C

Calibration

• Category: Evaluation & safety
• What it actually means: The agreement between a system's stated confidence and the observed frequency with which predictions at that confidence are correct.
• Why it matters: A system can be accurate on average yet dangerously overconfident on the cases where people rely on its score.
• In practice: Bucket predictions by confidence, compare confidence with empirical accuracy, and recalibrate or abstain when the gap is unacceptable.
• Common confusion: Calibration measures confidence reliability, not overall accuracy, factuality, or reasoning quality.
• Related terms: Softmax, Evaluation (Eval), Precision & Recall, Logits
• Sources: On Calibration of Modern Neural Networks

Canary Release

• Category: Reliability & operations
• What it actually means: A deployment strategy that exposes a new version to a limited slice of traffic or infrastructure before expanding the rollout.
• Why it matters: It limits the impact of defects and gives you production evidence before the new model, prompt, agent, or service reaches everyone.
• In practice: Route a small eligible cohort to the release, compare quality and operational metrics with the control, and stop or roll back on predefined failures.
• Common confusion: A canary release limits exposure; it does not replace pre-deployment tests, approval, or rollback preparation.
• Related terms: Evaluation (Eval), Observability, Rollback, Verification Gate
• Sources: Kubernetes Deployments: Canary Deployment

Chain of Thought (CoT)

• Category: Prompting & context
• What people say: Asking the model to show every step of its thinking.
• What it actually means: Intermediate reasoning used to decompose a task before producing an answer. A prompt can request a visible rationale, while some systems use internal reasoning that is not returned to the user.
• Why it matters: Decomposition can help on multi-step tasks, but a fluent rationale is not proof that the answer is correct or that the text faithfully represents the model's internal computation.
• In practice: Ask for a concise plan, independently check the result, and request verifiable calculations or citations instead of relying on a long reasoning transcript.
• Common confusion: Chain of thought is not a substitute for tools, tests, or external verification.
• Learn it: Few-Shot and Chain of Thought
• Related terms: Prompt Engineering, Verification Gate, Evaluation (Eval)

Checkpoint

• Category: Agents & tools
• What it actually means: A durable snapshot used to resume from a known boundary. In a workflow, it stores operational state and artifact references. In model training, it can store parameters, optimizer state, scheduler state, and the training position.
• Why it matters: Long-running workflows and training runs can recover from interruption without replaying completed work or losing expensive progress.
• In practice: Save an agent's accepted patch and test evidence after a verified step, or save a training run's weights, optimizer state, random state, and data position before shutdown.
• Common confusion: A workflow checkpoint and a model-training checkpoint serve the same recovery goal but preserve different state. Neither is merely a transcript or a weights file with no resume metadata.
• Learn it: Checkpoint Save and Resume; Repository Memory and State
• Related terms: Agent State, Durable Execution, Parameter, Optimizer

Chunked Prefill

• Category: Infrastructure & serving
• What it actually means: A serving technique that divides a long prompt's prefill work into smaller schedulable pieces so prompt processing can interleave with decode work from other requests.
• Why it matters: One long prompt can otherwise occupy the accelerator and delay active generations, producing poor tail latency even when total throughput looks healthy.
• In practice: Choose a chunk policy from measured workloads, account for scheduling overhead, and compare prefill completion, decode latency, and goodput under mixed prompt lengths.
• Common confusion: Chunked prefill changes how prompt computation is scheduled. It does not split the user's context into independent semantic chunks or change the model's context window.
• Learn it: vLLM Serving Internals
• Related terms: Prefill, Decode Phase, Dynamic Batching, Tail Latency
• Sources: Sarathi-Serve

Chunking

• Category: Retrieval & generation
• What people say: Splitting documents into pieces.
• What it actually means: Dividing source material into retrievable units before indexing. Chunk boundaries, overlap, metadata, and document structure determine whether retrieval returns enough context without flooding the prompt.
• Why it matters: The right chunking strategy depends on document shape, query type, embedding model, and evaluation results. There is no universal token size or overlap percentage.
• In practice: Keep headings and code blocks intact, attach source metadata, then measure retrieval quality on real questions before tuning size.
• Related terms: RAG (Retrieval-Augmented Generation), Reranker, Grounding

Circuit Breaker

• Category: AI-native development
• What it actually means: A reliability control that temporarily stops calls to a dependency after failures cross a threshold, then probes whether the dependency has recovered.
• Why it matters: It prevents repeated model or tool failures from consuming latency, budget, and capacity across the rest of the system.
• In practice: Open the breaker after repeated provider timeouts, fail over or return a controlled response, then allow a limited health probe after a cooldown.
• Common confusion: A circuit breaker reacts to dependency health. A rate limit controls allowed request volume.
• Related terms: Retry with Backoff, Rate Limit, Model Router, Backpressure

CNN (Convolutional Neural Network)

• Category: Models & inference
• What people say: A neural network for images.
• What it actually means: A neural network that uses convolution operations (sliding filters over the input) to detect local patterns. Stacking convolutions detects increasingly complex features: edges, textures, objects.
• Common confusion: Convolutions also work on audio, time series, and other grid-like data.
• Related terms: Feature, Inductive Bias, Activation Function

Coding Agent

• Category: AI-native development
• What it actually means: An agent specialized for software work that can inspect a repository, edit files, run development tools, and use their outputs to advance a scoped engineering task.
• Why it matters: Its value depends on repository context, tool permissions, review boundaries, and verification, not only code generation quality.
• In practice: Give the agent an issue, a scope contract, repository instructions, and a test command; review the resulting patch and evidence before accepting it.
• Common confusion: A coding assistant that only suggests text is not necessarily an agent. The agent acts through tools and observes results.
• Learn it: Skill Discovery and Progressive Disclosure
• Related terms: Agent Harness, Repository Map, Patch, Scope Contract, Reviewer Agent

Compensating Action

• Category: Agents & tools
• What it actually means: A deliberate operation that semantically counteracts a completed side effect when the original operation cannot be rolled back atomically.
• Why it matters: Multi-step agent workflows cross databases and external services where a later failure cannot undo earlier writes through one transaction.
• In practice: If a booking workflow charges a card but the reservation fails, issue a tracked refund and preserve both events rather than deleting history.
• Common confusion: Compensation is a new business action, not time travel. It can fail and therefore needs idempotency, monitoring, and escalation.
• Related terms: Durable Execution, Idempotency, Checkpoint, Approval Gate
• Sources: Sagas

Content Provenance

• Category: Security & governance
• What it actually means: Verifiable information about the origin and editing history of a piece of media or other digital content, including the actors, tools, transformations, and assertions attached to it.
• Why it matters: Generative systems make origin claims difficult to infer from appearance alone, so consumers and platforms need inspectable evidence about how content was produced.
• In practice: Bind provenance assertions to the content, sign them with controlled identities, preserve transformation history, and show clearly when evidence is missing or cannot be verified.
• Common confusion: Provenance can establish who asserted a history and whether the record was altered. It does not prove that the depicted event is true or that the content is harmless.
• Learn it: Watermarking, SynthID, Stable Signature, and C2PA
• Related terms: Data Provenance, Provenance Attestation, Audit Log, Grounding
• Sources: C2PA Technical Specification

Context Compression

• Category: Prompting & context
• What it actually means: Reducing the token footprint of source material while attempting to preserve the information required for a later model decision.
• Why it matters: Compression can make long tasks fit within budget, but every omitted detail creates a risk that the model loses evidence, constraints, or unresolved state.
• In practice: Preserve authoritative facts and identifiers verbatim, summarize redundant history, attach source pointers, and test the compressed context on representative tasks.
• Common confusion: Compression is lossy unless it retains the full original. A shorter summary is not automatically an equivalent context.
• Related terms: Token Budget, Context Engineering, Progressive Disclosure, Handoff
• Sources: LLMLingua

Context Engineering

• Category: Prompting & context
• What it actually means: Designing the full information environment supplied to a model at each step, including instructions, selected files, retrieved evidence, tool results, examples, state, and output constraints.
• Why it matters: Model performance often fails because relevant evidence is missing, stale, badly ordered, or overwhelmed by noise.
• In practice: Build a compact task packet with the goal, repository rules, relevant interfaces, recent tool output, and unresolved decisions, then update it as state changes.
• Common confusion: Prompt engineering focuses on instruction wording. Context engineering also decides what evidence and state enter the model's working context.
• Learn it: Context Engineering
• Related terms: Context Window, Progressive Disclosure, Agent State, Repository Map

Context Window

• Category: Prompting & context
• What people say: How much the model remembers.
• What it actually means: The maximum token capacity available to one model inference under a specific model and API contract. The capacity may include system instructions, messages, retrieved content, tool exchanges, and generated output, with provider-specific accounting and output limits.
• Why it matters: Conversation history is only available when the application sends or reconstructs it. A large window does not guarantee that every included detail will be used reliably.
• Common confusion: Context is temporary input to an inference. Durable memory is stored outside the model and selected back into later context.
• Learn it: Context Engineering
• Related terms: Token Budget, Context Engineering, Prompt Cache, Agent State

Continuous Batching

• Category: Infrastructure & serving
• What it actually means: A serving scheduler that adds and removes generation requests at iteration boundaries instead of waiting for every request in a fixed batch to finish.
• Why it matters: Autoregressive requests produce different output lengths, so continuous batching can keep accelerators utilized without forcing short requests to wait for the longest one.
• In practice: Admit new requests when capacity becomes available, track per-request latency, and apply backpressure when the live batch or KV-cache budget is full.
• Common confusion: Continuous batching is an inference scheduling policy, not gradient accumulation or a training batch-size technique.
• Related terms: Dynamic Batching, Decode Phase, Backpressure, Rate Limit
• Sources: Orca

Contrastive Learning

• Category: Math & training
• What people say: Learning by comparison.
• What it actually means: Training by pulling similar pairs closer and pushing dissimilar pairs apart in embedding space. CLIP uses this: matching image-text pairs vs non-matching ones.
• Related terms: Embedding, Cosine Similarity, Loss Function

Cosine Similarity

• Category: Data & representations
• What people say: How similar two vectors are.
• What it actually means: The normalized dot product of two vectors. It compares their direction rather than their magnitude and ranges from -1 to 1 for real-valued vectors.
• Common confusion: High cosine similarity only has meaning relative to the embedding model and the data distribution. It does not prove factual or semantic equivalence.
• Related terms: Embedding, Semantic Search, Reranker

Cost per Successful Task

• Category: AI-native development
• What it actually means: Total system cost divided by the number of tasks that satisfy a defined success criterion, including retries, failed runs, tool use, and evaluation overhead.
• Why it matters: A cheap model call can produce an expensive workflow if it fails often or requires repeated human correction.
• In practice: Measure provider charges and infrastructure cost across 100 repository tasks, then divide by the number whose patches pass tests and review.
• Common confusion: Cost per token measures usage. Cost per successful task measures useful outcomes.
• Related terms: Evaluation (Eval), Retry with Backoff, Model Router, Verification Gate

Cross-Attention

• Category: Multimodal systems
• What it actually means: Attention in which the query representation comes from one sequence or representation while keys and values come from another.
• Why it matters: It gives one stream a learnable way to retrieve information from another, such as language tokens attending to visual features.
• In practice: State which stream supplies queries, keys, and values, apply masks for missing or invalid positions, and inspect whether the model still performs when one modality is ablated.
• Common confusion: Cross-attention is not intrinsically multimodal. It can connect two text sequences or other representations; self-attention instead derives queries, keys, and values from the same sequence representation.
• Related terms: Attention, Self-Attention, Vision-Language Model (VLM), Multimodal Fusion
• Sources: Attention Is All You Need

Cross-Entropy

• Category: Math & training
• What people say: The classification loss.
• What it actually means: A loss based on the negative log probability assigned to the target outcome. In next-token training, it penalizes the model when it assigns low probability to the observed next token.
• Common confusion: Perplexity is the exponentiated average cross-entropy only when the averaging and logarithm base are defined consistently.
• Related terms: Loss Function, Softmax, Perplexity

CUDA

• Category: Models & inference
• What people say: GPU programming.
• What it actually means: NVIDIA's platform and programming model for general-purpose computation on compatible GPUs. Deep-learning frameworks use CUDA libraries and kernels to execute many tensor operations in parallel.
• Common confusion: GPU acceleration is not synonymous with CUDA; other hardware and software stacks exist.
• Related terms: Tensor, Mixed Precision, JAX