MediumContext Truncation◐Demonstrated
Context-window truncation & lost-in-the-middle
Model Accuracy & ReliabilityDescription
Long inputs exceed or are poorly attended within the context window, so key facts are dropped or de-weighted ('lost in the middle').
Example scenario
A long loan file is summarised but the disqualifying clause buried mid-document is omitted.
Real-world evidence◐Demonstrated
Error propagation in multi-stage NLP pipelines is a well-established phenomenon studied in machine translation, information extraction, and retrieval-augmented generation research, showing upstream errors compound through downstream stages. Specific production pipeline failures caused by AI error compounding (rather than single-model failure) have not been cleanly isolated and publicly documented as the root cause of a significant harm event.
Primary mitigations
- Chunking & re-ranking
- salience-aware retrieval
- context-length budgeting
- map-reduce summarisation
- position testing.
Detection signals
Long-document recall tests; needle-in-haystack evals; omission sampling.
Mitigating controls
9 Non-agentic controls
ZYC-ACCU-001Non-Agentic
Lexical Similarity
Accuracy
ZYC-XAI-001Non-Agentic
Field Justification
Explainability (XAI)
ZYC-ROB-001Non-Agentic
Out-of-Distribution Detection
Robustness
ZYC-ACCU-003Non-Agentic
Lexical Similarity
Accuracy
ZYC-ACCU-013Non-Agentic
Retrieval Quality
Accuracy
ZYC-ACCU-014Non-Agentic
Context Efficiency
Accuracy
ZYC-ACCU-015Non-Agentic
Accuracy
Accuracy
ZYC-ROB-008Non-Agentic
Structural Awareness
Robustness
ZYC-ROB-009Non-Agentic
Decision Integrity
Robustness