Chain of thought
A prompting technique where you ask the model to show its reasoning step by step before giving a final answer. Often improves accuracy on math and logic tasks. Costs more tokens than a direct answer.
Glossary
Glossary. 193 terms. Three categories.
AI, federal, and military terms with definitions, acronyms expanded, and links to the lessons where each term shows up. Search or browse by group.
AI / Foundations
Context window
The amount of text a model can read and write in a single turn, measured in tokens. A 200K context window holds roughly 150,000 words. Once you exceed it, the model drops or forgets the earliest content.
Few-shot prompting
Showing the model 2-5 examples of input and output before asking for the real answer. Steers tone, format, and accuracy. Often the fastest fix when a zero-shot prompt fails.
Generative AI
Models that produce new text, images, audio, or code rather than just classify existing inputs. ChatGPT, Claude, and Midjourney are generative AI. The output is generated token by token or pixel by pixel.
Hallucination
When a model produces confident output that is false. The model is not lying. It is pattern-matching to plausible-sounding text without a fact-check step. You verify before you trust.
Inference
Running a trained model to produce output. Every time you send a prompt and get a response, that is inference. It is the part you pay for at the API level.
LLM Large Language Model
A neural network trained on huge amounts of text to predict the next token. GPT-4, Claude, and Llama are LLMs. They are the engine behind most chat and assistant products today.
Model card
A short document published by a model maker that lists training data, intended uses, known limits, and evaluation scores. Read it before you deploy a model in production. It is the spec sheet for the model.
Parameters
The learned numbers inside a model that get tuned during training. A 70B model has 70 billion of them. More parameters often means more capability but also more cost and slower inference.
Pre-training
The first and most expensive training stage, where a model learns language patterns from trillions of tokens of web text, books, and code. It produces a base model that knows a lot but does not yet follow instructions well. Post-training fixes that.
Prompt
The text you send to a model to get a response. A good prompt has context, a task, and standards. A bad prompt has a vague question and hopes for the best.
System prompt
A high-priority instruction that sets the model's role, tone, and rules for the whole conversation. The user does not see it. It is where you put the operator guidance for an assistant.
Token
The unit of text a model reads and writes. One token is roughly 4 characters of English, or about 3/4 of a word. You are billed per token in and per token out.
Training
The process of adjusting a model's parameters by showing it examples and correcting its predictions. Training is expensive and slow. Inference is what you do after.
Zero-shot prompting
Asking the model to do a task with no examples, just instructions. Works well for common tasks like summarization. Falls down when format or tone matters and the model has to guess.
AI / RAG & embeddings
Chunk
A small piece of a longer document, usually 200-1000 tokens, that you embed and store for retrieval. You retrieve chunks, not whole documents. Chunk size affects both cost and answer quality.
Chunking strategy
The rules for splitting documents into chunks. Fixed-size, paragraph, semantic, and recursive are common strategies. The right choice depends on the document type and the questions you expect.
Cosine similarity
A math measure of how close two vectors point in the same direction, from -1 to 1. Higher means more similar. The default scoring method for most vector databases.
Dense retrieval
Finding relevant chunks by comparing embedding vectors. Catches meaning even when the words do not match. The opposite of keyword search, which only finds exact word matches.
Embedding
A list of numbers, usually 768 to 3072 long, that represents the meaning of a piece of text. Similar meanings produce similar vectors. Embeddings are how RAG systems decide what to retrieve.
Hybrid search
Combining dense retrieval with keyword search and merging the results. Catches both semantic matches and exact terms like product codes or names. Usually beats either method alone.
RAG Retrieval Augmented Generation
A pattern where you search a knowledge base for relevant chunks, then stuff them into the prompt before asking the model. Lets the model answer questions about your private data. The most common production AI pattern outside chat.
Reranking
A second-pass model that reorders the top 20-50 retrieval hits by relevance before sending the best 3-5 to the LLM. Costs extra latency. Often the biggest single quality win in a RAG pipeline.
Similarity search
Finding the chunks whose embeddings are closest to the query embedding. The core operation a vector database performs. Returns the top K results, where K is usually 5 to 50.
Vector database
A database built to store embeddings and find the closest matches fast. Pinecone, Weaviate, Qdrant, and pgvector are common picks. The storage layer under most RAG systems.
AI / Agents & tool use
Agent
An LLM in a loop that picks a tool, runs it, reads the result, and picks the next step until the task is done. Adds power and adds risk. You give up some control to get autonomy.
Agentic
Describes a system where the model decides what to do next, not a fixed script. An agentic workflow can plan, retry, and branch. A non-agentic one runs the same steps every time.
Autonomous agent
An agent that runs for many steps without a human in the loop. Useful for long tasks like research or code refactors. Higher risk because mistakes compound.
Function calling
A feature where the model outputs a structured request to call a specific function with specific arguments. Your code runs the function and sends the result back. The mechanism behind most tool use.
MCP Model Context Protocol
An open standard for connecting AI models to tools and data sources. Lets you write a tool once and use it from any MCP-compatible client. Replaces a pile of custom integrations.
Multi-agent system
A setup where multiple agents with different roles work together, often with a coordinator. Used for research, coding, and review tasks. Costs more tokens and adds coordination bugs.
Planner-executor
An agent pattern where one model writes the plan and another runs the steps. Keeps planning context separate from execution. Easier to debug than a single agent doing both.
ReAct Reasoning and Acting
An agent loop pattern: the model writes a thought, picks an action, reads the result, then thinks again. One of the earliest and still common agent designs. Easy to log and inspect.
Tool use
When the model calls external functions like search, calculators, or APIs to do things it cannot do alone. The model decides which tool to call and with what arguments. Your code runs the tool.
AI / Post-training & alignment
Alignment
The work of making a model behave the way humans want, including being helpful, honest, and harmless. Done with techniques like RLHF and DPO. Imperfect, which is why jailbreaks still work.
Constitutional AI
An alignment method where the model critiques and revises its own outputs against a written set of principles. Used by Anthropic to train Claude. Reduces the need for human labelers on every example.
DPO Direct Preference Optimization
An alignment method that trains a model directly on pairs of preferred and rejected responses, without a separate reward model. Simpler and cheaper than RLHF. Now the default for many open models.
Fine-tuning
Continuing training on a smaller, specialized dataset to teach a base model new skills or styles. Useful when prompting alone is not enough. Most teams should try prompting and RAG first.
Instruction tuning
Fine-tuning a base model on examples of instructions and good responses. Turns a raw text predictor into something that follows directions. A subset of SFT.
LoRA Low-Rank Adaptation
A fine-tuning method that trains small adapter matrices instead of all model weights. Cuts memory and compute by orders of magnitude. The standard way to fine-tune open models on one or two GPUs.
QLoRA Quantized Low-Rank Adaptation
LoRA done on top of a model that has been quantized to 4-bit precision. Lets you fine-tune a 70B model on a single consumer GPU. Trades a small accuracy hit for a huge cost cut.
RLAIF Reinforcement Learning from AI Feedback
An alignment method that uses another model, not humans, to rank responses for reward signal. Cheaper and faster than RLHF. Quality depends on how good the judge model is.
RLHF Reinforcement Learning from Human Feedback
An alignment method where humans rank model responses, a reward model learns from those rankings, and the model is trained to maximize the reward. The technique behind ChatGPT's launch. Expensive and slow to run.
SFT Supervised Fine-Tuning
Fine-tuning on labeled examples of correct input-output pairs. The first step of most post-training pipelines, before RLHF or DPO. The simplest kind of fine-tuning.
AI / Architecture
CNN Convolutional Neural Network
A network design built for grids of data like images. Uses small filters that slide across the input to detect patterns. Powered most image AI before transformers caught up.
Decoder
The half of a transformer that produces output one token at a time. GPT-style models are decoder-only. Each new token attends to all the tokens before it.
Diffusion model
A generative model that learns to turn random noise into a clean image, audio clip, or video by reversing a noise process step by step. Powers Stable Diffusion, Midjourney, and Sora. Different math from LLMs.
Encoder
The half of a transformer that reads input and turns it into a dense representation. BERT is encoder-only. Good for classification and embeddings, not for generating long outputs.
Encoder-decoder
A transformer with both an encoder and a decoder, often used for translation and summarization. T5 and the original Transformer paper are encoder-decoder. Less common than decoder-only today.
KV cache Key-Value cache
Stored attention vectors from earlier tokens so the model does not recompute them on each new token. Cuts inference cost a lot. Eats GPU memory, which is why long contexts get expensive.
LSTM Long Short-Term Memory
A type of RNN with gates that let it remember information across long sequences. Standard for language tasks before transformers. Still used in some time-series and audio work.
MLP Multi-Layer Perceptron
A stack of fully connected neural network layers. The simplest deep learning design and a building block inside transformers. Every transformer block contains an MLP after the attention step.
MoE Mixture of Experts
A design where the model has many expert subnetworks and a router that picks a few to run for each token. Lets total parameters grow without proportional cost. Used in Mixtral, DeepSeek, and GPT-4.
Multi-head attention
Running self-attention in parallel several times with different learned projections. Each head can focus on a different kind of relationship in the text. Standard inside every transformer block.
Multimodal
A model that handles more than one input type, such as text plus images, audio, or video. GPT-4o and Claude with vision are multimodal. The same model produces a response across modalities.
RNN Recurrent Neural Network
A network design that processes sequences one step at a time, carrying state forward. Used in language models before transformers. Slow to train because it cannot parallelize across time.
Self-attention
The mechanism that lets each token in a sequence look at every other token and decide what to weigh. The core operation inside a transformer. The reason long context is expensive: cost grows with the square of sequence length.
Transformer
The neural network design behind almost every modern LLM. Built from stacks of attention and MLP layers. Introduced in 2017 and still the dominant pattern.
AI / Production
Autoscaling
Adding or removing inference servers based on load. Keeps latency down at peak and cost down at idle. Tricky for GPUs because they take minutes to warm up.
Drift detection
Monitoring for changes in input data or model output over time. Catches problems like a new customer segment producing weird prompts, or quality slowly dropping. A standard part of mature LLM ops.
Observability
The ability to see what your AI system is doing in production through logs, traces, and metrics. For LLMs this includes prompts, responses, token counts, latency, and tool calls. Without it you debug blind.
Ollama
A simple tool for running open-source LLMs on your own machine or server. Good for prototyping and air-gapped use. Not built for high-throughput production.
OpenTelemetry
An open standard for collecting traces, metrics, and logs from any system. The default for LLM observability tools today. Lets you switch monitoring vendors without rewriting instrumentation.
SGLang Structured Generation Language
An open-source serving framework focused on fast structured output and complex prompt patterns. Strong on constrained decoding and prefix caching. A vLLM alternative for some workloads.
SLI Service Level Indicator
A measurement of one aspect of service health, like p95 latency or error rate. SLIs are the raw numbers you track. SLOs are the targets you set against them.
SLO Service Level Objective
A target you commit to, like 99 percent of requests under 2 seconds. SLOs make trade-offs explicit. If you miss them, that is a real signal, not a vibe.
TGI Text Generation Inference
Hugging Face's open-source LLM serving framework. Supports many model formats and quantization options. A common pick alongside vLLM for self-hosted inference.
vLLM
An open-source LLM serving framework built around paged attention and continuous batching. Squeezes a lot more throughput out of a GPU than naive serving. The default for self-hosted production today.
AI / Cost & ops
Batch API
A way to submit many prompts at once and get results within hours instead of seconds. Usually 50 percent cheaper than the live API. Use it for offline jobs like classifying old tickets or generating embeddings.
Latency
How long a request takes from send to last token. Measured in milliseconds. Driven by prompt length, output length, model size, and serving setup.
p95 latency
The latency that 95 percent of requests come in under. Catches slow outliers that an average hides. The number you usually set SLOs against.
Prompt caching
Reusing the model's computation for a long prompt prefix that you send over and over. Cuts both cost and latency, often by 80 to 90 percent on cached tokens. Worth setting up any time you have a stable system prompt.
Sampling
The method for picking the next token from the model's probability distribution. Temperature, top-p, and top-k are sampling controls. Affects creativity and consistency of output.
Streaming
Sending tokens to the client as they are generated instead of waiting for the full response. Cuts perceived latency a lot. Standard for chat interfaces.
Temperature
A sampling control between 0 and 2 that sets how random the output is. 0 is deterministic and repetitive. 0.7 to 1.0 is normal chat range. Above 1.2 starts to break down.
Throughput
How many requests or tokens per second your system can serve. The lever you pull to bring down cost per request at scale. Often trades against latency.
Top-k
A sampling control that limits the next token choice to the K most likely options. K of 40 is common. Simpler than top-p but less adaptive to confidence.
Top-p
A sampling control, also called nucleus sampling, that picks from the smallest set of tokens whose total probability hits P. 0.9 to 0.95 is typical. Cuts off long-tail nonsense without being rigid.
AI / Security
Adversarial input
Input designed to make a model misbehave, leak data, or skip safety rules. Includes prompt injection, jailbreaks, and crafted images. Assume adversarial input will hit anything you ship to the public.
Guardrails
Checks that sit before, during, or after a model call to block bad input or output. Can be regex, classifiers, or other LLMs. The seatbelt around your LLM, not the brakes.
Indirect prompt injection
Prompt injection delivered through content the model reads, like a web page, email, or document, not through the user's direct input. Especially dangerous for agents that browse or read files. Hard to defend against fully.
Jailbreak
A prompt trick that gets a model to ignore its safety training. Roleplay scenarios, encoded requests, and refusal-suppression are common patterns. New jailbreaks appear constantly.
Model extraction
An attack where someone queries your model enough times to clone its behavior into their own model. A real risk for fine-tuned models exposed via API. Rate limits and watermarking help.
OWASP Top 10 for LLMs OWASP
A community list of the most common security risks in LLM apps, including prompt injection, sensitive info disclosure, and supply chain attacks. The starting point for an LLM threat model. Updated as new attacks emerge.
Prompt injection
An attack where input text tells the model to ignore its instructions and do something else. Hard to fully block. The number-one risk on the OWASP Top 10 for LLMs.
Red team
A group that attacks your AI system on purpose to find weaknesses before adversaries do. Tries jailbreaks, prompt injection, data exfiltration, and abuse patterns. Standard practice before launching a model to the public.
SBOM Software Bill of Materials
A list of every component in your software, including model weights, datasets, and dependencies. Required for federal AI procurement under recent rules. Lets you check fast when a vulnerability is disclosed.
Supply chain attack
An attack delivered through a component you depend on, like a poisoned model on Hugging Face or a backdoored Python package. Affects every system that pulls in the bad component. SBOMs and pinned versions help.
Federal / Policy & frameworks
AI use case inventory
A list every federal agency must publish of how it uses AI. Required by OMB M-24-10. If your system isn't on it, you have a paperwork problem.
Executive Order on AI
The White House directive setting AI policy across the federal government. The current version drives agency AI strategies, hiring, and safety testing requirements. Expect the exact EO number to change between administrations.
GOVERN function
The NIST AI RMF function that sets up the policies, roles, and accountability for AI risk. It's the org chart and the rulebook. Without GOVERN, the other functions have nowhere to land.
MANAGE function
The NIST AI RMF function for acting on identified risks. You prioritize, treat, monitor, and respond to incidents. It's where the playbooks live.
MAP function
The NIST AI RMF function for naming the system's context, purpose, and risks. You write down who uses it, what it decides, and what could go wrong. Most teams skip this and pay for it later.
MEASURE function
The NIST AI RMF function for testing the system against the risks you mapped. Accuracy, bias, robustness, security. If you can't measure it, you can't claim it's safe.
NIST AI RMF National Institute of Standards and Technology AI Risk Management Framework
The voluntary federal framework for managing AI risk. Four functions: GOVERN, MAP, MEASURE, MANAGE. It's the spine that most agency AI policies and contract clauses now reference.
OMB M-24-10
The 2024 OMB memo telling federal agencies how to govern their use of AI. It requires a Chief AI Officer, an AI use case inventory, and minimum safeguards for rights- and safety-impacting AI. This is the operational rulebook for civilian agencies.
OSCAL Open Security Controls Assessment Language
A NIST-built machine-readable format for security control documentation. The point is to stop shipping SSPs as Word docs. FedRAMP and cATO programs are moving toward OSCAL for faster review.
Risk impact assessment
A written analysis of how an AI system could harm people, mission, or rights. Required by OMB M-24-10 for rights- and safety-impacting use cases. It forces you to name the harm before you ship.
Federal / Compliance & ATO
ATO Authority to Operate
A formal sign-off from an Authorizing Official that lets a system run on a federal network. Without an ATO, your code is a demo. With one, it's a system of record.
cATO Continuous Authority to Operate
An ATO that stays valid as long as you keep proving controls are working through continuous monitoring. You earn the right to push code without re-authorizing every change. DoD pushes this hard for software factories.
FedRAMP Federal Risk and Authorization Management Program
The federal program that authorizes commercial cloud services for government use. One authorization, many agencies. If your SaaS isn't FedRAMP, civilian agencies can't buy it for sensitive workloads.
FedRAMP High
FedRAMP for systems where a breach causes severe harm: law enforcement, emergency services, financial systems, health records. Roughly 421 controls. The bar is high and the audit is brutal.
FedRAMP Low
FedRAMP for systems where a breach causes limited harm. Public-facing data and low-impact workloads. Fewer controls, faster path, but most real federal AI work needs Moderate or higher.
FedRAMP Moderate
FedRAMP for systems where a breach causes serious harm: most PII, most CUI, most agency line-of-business apps. The default tier for civilian federal SaaS. Around 325 controls.
IL2 Impact Level 2
DoD cloud tier for public or non-critical unclassified data. Maps roughly to FedRAMP Moderate. Rarely used alone for real DoD AI work because most DoD data is CUI.
IL4 Impact Level 4
DoD cloud tier for CUI and non-critical mission data. The most common tier for DoD unclassified AI work. Requires a FedRAMP Moderate baseline plus DoD-specific controls.
IL5 Impact Level 5
DoD cloud tier for higher-sensitivity CUI, mission-critical data, and National Security Systems at the unclassified level. Stronger isolation than IL4. Cleared US persons only for admin access.
IL6 Impact Level 6
DoD cloud tier for classified data up to Secret. Runs on SIPR-connected cloud regions. Cleared workforce, physical isolation, and a much smaller list of authorized providers.
SCAP Security Content Automation Protocol
A NIST standard for automating security configuration checks. SCAP scanners read STIG content and report what passed and failed. It's how compliance moves from a spreadsheet to a pipeline.
STIG Security Technical Implementation Guide
DISA's hardening checklists for operating systems, databases, browsers, and now containers. If you ship to DoD, your stack gets STIG'd. Plan for the findings before the assessor shows up.
Federal / Classification & data
CUI Controlled Unclassified Information
Unclassified information the government still requires you to protect. Examples: contract data, PII, export-controlled tech data, law enforcement info. CUI has rules but no clearance. Most federal AI work touches CUI.
DD-254
The contract security classification specification. It tells a contractor what classified work they can do, at what level, and where. If you're cleared and working a contract, the DD-254 sets the rules of the road.
EAR Export Administration Regulations
Commerce Department rules on exporting dual-use technology, including most commercial software and AI models. Sharing a model weight with a foreign national can be an export. EAR governs the commercial side; ITAR governs the defense side.
Foreign disclosure
The formal process for sharing classified or controlled information with foreign governments or nationals. Has its own officers and approval chains. Don't paste anything into a chat tool that a Foreign Disclosure Officer hasn't cleared.
ITAR International Traffic in Arms Regulations
State Department rules on exporting defense articles and technical data. Covers weapons, military systems, and the technical data behind them. ITAR violations are personal and criminal, not just corporate fines.
SAP Special Access Program
A classified program with access controls beyond a normal clearance. You get read in, by name, to a specific program. Most cleared people will never see one. The ones who do don't talk about it.
SCI Sensitive Compartmented Information
Classified intelligence information handled in compartments. You need TS clearance plus a read-on to a specific compartment. SCI work happens in a SCIF, not at your laptop.
Secret
The classification level for information whose unauthorized disclosure would cause serious damage to national security. Handled on SIPR. The most common clearance level in DoD.
Top Secret
The classification level for information whose unauthorized disclosure would cause exceptionally grave damage to national security. Handled on JWICS or in SCIFs. Often paired with SCI access.
Federal / Contracting
BPA Blanket Purchase Agreement
A pre-negotiated agreement that lets an agency place repeat orders without a new solicitation each time. Faster than a full procurement. Common for staffing, training, and recurring cloud spend.
FAR Part 15
The Federal Acquisition Regulation part governing contracting by negotiation. The default rules for large, complex, source-selection awards. If your proposal has a written technical volume and oral presentations, you're in Part 15.
GWAC Government-Wide Acquisition Contract
A contract one agency runs that other agencies can buy off of. Alliant, CIO-SP, and 8(a) STARS are examples. Vendors compete once for the GWAC, then compete again for each task order.
IDIQ Indefinite Delivery, Indefinite Quantity
A contract that sets terms and ceiling price, then lets the government order specific work over time as task orders. Most large federal services work runs on IDIQs. You win the IDIQ to get the right to compete for task orders.
OASIS
A GSA government-wide contract vehicle for complex professional services. OASIS+ is the current generation. If you do federal consulting or systems integration, this is one of the big front doors.
OTA Other Transaction Authority
A non-FAR contracting authority DoD and a few other agencies use for prototypes and research. Faster, more flexible, and friendlier to non-traditional vendors. Most fast-moving DoD AI work rides on OTAs.
PWS Performance Work Statement
A contract document that describes the outcomes the government wants, not the steps to get there. The vendor proposes how. Good for outcome-based AI work.
SBIR Small Business Innovation Research
A federal program that funds small business R&D in three phases. Phase I proves the idea, Phase II builds it, Phase III commercializes it. Common entry point for AI startups working with DoD.
SeaPort-NxG
The Navy's main IDIQ for engineering and professional services. NxG is the current generation. If you're chasing Navy AI work, this is one of the big vehicles.
SOO Statement of Objectives
A short contract document that names the government's objectives and lets the vendor propose the full approach. Even more outcome-focused than a PWS. Common in OTA and innovation-style buys.
SOW Statement of Work
A contract document that prescribes what the vendor will do, step by step. Tells you the tasks, deliverables, and schedule. The most rigid of the three; SOO and PWS give vendors more room to propose.
STTR Small Business Technology Transfer
A federal R&D program like SBIR, but the small business must partner with a research institution. Phase structure mirrors SBIR. Useful when university IP is part of the AI work.
Federal / Money
Continuing Resolution
A temporary law that keeps the government funded at last year's levels when Congress hasn't passed a new budget. Under a CR, you usually can't start new programs. Plan your AI roadmap around the CR risk, not against it.
FYDP Future Years Defense Program
The five-year DoD budget plan that flows out of the POM. Shows what the department intends to spend across programs. If your AI line item isn't in the FYDP, it doesn't really exist yet.
O&M Operation and Maintenance
Single-year money that funds day-to-day operations: salaries, training, sustainment, cloud bills. You have to spend it in the fiscal year it's appropriated. Most production AI workloads run on O&M.
OCO Overseas Contingency Operations
A separate appropriation that funded war operations outside the base budget. Largely folded back into base accounts in recent budgets, but the term still shows up. Treat it as legacy unless your contract says otherwise.
POM Program Objective Memorandum
The DoD document where a service proposes how it will spend money across the next five years. The output of the second P in PPBE. POM season is when AI programs get cut or saved.
PPBE Planning, Programming, Budgeting, and Execution
The DoD process for deciding what to fund and then spending it. Four phases, multi-year cycle. If you want money for an AI program in FY28, you fight for it now.
Procurement
Multi-year money for buying things that last: aircraft, ships, large IT systems. You usually can't use it for salaries or cloud subscriptions. For AI, it tends to show up only when hardware is part of the deliverable.
RDT&E Research, Development, Test, and Evaluation
Multi-year money for research and prototypes. You can spend it across two fiscal years, which makes it the friendliest color of money for early-stage AI work. Different from O&M, which is single-year and meant for sustainment.
Supplemental appropriation
Extra funding Congress passes outside the normal budget cycle, usually for emergencies or new conflicts. Comes with its own rules and deadlines. Plan AI work funded by supplementals around its short shelf life.
Federal / Roles
AO Authorizing Official
The senior official who signs the ATO and accepts the risk of running a system. The buck stops here. Win the AO early or your AI system won't ship.
AOR Area of Responsibility
The geographic or functional area a command or person owns. In contracting, it can also mean the scope a COR oversees. Know whose AOR your AI work sits in before you ask for a decision.
CDAO Chief Digital and Artificial Intelligence Officer
The DoD office that sets data and AI policy across the department and runs cross-service AI programs. Civilian agencies have Chief AI Officers under OMB M-24-10. If your work touches DoD-wide AI, CDAO is in the room.
CO Contracting Officer
The only person who can legally bind the government to a contract or modify one. Engineers and program managers can ask, but only the CO can sign. If a government employee promises you scope outside the contract, get the CO to confirm in writing.
COR Contracting Officer's Representative
The technical lead the CO delegates day-to-day contract oversight to. The COR signs off on deliverables and flags problems. They can't change the contract, but they can make your life easy or hard.
ISSO Information System Security Officer
The person responsible for the day-to-day security of a specific system. Owns the SSP, the POA&M, and the relationship with the assessor. Befriend your ISSO early in any AI build.
PMO Program Management Office
The government office that owns a program: scope, schedule, budget, contracts. Vendors deliver to the PMO. If you're a cleared engineer, the PMO is usually your customer's customer.
SCA Security Control Assessor
The independent assessor who tests a system's security controls and writes the report the AO uses to decide on an ATO. Treat SCA findings as the floor, not the ceiling. They see what your team stopped looking at.
Military / Service roles & ranks
MOS Military Occupational Specialty
An Army or Marine Corps job code. It identifies the work a person was trained to do, like infantry, intelligence analyst, or network administrator. Civilian equivalent: job title or job family code.
AFSC Air Force Specialty Code
An Air Force or Space Force job code. Five characters that name the career field and skill level. Civilian equivalent: job title or job family code.
Rate
The Navy and Coast Guard term for an enlisted job. Different from rank. Examples: Hospital Corpsman, Information Systems Technician, Cryptologic Technician. Civilian equivalent: job title.
NCO Noncommissioned Officer
An enlisted leader, typically at the sergeant or petty officer level. Runs day-to-day work and leads small teams. Civilian equivalent: front-line supervisor or team lead.
SNCO Senior Noncommissioned Officer
A senior enlisted leader, usually E-7 through E-9. Manages larger teams, runs programs, and advises officers. Civilian equivalent: senior manager or operations manager.
Warrant Officer
A technical specialist who sits between enlisted and officer ranks. Deep expertise in one trade, like helicopter pilot or signals intelligence. Civilian equivalent: principal engineer or senior technical specialist.
Chief Warrant Officer
A senior warrant officer (CW2 through CW5). Top technical authority in their field, often with 15-plus years of experience. Civilian equivalent: distinguished engineer or technical fellow.
Field-grade officer
A mid-to-senior commissioned officer at the major, lieutenant colonel, or colonel level (O-4 through O-6). Runs battalions, squadrons, or larger units. Civilian equivalent: director or VP.
Company-grade officer
A junior commissioned officer at the lieutenant or captain level (O-1 through O-3). Leads platoons and companies, typically 20 to 150 people. Civilian equivalent: line manager or first-line manager.
Flag and general officers
The most senior commissioned officers. Generals in the Army, Air Force, Marines, and Space Force. Admirals in the Navy and Coast Guard. O-7 and above. Civilian equivalent: C-suite executive.
Enlisted vs commissioned
Two parallel career tracks. Enlisted personnel do the hands-on work and supervise small teams. Commissioned officers hold a presidential commission and lead larger organizations. Most veterans are enlisted.
NCOIC Noncommissioned Officer in Charge
The senior enlisted person running a specific shop, shift, or section. Owns the daily work and the people doing it. Civilian equivalent: senior individual contributor or working-level team lead.
Military / Performance docs & records
OER Officer Evaluation Report
The Army's annual performance review for commissioned officers. Documents duties, results, and potential. Civilian equivalent: annual performance review.
NCOER Noncommissioned Officer Evaluation Report
The Army's annual performance review for sergeants. Lists duties, ratings, and bullet-point accomplishments. Civilian equivalent: annual performance review.
FITREP Fitness Report
The Navy and Marine Corps performance review for officers and senior enlisted. Ranks the member against peers in the same grade. Civilian equivalent: annual performance review with stack ranking.
EPR Enlisted Performance Report
The Air Force's annual performance review for enlisted members. Documents job performance and leadership. Civilian equivalent: annual performance review.
LES Leave and Earnings Statement
A monthly pay statement for service members. Shows base pay, allowances, deductions, and remaining leave days. Civilian equivalent: pay stub.
DD-214 Certificate of Release or Discharge from Active Duty
The official record of military service. Lists dates, awards, training, character of discharge, and reenlistment eligibility. Treat it as the source of truth when verifying a veteran's service.
DD-2870 Authorization for Disclosure of Medical or Dental Information
A signed release that lets the Department of Defense share a service member's medical or dental records. Used during VA claims and some background investigations.
SF-86 Questionnaire for National Security Positions
The federal form used to apply for a security clearance. Asks for residences, jobs, foreign contacts, finances, and other background details going back seven to ten years. Takes most people several days to complete.
Periodic reinvestigation
A scheduled re-vetting of a cleared person, historically every five or ten years depending on clearance level. Largely being replaced by continuous evaluation.
Continuous evaluation
An automated, ongoing review of cleared personnel. Checks credit, criminal records, and other data streams between formal reinvestigations. Now the standard model under Trusted Workforce 2.0.
Military / Transition programs
SkillBridge
A DoD program that lets service members work a civilian internship during their final 180 days of active duty. The military keeps paying them while the host company evaluates them at no cost. A common hiring pipeline for veterans.
VET TEC 2.0 Veteran Employment Through Technology Education Courses
A VA program that funds short-term technology training for veterans, including software development, data science, and cybersecurity bootcamps. Does not use GI Bill entitlement.
VR&E Veteran Readiness and Employment
A VA program (also called Chapter 31) that helps veterans with service-connected disabilities prepare for, find, and keep a job. Can cover tuition, certifications, and job-search support.
Post-9/11 GI Bill
The main education benefit for people who served after September 10, 2001. Pays tuition, fees, a housing allowance, and a book stipend for up to 36 months of school. Can sometimes be transferred to a spouse or child.
TAP Transition Assistance Program
The mandatory training service members complete before leaving the military. Covers resumes, benefits, and basic financial planning. Quality varies widely by installation.
COOL programs Credentialing Opportunities On-Line
Service-run programs that pay for civilian certifications related to a member's military job. Each branch runs its own version (Army COOL, Navy COOL, and so on).
MyCAA My Career Advancement Account
A DoD scholarship for spouses of junior service members. Pays up to $4,000 toward a license, certification, or associate degree in a portable career field.
Yellow Ribbon
A program where schools and the VA together cover tuition costs above the Post-9/11 GI Bill cap. Makes private universities and out-of-state tuition affordable for eligible veterans.
Military / Clearance terms
TS/SCI Top Secret / Sensitive Compartmented Information
The highest standard clearance level for most cleared jobs. Top Secret covers classified national security information. SCI adds access to specific intelligence programs after extra read-ins.
Active clearance
A clearance currently in use at a cleared job. The person is enrolled in continuous evaluation and can access classified material at their level.
Current clearance
A clearance that has been used within the past 24 months but is not active right now. Usually faster and cheaper to reactivate than starting a new investigation.
Inactive clearance
A clearance the person held but has not used in more than 24 months. Often requires a new investigation to reinstate, depending on the gap and the agency.
CI Poly Counterintelligence Polygraph
A polygraph exam focused on espionage, sabotage, terrorism, and unauthorized contact with foreign entities. Common for DoD and intelligence community roles above Top Secret.
FS Poly Full Scope Polygraph
A polygraph covering both counterintelligence topics and personal conduct like drug use, theft, and serious crimes. Required for many NSA and CIA positions. Also called Lifestyle Polygraph.
JTF Joint Task Force
A temporary team built from multiple services or agencies to handle a specific mission. Common phrasing on cleared resumes (for example, JTF-Bravo). Civilian equivalent: cross-functional task force.
CONUS Continental United States
The 48 contiguous states plus DC. Used in cleared job postings to say the work or candidate must be inside the lower 48.
OCONUS Outside the Continental United States
Anywhere outside the lower 48, including Alaska, Hawaii, and overseas. Cleared OCONUS roles often pay more and include hardship or hazard pay.
Military / Universal military analogies that travel
Basic training
The initial 7 to 13 weeks of training every enlisted member completes. Teaches military basics, fitness, and discipline. Civilian equivalent: a mandatory company-wide onboarding bootcamp.
Deployment
An extended assignment away from home station, often overseas. Typically 6 to 12 months. Civilian equivalent: a long-term project rotation in another city or country.
Chain of command
The line of supervisors from a service member up to the President. Decisions and information move along this line. Civilian equivalent: reporting structure or org chart.
Brief-back
When a subordinate restates the plan to confirm they understood the orders. Catches misunderstandings before execution. Civilian equivalent: a confirmation readback at the end of a meeting.
AAR After-Action Review
A structured debrief after a mission or exercise. Covers what was planned, what happened, what went well, and what to fix. Civilian equivalent: blameless post-mortem or retrospective.
TDY Temporary Duty
A short trip away from home station for training, a meeting, or a short mission. Usually days to a few months. Civilian equivalent: business travel.
No terms match your search. Try a different word, or clear filters.