors reduce forecast or estimation loss versus strong adaptive, drift-aware, regime-switching and modern ML baselines.	Oracle ceiling is diagnostic; operative claim is live, past-only, frozen-split and leakage-audited performance versus strong baselines.	Oracle prediction ceiling: 0–5% low/no regime mixing; 5–20% base; 20–35% high; 35–45% rare/oracle-favourable. Net live gains should be materially lower after base-discovery and error costs.
H3 · Compute efficiency	Whether regime-compatible context reduces avoidable processing of stale or incompatible history while quality is held fixed.	Compute at equal quality: CPU/GPU seconds, kWh, cloud cost, latency, memory, evaluated windows/jobs; net of base-discovery, maintenance and error costs.	Oracle running-cost ceiling: 0–10% weak/low-mixing; 10–20% conservative viable; 20–35% base; 35–45% high-case. Live result must capture a material share of the ceiling after base costs.

Claim boundary and adjacent families

What is claimed	A formalizable and testable context-validity layer for history-dependent cognitive tasks under regime change.
What is not claimed	Not AGI, not a universal predictor, not a replacement model, not a production-ready product and not a guarantee that all host models improve.
Evidence role	Operational lineage motivates realistic failure modes and demonstrator design; it does not prove the generalized method.
Validation rule	Every out-of-sample claim must be past-only, leakage-audited, frozen-split, compared against strong native baselines and charged for action costs.

Adjacent area	Difference to test
Change-point / drift detection	Typically asks where a distribution changes; context validity asks whether a boundary produces admissible memory for downstream cognitive operation.
Adaptive windows	Often optimise statistical or predictive criteria; semantic windows focus on regime compatibility, boundary contamination and task validity.
Adaptive conformal / selective prediction	Close comparators for coverage and abstention; they do not by themselves validate the historical context used by the host model.
World models / MLOps monitoring	World models encode experience and monitoring observes degradation; this layer decides which experience remains admissible before reasoning or planning uses it.

Regime-Aware Context Validity Layer. The preferred role is a bounded scientific and technical work package inside a strong DeepRAP consortium led by partners in cognitive AI, neuro-symbolic reasoning, planning, uncertainty, world models, trustworthy AI or human-AI teaming.

Work package	Scientific / technical purpose	Outputs
WP1 · Formal primitives	Define context validity, regime compatibility, boundary contamination and admissible memory for cognitive tasks.	Definitions; claim register; assumptions; negative cases.
WP2 · Oracle benchmark	Use ex post semantic windows deliberately to test whether correct context boundaries matter for downstream RAP tasks.	Oracle benchmark; contamination tests; transition-resilience endpoints.
WP3 · Past-only inference	Develop online approximations using only information available up to time t, with leakage audits and error margins.	Past-only selector; uncertainty calibration; failure-mode taxonomy.
WP4 · RAP architecture integration	Connect context validity to reasoning, abstraction, planning, world models or neuro-symbolic components, using Phylons as explainable-architecture lineage.	Interfaces; host-model contracts; accepted/rejected context explanations.
WP5 · Safety-governed adaptation	Define confidence, abstention, escalation, re-planning and bounded response when context validity degrades.	Safety Governor states; human-in-the-loop triggers; recovery metrics.
WP6 · Portfolio contribution	Contribute to shared benchmarks, interoperability, evaluation protocols and DeepRAP portfolio activities.	Dedicated portfolio contribution; benchmark and protocol package.

Pathfinder dimension	Programme relevance
High risk	Ex ante context validity may be too error-prone, too domain-specific or too costly to integrate into RAP architectures.
Scientific gain	A successful proof-of-principle would add admissible memory/context management as a new primitive for trustworthy cognitive AI.
Technology gain	Host cognitive systems could reason, plan, abstain and re-plan over validity-aware context rather than undifferentiated historical memory.

Evaluation logic

Use the same host reasoner, planner or model where possible; change the context interface, not the whole cognitive stack. Compare fixed windows, adaptive windows, change-point baselines, concept-drift methods, adaptive conformal/selective baselines and semantic context selection under each method’s native online policy.

• Primary safety endpoints: WICE at matched FRR, invalid-context exposure, TTSP, ICB, tail loss and action-cost-adjusted safety value.
• Prediction endpoints: transition peak loss, cumulative excess loss, recovery time, forecast loss, calibration and robustness to contaminated context.
• Compute endpoints: CPU/GPU seconds, kWh, cloud cost, evaluated windows/jobs, online latency and memory at equal quality.
• Cognitive endpoints: explanation quality, validity of abstractions, plan repair speed, multi-agent coordination stability and robustness to contaminated memory.
• Negative and inconclusive results are valid outcomes if they identify limits, circularity, excessive assumptions or non-transferability.

Demonstrator A — Passenger transport orchestration under regime change

A passenger-transport planning environment — many vehicles, hubs, service commitments and passengers under hard capacity — is exposed to regime changes: capacity loss, demand shock, traffic disruption, no-show surges, policy shifts or sensor unreliability. This is the operational world behind xSeil, where centralized coordination was designed for up to roughly 15,000 interacting operational entities under continuous disruption.

Cognitive task	Passenger-flow planning and multi-vehicle coordination under shifting constraints.
Novel test	Can context validity prevent stale operating assumptions from propagating through passenger plans and transfers?
Grounded lineage	Realistic failure modes drawn from field-proven passenger orchestration; not proof of the generalized method.
Measures	Plan repair time, cascading-disruption reduction, coordination stability, explanation of context rejection.

Demonstrator B — World-model adaptation with contaminated memory

A cognitive system maintains or queries a world model built from historical observations. When the regime changes, parts of memory become obsolete. The demonstrator tests whether semantic windows help the system separate admissible context from contaminated memory before reasoning or abstraction proceeds.

Cognitive task	Formal validity, reasoning and abstraction over a history-derived world model.
Novel test	Can the system identify which memory remains valid for the current regime?
Measures	Formal validity accuracy under shifts, abstraction validity, abstention usefulness, recovery after regime break.

Consortium profile sought

Role	Contribution
Coordinator	Top group in cognitive AI, neuro-symbolic reasoning, planning, world models or trustworthy AI.
RAP core partner	Owns host reasoning/planning architecture and evaluates context-validity integration.
Statistical learning partner	Supports online inference, baselines, uncertainty and leakage-audited evaluation.
Human-centred AI partner	Covers explainability, oversight, escalation, trustworthiness and AI Act alignment.
Application / scenario partner	Provides a cognitively rich task environment; not merely a data source.
Tegrity.AI / IMSV / OÜ JUBAP	Contributes context-validity theory, semantic-window object, Safety Governor logic, Phylons explainable-architecture lineage, field-derived failure modes and demonstrator design.

The Pathfinder work package can be prepared in parallel with shorter Swiss and EU validation tracks. The purpose is not to prove all claims before Pathfinder, but to reduce conceptual risk and produce a disciplined claim register before joining or co-writing a consortium work package.

Track	Object	Boundary
Semantic Windows / H1–H2–H3	Formal and empirical plausibility of context-validity windows, Safety Governor, prediction-support and compute-efficiency claims.	Does not certify product readiness or universal model improvement.
Structural Awareness Foundations	Formalization loss, representation maps, context validity and the bridge from world-to-record limits to organisational measurement.	Does not replace the Semantic Windows validation; supplies a foundation and bridge.
QAVA / Universitat de València	Historical xSeil/Phylons-derived computational structures from quantum, quantum-inspired or hybrid perspectives.	Informs computational strategy; does not validate the current semantic-window mathematics.
TalTech / Estonia RUP	Practical exploration of information failure, Cost of Clarity and cascade risk in enterprise architecture and transformation contexts.	Tests organisational applicability; does not prove the formal mathematics.

What an independent validation partner reviews

Mathematical object	Semantic windows as context-validity objects: regime-compatible historical support, admissibility weights, boundary contamination, oracle-versus-live bases and safety-governed action.
Type of reasoning	Finite, time-indexed, causal/past-only definitions; boundary-contamination identities; admissibility weights; stability/false-reanchoring logic; leakage control; action-cost and abstention discipline.
Validation-track time	Preliminary desk assessment: 4–8 weeks. Theorem-level validation and reference implementation: 6–12 months. Empirical protocol: separate, longer track.
Articles / notes to validate	Semantic Windows note; H1 Safety Governor; H2 Prediction Improvement; H3 Compute Efficiency; Paper A1/A2/B as controlled package; operational provenance as context only.

Guideline	Language
Do say	Context validity is a missing primitive for RAP systems under regime change.
Do say	We seek a bounded scientific work package in a strong DeepRAP consortium.
Do say	The goal is proof-of-principle, benchmarks, formalisation and cognitive demonstrators.
Do say	We bring field provenance in mission-critical systems, but not product-readiness claims.
Do not say	We have a product, a full EWS, a universal forecasting improvement or an AGI component.
Do not say	This is just window selection; it is about admissible context for cognitive operations.

Six-week action plan

Timing	Action
Week 1	Finalize DeepRAP v3.0; prepare short outreach email, PDF package and 30-minute technical scoping agenda.
Week 1–2	Map 20–30 coordinators in neuro-symbolic AI, planning, world models, trustworthy AI and human-AI teaming.
Week 2–3	Contact only strong coordinators or consortia already moving; position as bounded WP contributor.
Week 3–4	Qualify fit: DeepRAP vision, consortium strength, WP gap, demonstrator, budget and ownership boundary.
Week 4–6	Co-write one WP plus evaluation language; align portfolio activity, benchmarks and demonstrator role.
Before submission	Keep claims TRL 1–4, proof-of-principle, high-risk/high-gain and avoid operational/product overreach.

Contact and official sources

Contact: Tegrity.AI / IMSV / OÜ JUBAP (JubAp.EU) — info@jubap.eu · tegrity.ai · jubap.net · jubap.eu

Official call anchors: EIC Pathfinder Challenges 2026 page; EIC Pathfinder general page; DeepRAP Challenge Guide; EIC Work Programme 2026.

• EIC Pathfinder Challenges 2026
• EIC Pathfinder (general)
• DeepRAP Challenge Guide (PDF)
• EIC Work Programme 2026 (PDF)
• Minimalistic Regime-Aware Early Warning Systems
• Evolution of Regime-Awareness Capability
• Structural Awareness Program
• Systems Orchestration & Regime-Awareness Capabilities
• JubAp.Net — xSeil whitepaper
• JubAp.Net — case studies
• xSeil / Experiencias Xcaret case study
• GEPLAN / PEMEX case study (2006)
• Mexico before Industry 4.0 — GEPLAN case

JubAp.EU · Strategic Intelligence & Transformation · OÜ JUBAP (Tallinn) · Tegrity.AI / IMSV (Geneva)
Pathfinder-native consortium discussion brief · Field provenance only; claims fenced as TRL 1–4 proof-of-principle.