Concept Glossary: P¶

This page lists English glossary entries for this letter. Entry bodies are generated from per-term source files with includes.

principal component analysis (PCA)¶

Meaning: Principal component analysis is a dimensionality-reduction method that finds new orthogonal axes in directions where the data show large variance.
Why it matters: PCA shows that dimensionality reduction is not just deleting columns. It can rebuild the representation through new axes, preserve large variation first, and make high-dimensional structure easier to inspect.
Related concepts: dimensionality reduction, variance, linear transformation
Core Section: P4-18.1
Appears in:

parameter¶

Meaning: A parameter is a named input position in a function definition. Before a real value is passed in, the parameter records what kind of value the function is prepared to receive.
Why it matters: Parameters help readers distinguish the input names defined inside a function from the actual argument values supplied at call time. The term also needs to be separated from model parameters, which belong to a different layer of AI explanation.
Related concepts: function, model input, model
Core Section: P2-8.5
Appears in: P2-8.6, P2-10.3, P2-15.2, P4-9.1

partial derivative¶

Meaning: A partial derivative is the rate of change computed by varying one input while holding the other inputs fixed. It lets a multivariable problem be examined one variable or direction at a time.
Why it matters: Gradients in models with many parameters depend on reading sensitivity in each direction. This concept separates total change from the effect of one parameter and explains why backpropagation can provide parameter-level update signals.
Related concepts: derivative, gradient, parameter
Core Section: P2-4.3
Appears in: P2-4.4, P2-6.3

partial observability¶

Meaning: Partial observability is a condition where a system must judge from only part of the full state. Some information is visible, but the whole situation is not.
Why it matters: It separates model weakness from information limits. An AI agent, sensor system, or dialogue system may fail because important state is outside the current observation window, not because the rule or model form is wrong.
Related concepts: incomplete information, noise, search
Core Section: P1-6.1
Appears in: P1-6.2

tool execution permission¶

Meaning: Tool execution permission is the executable scope allowed for a tool or action. The same capability can have different permission levels, such as read-only access, write access, or the ability to cause real external changes. Permission defines a boundary before execution and is separate from approval, which asks whether a specific action should proceed now.
Why it matters: The same tool call can carry different risk depending on whether it reads, writes, or deploys. This concept separates can execute from is allowed to execute, which clarifies operational judgment and responsibility. It also explains why permission boundaries and approval steps are both needed.
Related concepts: least privilege, tool use, security, accountability
Core Section: P7-6.2
Appears in: P6-14.1, P6-16.2, P6-18.2, P7-6.1

policy-based reinforcement learning¶

Meaning: Policy-based reinforcement learning directly adjusts the policy that chooses actions instead of first building a scoreboard of values for states or actions. It is closer to changing the behavior rule itself than to calculating a value and then choosing from it. Updating policy parameters toward higher average reward is a representative submethod of this approach.
Why it matters: This concept lets readers separate value-based methods from policy-based methods inside reinforcement learning. It also explains why directly adjusting the policy is natural for continuous actions, stochastic action choices, and methods that update a policy with help from value estimates.
Related concepts: reinforcement learning, reinforcement learning policy, reward
Core Section: P4-19.2
Appears in: P4-2.3, P4-19.3, P4-19.4

population¶

Meaning: A population is the full set of subjects we want to understand. It is not the few data points currently in hand, but the broader set that the data is assumed to represent.
Why it matters: Estimation and error can only be interpreted correctly when it is clear what the sample is part of. This concept prevents treating a dataset as the whole reality and encourages asking how far results can be generalized.
Related concepts: sample, estimation, error
Core Section: P2-5.3
Appears in: P2-5.4, P2-6.1

precision¶

Meaning: Precision is the share of predicted positive cases that are actually positive.
Why it matters: Precision is important when false positives are costly. It helps distinguish a model whose positive predictions are trustworthy from a model that raises many unnecessary alarms.
Related concepts: recall, accuracy, confusion matrix, false positive
Core Section: P1-13.4
Appears in: P4-4.1, P4-4.2, P4-6.1, P4-6.4, P4-8.1

prediction¶

Meaning: Prediction is estimating a next value, state, category, or event from given input or current information. In machine learning, both regression and classification can be understood as forms of prediction. The concept includes not only what will happen later?, but also which result does this input correspond to?
Why it matters: Prediction helps distinguish problem types by asking what kind of result is being estimated from the available input. It also prevents treating prediction only as future forecasting. The output may be a number, category, ranked option, or action choice, and those differences lead to different tasks.
Related concepts: classification, regression, forecast, probability estimate
Core Section: P1-10.1
Appears in: P1-5.2, P1-5.3, P1-7.4, P1-17.1, P1-17.3, P2-15.2, P3-9.9, P4-2.1, P4-5.2, P4-10.1

prediction contract¶

Meaning: An explicit agreement in a prediction problem about which values count as inputs, which result is being predicted, and what information is available up to the prediction time. It includes not only column names, but also when each column is created and whether it can be rebuilt the same way in operations.
Why it matters: If a value created after prediction time leaks into the inputs, the score may look better while the operational prediction problem is broken. A prediction contract keeps features, target candidates, time boundaries, leakage prevention, and reproducibility checked together.
Related concepts: feature, target candidate, data leakage, reproducibility
Core Section: P3-9.7
Appears in: P3-9.7

preprocessing¶

Meaning: Preprocessing is the preparation step that makes input representations more stable before they enter a model, such as handling missing values, scaling features, or encoding categories. It turns raw data into a form the current model can compute with.
Why it matters: The same model can behave very differently depending on how its inputs are prepared. Preprocessing affects distance calculations, training stability, and evaluation results. This concept prevents treating we used data and we prepared data in a model-readable form as the same thing.
Related concepts: feature, feature selection, data leakage
Core Section: P4-7.2
Appears in: P2-12.3, P4-index, P3-5.5, P4-7.1, P4-7.3, P4-9.2, P4-12.2, P4-12.3, P4-3.1, P4-3.2, P4-4.2

pretraining¶

Meaning: Pretraining is the stage in which a model first learns broad language patterns and representations from a large corpus before being adapted to a specific task. It builds a general foundation that can transfer to many tasks, rather than training the model for only one job from the start.
Why it matters: Pretraining explains why modern LLMs can show general language ability before seeing task-specific examples. It helps separate the broad capability formed during large-scale training from later adaptation stages such as fine-tuning or in-context learning.
Related concepts: fine-tuning, in-context learning, language model
Core Section: P1-11.3
Appears in: P6-5.1, P6-5.2, P6-8.1, P6-8.2, P6-10.1, P6-20.1, P6-6.1

proxy target¶

Meaning: A proxy target is a substitute column used like a target when the actual target cannot be observed directly or appears too late. It can make a learning problem possible, but it is not automatically equivalent to the actual target.
Why it matters: Good performance on a proxy target does not guarantee that the original goal is being predicted well. This concept separates we can build a problem now from we are directly solving the original goal, and it forces the distance between the proxy and the actual target to be recorded.
Related concepts: actual target, proxy label, target candidate, supervised learning label
Core Section: P3-9.9
Appears in: P3-9.9

privacy¶

Meaning: Privacy concerns information that can identify or track a person and the rules for how that information is collected, stored, retrieved, reused, and exposed.
Why it matters: AI inputs, retrieval documents, logs, and evaluation samples can contain personal information. Privacy keeps model performance decisions separate from decisions about whether data may be collected, stored, searched, reused, or shown.
Related concepts: security, sensitive information, provenance
Core Section: P1-15.3

probabilistic model¶

Meaning: A probabilistic model represents uncertain states or outcomes with probabilities, conditional probabilities, or distributions. It defines what the number means, not only how high the number is.
Why it matters: Probabilistic models distinguish strict probability interpretation from heuristic scores. They help explain how beliefs or predictions can be updated when new evidence arrives and why the same numeric value can mean different things in different model contexts.
Related concepts: probability, probabilistic reasoning, probability calibration, probability estimate, heuristic, supporting evidence
Core Section: P1-7.3
Appears in: P1-7.2, P1-7.4, P4-1.2, P1-summary

probabilistic reasoning¶

Meaning: Probabilistic reasoning estimates how plausible possible conclusions are when information is incomplete, noisy, or uncertain. It asks which conclusion is more likely under current evidence rather than whether one conclusion is certainly true.
Why it matters: Many real problems cannot be handled by true-or-false rules alone. Probabilistic reasoning explains why AI systems may need scores, probabilities, or distributions when conclusions do not close into one certain answer.
Related concepts: probability, uncertainty, inference
Core Section: P1-2.2
Appears in: P1-index, P1-2.3, P1-3.1, P1-summary, P4-1.1

probability¶

Meaning: Probability expresses uncertainty as a number between 0 and 1. It does not say that an outcome is guaranteed; it describes how plausible an event is among possible outcomes.
Why it matters: Prediction, risk, and classification results often need to be read as possibilities rather than certain answers. Probability helps separate scores, frequencies, probability estimates, and calibrated probabilities.
Related concepts: uncertainty, event, distribution, probability estimate, estimation
Core Section: P2-5.1
Appears in: P1-6.2, P2-5.2, P2-5.3, P2-6.2

probability calibration¶

Meaning: Probability calibration checks whether a model probability score matches real observed frequency. If cases scored around 0.70 are correct about 70 percent of the time, the score is better calibrated.
Why it matters: A high score is not automatically a reliable probability. Probability calibration matters when scores are turned into human review rules, automatic approval thresholds, or risk alerts because the operating decision depends on whether the number can be trusted as a probability.
Related concepts: probability, classification, model evaluation design
Core Section: P1-6.2
Appears in: P1-6.3, P1-7.3, P4-6.4, P4-11.1, P4-15.3

probability estimate¶

Meaning: A probability estimate is a value that represents how likely a particular result or class is. It may look like a probability, but it should not be treated as a reliable probability unless its calibration and meaning are checked.
Why it matters: A number such as 0.9 can be a ranking score, a model confidence value, or a calibrated probability estimate. This distinction matters when thresholds trigger review, action, or risk decisions.
Related concepts: probability, probabilistic model, probability calibration, threshold, human oversight
Core Section: P1-7.3
Appears in: P1-6.3, P3-9.8, P4-6.4, P4-11.1, P4-15.3

prompt¶

Meaning: A prompt is the full input given to a model to produce the current response. It includes not only a one-line question, but also instructions, context, examples, constraints, and output format. It is best understood as a designed bundle of current task conditions.
Why it matters: Prompting is the basic way to adjust the conditions of the current output without retraining the model. This concept separates changing the model itself from changing the input conditions for one response. It also explains why the same model can respond differently depending on context, examples, and requested format.
Related concepts: instruction, input context, in-context learning
Core Section: P1-12.1
Appears in: P1-12.1, P6-2.4, P6-10.1, P6-10.3, P7-5.1

prompt engineering¶

Meaning: Prompt engineering is the practical work of designing instructions, context, examples, and formats, then observing current model behavior and adjusting it more reliably. It is not just writing polished sentences; it is an iterative experiment that changes input conditions based on observed failures.
Why it matters: Prompt engineering is often the fastest first handle for experimentation because it does not require retraining the model. It also helps separate problems that can be reduced through prompting from problems that need retrieval, tool use, model changes, or workflow redesign.
Related concepts: prompt, instruction, instruction tuning
Core Section: P6-11.1

prompt injection¶

Meaning: Prompt injection is an attack in which hidden instructions in user input or external documents try to override or bypass the original system instructions and cause unintended model or reinforcement learning AI agent behavior. The key risk is that text that appears to be content can act as an instruction that changes behavior.
Why it matters: In AI systems connected to search, browsing, file reading, or tool use, external documents can become inputs for action decisions. A hidden instruction can therefore lead to permission abuse or information leakage. This concept explains why least privilege, approvals, execution boundaries, and logs are needed alongside model capability.
Related concepts: security, least privilege, permission, retrieval-augmented generation, RAG, tool use
Core Section: P1-15.3
Appears in: P1-10.3, P1-14.2, P1-14.4, P1-14.5

prompt structuring¶

Meaning: Prompt structuring is the practice of separating roles inside a prompt, such as task, context, examples, constraints, and output format. Instead of placing everything in one block, it arranges information so the role of each part is visible.
Why it matters: In long requests, separating instructions from reference information improves reproducibility and reviewability. It also helps explain why prompt quality is not just a matter of one well-written sentence, but of making the internal roles of the input clear.
Related concepts: prompt, instruction, input context
Core Section: P1-12.2

provenance¶

Meaning: Provenance is source-tracking information that makes it possible to trace which documents and passages influenced an answer or generated result. It is not just a list of links; it records the evidence path behind the result and which material was actually used for judgment.
Why it matters: In RAG, attaching documents is not enough. Review and accountability require knowing which evidence was used. Provenance makes freshness checks, error correction, and later explanation easier by separating generic references from the evidence path that shaped the answer.
Related concepts: supporting evidence, retrieval-augmented generation, RAG, human oversight, document retrieval metadata
Core Section: P1-13.3

proxy label¶

Meaning: A label used as a practical substitute when the true target is delayed, unavailable, or hard to observe directly.
Why it matters: It helps readers see that available labels may not exactly match the real goal.
Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
Core Section: P1-8.1

protected expression¶

Meaning: The concrete expressive form of a copyrighted work, as distinct from unprotected facts, ideas, or methods.
Why it matters: It matters because AI learning material must separate reusable ideas from protected wording, images, structure, or other expression.
Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
Core Section: P1-15.2

probabilistic choice¶

Meaning: A choice made from a probability distribution rather than by always selecting the single highest-scoring option.
Why it matters: It explains why generative systems can produce different outputs from the same prompt under sampling settings.
Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
Core Section: P1-10.2

point prediction¶

Meaning: A prediction that returns one specific value rather than a distribution or interval.
Why it matters: It is easy to read as a final answer, so it must be distinguished from uncertainty-aware predictions.
Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
Core Section: P1-6.3

probabilistic prediction¶

Meaning: A prediction that expresses possible outcomes with probabilities rather than returning only one fixed answer.
Why it matters: It helps separate confidence, uncertainty, calibration, and decision thresholds.
Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
Core Section: P1-6.3

probability distribution¶

Meaning: A description of how probability is assigned across possible values or outcomes.
Why it matters: It matters because models often compute a distribution first and only later choose or report one result.
Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
Core Section: P2-5.2