Plain-language definitions for the terms you keep running into.

An append only file at the root of a project folder where every working session ends with a one entry summary: what was done, decisions made, files changed, next step, known issues.

Our take. We never start a session without reading this file. It is the single most reliable session handoff document in any AI assisted workflow.

An AI system that can take actions, not just answer questions. An agent reads input, makes decisions, and triggers tool calls or follow up steps without waiting for a human at every turn.

Our take. Most teams reach for agent frameworks too early. A folder structure plus one well configured agent does the work of a multi agent stack with less complexity.

Application Programming Interface. The contract that lets one piece of software talk to another. In AI work, an API key authenticates your application against a model provider so you can send prompts and receive completions.

The 01 stage in our folder system. Where the system is designed on paper before any code gets written. Inputs come from discovery. Outputs are a written specification, an architecture diagram, a compliance map, and a fixed price quote.

Business Associate Agreement. A contract required under HIPAA whenever a covered entity shares protected health information with a vendor. The BAA spells out how the vendor will protect the data and what happens if it is breached.

Our take. If a vendor will not sign a BAA, you cannot use them for anything that touches PHI. Period.

A ByteWorthy product term for a small reusable system you can buy outright, fork, and run yourself. Examples: Sovra, Klienta, Clynova. The blueprint is the foundation. Customization sits on top.

The 02 stage in our folder system. Where the architecture spec becomes working software. We ship a working increment every week and demo every Friday. Feedback gets incorporated within the same sprint.

The locked configuration folder in our folder system. Holds brand, voice, integrations, credentials, and anything else that must stay fixed for the duration of the project. Read only during execution.

Our take. The single most common cause of inconsistent AI assisted work is configuration drift. Lock _config at kickoff. Any mid project change requires explicit approval and a versioned commit.

Claim Adjustment Reason Code. The code a payer attaches to a claim adjustment to explain why they paid less than billed. CARCs are paired with RARCs to give the full reason for a denial or adjustment.

A large language model hosted by a provider like Anthropic, OpenAI, or Google. Your application sends prompts over the internet and receives completions. Convenient, fast, and the right choice for many use cases.

Our take. If your data is sensitive, a cloud LLM may not be the right call regardless of cost. Read the local LLM entry next.

Centers for Medicare and Medicaid Services. The federal agency that runs Medicare and Medicaid in the United States. CMS publishes coding rules, fee schedules, and policy changes that affect every healthcare practice.

The maximum amount of text a language model can read in a single turn. Measured in tokens. When you exceed the window, older context gets dropped and the model loses awareness of it.

Our take. Most teams overload the context window with whole project history when each stage needs only its own context. That is a system design failure, not a model limitation.

The 03 stage in our folder system. Where the working build moves to production. We migrate from staging, run smoke tests, configure monitoring, train the operating team, and stand by for 48 hours post launch.

The 00 stage in our folder system. Where the problem gets defined before any solution gets designed. We map the current process, identify what is automatable, and produce a written assessment.

Electronic Health Record. The digital system a healthcare practice uses to manage patient charts, encounters, orders, and billing. Common examples: Epic, athenahealth, Cerner, eClinicalWorks.

A numeric representation of text that captures its meaning. Used by vector databases to find semantically similar content. The phrase 'cardiac arrest' and 'heart attack' produce embeddings close together because they mean similar things.

Data encrypted when stored on disk, so anyone who steals the disk cannot read the data without the key. A HIPAA technical safeguard requirement for any system that stores PHI.

Electronic Protected Health Information. PHI in digital form. Subject to the same HIPAA protections as paper PHI plus the technical safeguard rules that apply to electronic systems.

Fast Healthcare Interoperability Resources. A standard for exchanging healthcare data between systems. Most modern EHRs expose a FHIR API for reading and writing patient data.

Continuing the training of a pre trained language model on your own data, so it learns your domain, voice, or style. Different from RAG, which retrieves relevant documents at query time without changing the model.

Our take. Most teams who think they need fine tuning actually need RAG. Try retrieval first. Fine tune only when retrieval is not enough.

Our methodology. A numbered folder structure (00 discovery through 04 operate plus _config) that holds every artifact, decision, and handoff for a project. The folder is the contract. The structure is the routing logic. The output is the promise.

Our take. Read /the-folder-system for the full argument. Free starter kit included.

When an AI model generates plausible sounding text that is factually wrong. Common in language models that lack access to ground truth data for the question being asked.

Our take. RAG with cited sources is the most reliable way to reduce hallucinations in production AI applications.

Health Insurance Portability and Accountability Act. The US federal law that governs how healthcare organizations protect patient data. Sets requirements for privacy, security, and breach notification.

The Health Information Technology for Economic and Clinical Health Act. A 2009 law that strengthened HIPAA, added breach notification rules, and increased penalties for violations.

Health Level Seven. A family of healthcare data exchange standards that predate FHIR. Many legacy EHR integrations still use HL7 v2 messages.

Running a trained model to generate output. When you send a prompt to a language model and get a completion back, that is inference. Different from training, which builds the model in the first place.

Large Language Model. A neural network trained on massive amounts of text that can generate, summarize, translate, and reason about language. GPT, Claude, Gemini, Llama, and Mistral are all LLMs.

A large language model that runs on hardware you control. The model weights live on your server or workstation. Your data never leaves your network. Examples: Llama 3, Mistral, Phi.

Our take. For HIPAA, attorney privilege, and other high sensitivity workflows, a local LLM is often the only path that satisfies compliance without a vendor BAA.

A document that describes a machine learning model: what it was trained on, how it performs on benchmarks, what its limitations are, and what it should not be used for. Released alongside the model itself.

Medical Record Number. A unique identifier a healthcare practice assigns to each patient. Used to look up that patient's chart across visits and providers within the same organization.

National Provider Identifier. A 10 digit number CMS issues to every healthcare provider in the United States. Required on every claim and used to look up provider credentials.

National Plan and Provider Enumeration System. The CMS database that holds every NPI and the provider details associated with it. Free public API. Useful for verifying providers and credentials.

The 04 stage in our folder system. Where the live system gets monitored, patched, and improved on a rolling basis. Monthly check ins. Quarterly reviews. The retainer never auto renews silently.

A ByteWorthy product term for a small focused automation that solves one specific workflow. Smaller in scope than a Blueprint. Faster to deploy. Examples: prior auth pack, denial appeal pack.

Protected Health Information. Any health information about an identifiable patient. The HIPAA Privacy Rule lists 18 identifiers that turn ordinary health information into PHI. Names, dates, addresses, account numbers, photos.

Prior Authorization. A process where a healthcare provider must get approval from a payer before delivering a service or prescription, or the payer will not cover it. Notorious source of administrative burden and care delays.

The input you give to a language model. Includes the system message (what role the model is playing) plus the user message (the question or task). Quality of the prompt heavily influences quality of the output.

An attack where untrusted content (a user message, a web page, a document) contains hidden instructions that try to override the application's intended behavior. A serious security concern for any AI system that processes external content.

Our take. Treat every external input as potentially adversarial. Never let an AI make irreversible decisions based on content you do not control.

Retrieval Augmented Generation. A pattern where the application retrieves relevant documents from a vector database at query time and includes them in the prompt, so the model can answer based on your data instead of just its training.

Our take. RAG is the default for most enterprise AI use cases. Cheaper, faster, and more controllable than fine tuning.

Remittance Advice Remark Code. The code a payer adds alongside a CARC to provide more specific information about a claim adjustment. CARC says why; RARC says what specifically.

The property of a project being pickable up by a new agent or human without context from the prior session. The folder explains itself, the activity log explains the state, and the stage CONTEXT files explain what is happening.

Our take. If a project requires a brain dump from the founder to continue, the system has failed. Resumability is the test of any folder system.

A monthly engagement where a vendor keeps watching a system, applying patches, responding to incidents, and shipping small improvements. Different from a fixed price build.

When a project's requirements expand beyond what was originally agreed without a corresponding change to budget or timeline. The most common cause of failed projects.

Our take. The fix is not better contracts. The fix is explicit human review gates between stages so changes get acknowledged and priced.

ByteWorthy's open source MIT licensed Blueprint. Next.js plus Go plus Supabase. The foundation every ByteWorthy build starts from. Free to fork. Free to ship.

A written specification of what a system does. The output of the architecture stage. Includes data model, integrations, screens, automation rules, and compliance surface. Must be specific enough that two builders would produce the same system from it.

The instructions you give a language model to set its role, voice, and constraints before any user message arrives. The system prompt typically stays fixed across a session while user messages change.

The unit a language model uses to count input and output. Roughly 4 characters of English text, or 0.75 of a word. Models charge per token and have a maximum context window measured in tokens.

The process of building a machine learning model from data. Different from inference (running an existing model). Training a large language model from scratch costs millions; fine tuning an existing one is much cheaper.

A database optimized for storing and searching embeddings. The backbone of RAG. Examples: Qdrant, Weaviate, Pinecone, pgvector. You insert documents as embeddings, then query for nearest neighbors at runtime.