NCE · v2.4 · 12 / 14 active dimensions

The engine that measures.
Not the one that proclaims.

NCE (Neural Context Engine) is the quality engine embedded in KnowWeave. It measures the reliability of every answer across 14 dimensions — and refuses to answer when confidence is insufficient.

A measurable, auditable quality score, communicated without filtering — not an opaque probability.

/ NCE · architecture

Extract. Structure. Validate.
Three steps. Zero approximation.

⚙️ Guided extraction Sector domain schema · typed entities · semantic relations · per-extraction confidence
🕸️ Structured network Relation network · business conceptual model · deduplication · quality scoring
Validated answer Hybrid retrieval · entity whitelist · fidelity validation · complete source traceability

Architectural traceability: At every step, NCE traces what it extracts, why, and from which document. Traceability is architectural — not added after the fact.

/ reliability · NCE score

14 dimensions.
A score. Not an intuition.

The NCE reliability score aggregates 14 dimensions measured continuously on the knowledge network. Each NCE version is compared to the previous one on these same dimensions. These 14 dimensions are not a final state — they are a permanent R&D axis: each engine iteration is built to surpass the previous one, dimension by dimension, towards an ever-higher reliability score.

A — Domain model coherence
A1 Domain / scope alignment Does each extracted link correspond to a relation authorised within your business domain?
A2 Cardinality compliance Does the engine respect the cardinality rules of your domain — e.g. a substance can only have one regulatory identifier?
A3 Honest answers on the unknown Does the engine recognise what it does not know — and explicitly answer "information not available" rather than inventing?
B — Document grounding
B1 Semantic proximity Are similar concepts from your domain well grouped in the engine's semantic memory?
B2 Documentary support Is every extracted fact anchored to an identified source text extract?
C — Query stability
C1 Determinism Does the same question produce the same answer on the same graph?
C2 Alignment drift Do the domain model and knowledge network remain coherent when you add new documents?
C3 Repair rate What proportion of schema discrepancies is detected and corrected automatically?
D — Graph structure
D1 Relation coverage Are the link types defined in your domain model effectively populated in the knowledge network?
D2 Community coherence Do the graph's thematic clusters correspond to the expected business domains?
D3 Temporal validity Are dated items tagged with their validity period?
D4 Suggested enrichment Does the engine identify which graph areas should be completed as a priority?
E — Under active research R&D
E1 PRD · Path Reliance Degree Structural robustness of answer paths — identifies entities connected by a single path (fragile answers).
E2 SAS · Semantic Alignment Score Is every entity in the network positioned coherently with the other members of its category? An invisible misalignment can generate incorrect answers — difficult to detect without this score.
/ anti-hallucination · mechanisms

Three locks.
Zero invented answers.

Entity whitelist

Only entities present in the knowledge network can appear in an NCE answer. No interpolation. No extrapolation. If the entity is not in the base, it is not in the answer.

Out-of-domain refusal

If no source in the network is relevant to the question asked, NCE refuses to answer — without LLM call, without fabrication. The refusal message is explicit and customisable by business domain.

Fidelity validation

Every generated answer is submitted to an independent verifier that compares the answer to its cited sources. Binary verdict: FAITHFUL or CONTRADICTION. Production deployment in progress.

Layer 4 · deployment in progress
/ R&D · NCE v3.0

Two dimensions under research.
The next quality frontier.

PRD and SAS are the two reliability dimensions on which the NCE engine is under active research. Their full integration completes the score to 14/14 and unlocks agentic missions.

dimension E · 01

Answer path robustness

An entity connected to the rest of the graph by a single path is a fragile entity. If that path is removed or modified, the answer fails silently. PRD identifies and scores all single-path entities — allowing the agent to qualify the robustness of its own answer.

● Path Reliance Degree
dimension E · 02

OWL semantic alignment

An entity poorly positioned in the engine's semantic memory will be retrieved on the wrong questions and missed on the right ones. Result: answers that seem correct but are not — a silent cause of errors, difficult to diagnose without this specific score.

● Semantic Alignment Score · ref. academic research 2025
/ evaluation · protocol

How NCE is evaluated.
Every version. Every corpus.

Sector golden tests

Each sector has a set of reference questions with expected answers (ground truth). Standardised format — the same battery is replayed at each NCE version on the same corpus.

NCE vN vs vN−1

Every new version is benchmarked against the previous one. Measured KPIs: average score, pipeline confidence, out-of-domain refusal rate, contradiction detection rate. No figures published without this protocol.

Auditable protocols

All test sets are documented and available under NDA for clients in evaluation. The evaluation methodology is part of the delivery contract.

/ versions · NCE

NCE — Version History

NCE v1.0 delivered

Entity and relation extraction from text corpus. Generic relations. First knowledge graph queryable in natural language.

NCE v2.0 delivered

Sector domain schema. Extraction guided by conceptual model. Per-extraction confidence scoring. 12 active reliability dimensions.

NCE v2.4 current version

Entity deduplication. Noise node filtering. Semantic cache. Anti-hallucination whitelist. Fidelity validation implemented. Structural table parsing.

NCE v3.0 in development

Full reliability score on 14 dimensions. Autonomous cognitive agents capable of operating on continuously evolving living corpus. End-to-end auditable answers — every assertion traced to its source, ready for regulatory inspection or external audit. The platform that answers the question your teams have been asking for ten years.

/ next step

See NCE on your corpus.
30 minutes. Real diagnostic.

Bring a bounded corpus. We show how NCE ingests it, what it extracts, and how it answers your domain questions.

14 reliability dimensions Anti-hallucination Complete source traceability