The Network Failed. The Fraud Model Saw Fraud.
A Nigerian POS fraud study shows how uncertainty-specific human review can improve recall and narrow a simulated rural accuracy gap without sending most transactions to analysts.
Cognaptus Insights
Industry thoughts, trends, and commentary.
The Training Rhythm Survives Encryption
FLINT shows how 5G scheduling metadata can expose federated-learning architecture families—and why strong closed-world accuracy is not enough for operational security.
A Multilingual Research Assistant Is Still an Infrastructure Project
ReSearch_SSH shows how specialised research platforms can combine domain adaptation, behavioural retrieval, knowledge graphs, expert evaluation, and legal controls without discarding their existing infrastructure.
Attention Is a Connection Walk, Not Automatically a Laplacian
A precise operator view shows why attention maps capture token routing but omit the feature transformations that determine what the layer actually computes.
The Probe Saw the Prompt Before It Saw the Fake
A hidden-state monitor can detect alignment-faking signals, but only after controls separate strategic compliance from prompt identity, query leakage, and configuration artifacts.
Let the Model Design the Poster—Not the Evidence
PosterHarness shows how to make scientific poster generation auditable by separating visual composition from evidence-bearing figures.
The Catalog Grew. The Agent Needed a Call Stack.
A hierarchical agent architecture sharply reduces tool-schema exposure at scale, but only when taxonomy, validation, and latency controls are engineered with equal care.
The Memory Score Changed Before the Memory Did
MemTools shows why agent-memory evaluation must separate component quality from interface compatibility, execution timing, and representation coordination.
Stored Is Not Reachable: Why Continual Fact Writing Breaks Under Later Updates
Controlled experiments show that a fact can remain statistically present in an LLM’s weights while becoming inaccessible, unusable, and vulnerable to later updates.
The Model Saw Every Scene. The System Had to Remember the Story.
StoryTeller shows why coherent long-form narration depends on verified external state, not only a stronger vision-language model or a longer prompt.