LLM
Large language model. A model trained to read and generate text by learning patterns across many examples.
Example: A chat assistant that answers a question is usually powered by an LLM.
learn
AI glossary
plain definitions for technical words.
SIGNALWATCH uses terms from AI safety, evaluation, perception, and observability. This glossary explains them without assuming a machine learning background.
Simple first
Each definition starts with the practical meaning, not formal math.
Evidence-focused
Terms are explained in the way they appear inside SIGNALWATCH.
No hype
The glossary avoids magical claims and keeps uncertainty visible.
Model basics
Token
A small piece of text the model processes. It can be a word, part of a word, punctuation, or a code fragment.
Example: The word 'monitoring' may be split into several tokens.
Pretraining
The first large training stage where a model learns broad patterns by predicting text.
Example: This is where a model learns language structure and many general facts.
Fine-tuning
Additional training that changes how the model behaves for a narrower purpose or style.
Example: A model can be tuned to follow instructions more clearly.
RLHF
Reinforcement learning from human feedback. A method where human preferences help steer model responses.
Example: People compare answers, and that preference signal can shape future behavior.
Evaluation and safety
Benchmark
A test set used to compare model performance on specific tasks.
Example: A math benchmark checks math ability, not overall safety.
Evaluation
A structured test of how a system behaves under known conditions.
Example: SIGNALWATCH treats evaluation as evidence, not just a score.
Red teaming
Deliberately trying to make a system fail so weaknesses can be found before deployment.
Example: A team may test whether a model can be pushed into unsafe instructions.
Hallucination
When a model produces information that sounds plausible but is not grounded in truth.
Example: A model inventing a fake citation is hallucinating.
Distribution shift
When real-world inputs differ from the conditions used during training or testing.
Example: A camera model trained in daylight may fail in low light.
Perception and robustness
Detection
A model output saying it found an object or pattern in an input.
Example: COCO-SSD may report 'person' or 'car' with a bounding box.
Confidence
The model's reported score for an output. It is not a guarantee that the output is correct.
Example: A 70% confidence detection can still be wrong.
Robustness
How well a system continues working when conditions get harder or change.
Example: Blur, compression, motion, and occlusion test robustness.
Temporal consistency
Whether outputs stay stable across time instead of changing unexpectedly frame to frame.
Example: A detected object disappearing and reappearing can show instability.
Frame integrity
Whether recent inference frames produced usable detections or became empty.
Example: A sequence with many empty frames has weak frame integrity.
Evidence and operations
Provenance
The trace of where information came from and how it reached the interface.
Example: A source link, fetch time, and source name are provenance.
Telemetry
Operational signals about system behavior, such as latency, connection status, or collector health.
Example: Telemetry can show whether a source collector is delayed.
Evidence packet
A structured export containing timestamps, outputs, settings, and derived observations from a run.
Example: The perception lab exports evidence packets after real inference.
Observation window
The time span or frame span used to compute a measurement.
Example: Twelve inference frames over ten seconds form an observation window.
Unavailable state
A visible state showing that data or model output is missing instead of filling the gap with guesses.
Example: If the model emits no detection, SIGNALWATCH says so.
where to go next