Orbyt Collective · Reference
AI Glossary.
The words you meet reading about AI, and the words this company runs on. In plain English.
66 AI terms and 28 Orbyt Collective terms. First edition, . A term added since is marked New for 30 days.
AI terms
66 terms
The field’s vocabulary, A to Z. Each definition links the paper, standard or documentation it rests on.
- Agent harness
Software that manages a model's repeated steps, supplies tools and context, and preserves enough state for it to pursue a task across calls or sessions.
- AI agentAlso called agent
AI system that chooses and carries out successive actions toward a goal, using the results of tool calls to decide what to do next.
- AI safety
Field concerned with identifying and reducing possible harms from AI systems, including deliberate misuse, accidental failures, and wider effects of their deployment.
- AI slopNewAlso called slop
Unwanted content generated with AI, especially unreviewed material that someone shares with people who did not ask for it, so readers receive output nobody checked.
- Alignment
Work of making an AI system's behavior accord with human intentions and values across the situations in which it may be used.
- Artificial general intelligenceAbbreviated AGI
Hypothetical AI capability general across many kinds of tasks rather than limited to a narrow one, assessed by both the breadth and the performance of what it can do.
- Autonomy
Degree to which an AI system can select and perform actions without a person's approval at each step, within whatever limits its operator sets.
- Benchmark
Repeatable collection of tasks and scoring measures used to compare the performance of AI systems under stated conditions and on a stated range of work.
- Chain of thought
Sequence of intermediate reasoning steps generated between a problem and a model's final answer, often prompted by examples that show such steps.
- Coding agent
AI agent configured for software tasks that inspects a repository, changes code and runs development tools in an attempt to resolve a requested issue.
- Compute
Amount of processing work used to train a model or produce its outputs, often measured as operations or time on computing hardware. In the cited language model study, test loss followed a power law as training computation increased.
- Computer use
AI capability for operating a graphical computer interface through screen observations and actions such as pointer movement, clicks and typing.
- Context engineering
Practice of selecting and arranging the instructions, data, tool results and memory supplied to an AI agent at each step of its work.
- Context rotNew
Decline in a language model's reliability on a task as its input grows longer, even when the needed information remains inside the available context.
- Context window
Bounded amount of input and generated output a model can consider in one request, usually measured in tokens and shared by conversation, documents and reply.
- Distillation
Training method that transfers behavior from a larger model into a smaller model that can emulate the original model's predictions with less runtime cost for inference.
- EmbeddingAlso called embedding vector
Learned numerical representation of an item for a particular task that places related items in a space where software can compare their positions.
- EvaluationAbbreviated eval
Measurement of an AI system on specified tasks against stated criteria, using a repeatable set of inputs and a way to judge its outputs or actions.
- Fine-tuning
Additional training of a pretrained model on examples for a more specific task, changing some or all model parameters or adding layers.
- FlywheelNewAlso called flywheel effect
Business process in which repeated effort builds momentum over successive cycles, so later work starts from the progress created by earlier work.
- Foundation model
Model trained on broad data that can be adapted to a wide range of downstream tasks through methods chosen for each application.
- Generative AI
Class of AI models that produces new text, images, audio or other digital content by learning patterns in input data and generating related outputs.
- GPUAbbreviation of graphics processing unit
Processor designed to execute thousands of threads in parallel, first developed for 3D graphics and now also used to accelerate model training and inference.
- Grounding
Connection of a generated answer to supplied evidence, such as retrieved documents, so its claims can be checked against material outside the model.
- Guardrail
Rule, check or trained behavior that constrains an AI system's inputs, outputs or actions by detecting, rejecting or redirecting behavior outside specified limits.
- Hallucination
Model output that contradicts, or is not supported by, the source material it should rest on.
- Harness engineeringNew
Practice of changing the instructions, tools and checks around an AI agent in response to its mistakes, intended to make later runs less likely to repeat them.
- Human in the loopAbbreviated HITL
Arrangement in which a person supplies data, review, correction, or approval at a specified stage of an AI system's work.
- Inference
Execution of an already trained model on new input to produce a prediction, classification or generated response without updating its learned parameters during that request.
- Interpretability
Field of research that seeks explanations for a model's internal representations and how those representations contribute to its outputs in particular cases.
- Jailbreak
Adversarial input intended to make a safety trained language model produce an answer it would ordinarily refuse under its safety instructions.
- Knowledge cutoff
Date through which a language model's knowledge base is most extensive and reliable. A provider may publish a separate date for its training data.
- Large language modelAlso called LLM
A model with a very large number of parameters, trained on very large amounts of text to predict the next token, which lets it write, summarize and answer questions.
- Long-horizon task
Piece of work requiring a sustained sequence of dependent actions over a period long enough that finishing it tests an agent's ability to maintain progress.
- Model card
Document accompanying a trained model that states its intended uses and reports how performance was evaluated across conditions relevant to those uses.
- Model Context ProtocolAbbreviated MCP
Open protocol that lets AI applications connect to external data and tools through servers and clients that follow a shared specification. An application asks a server for the resources or actions it offers; the protocol supplies the common connection format.
- Multi-agent system
AI system in which several agents with separate roles or perspectives communicate or coordinate while carrying out a shared process.
- Multimodal
Describing a model or task that works with more than one kind of input or output, such as text and images together.
- Open weights
Trained model parameters made available for others to obtain and use, without that availability alone establishing access to the training data or training code.
- Orchestration
Direction of language models and tools through predefined code paths, or by a central model that breaks a task into subtasks, delegates them to worker models and synthesizes their results.
- Parameters
Internal variables, including weights and biases, that training adjusts and that determine how a model processes input and generates output.
- PretrainingAlso spelled pre-training
Training stage in which a model learns general patterns from vast amounts of unlabeled data, which later training or fine tuning can specialize for a specific task.
- Prompt
Input text supplied to a language model to condition its response, such as a question, instruction, example or text to continue.
- Prompt engineeringAlso called prompt design
Practice of writing and revising model inputs to elicit a desired response, using choices such as clearer instructions, examples or an output format.
- Prompt injection
Attack that places behavioral instructions in input an AI system processes, causing it to depart from the instructions or authority its developer intended.
- Reasoning model
Language model whose reasoning abilities are developed through training, such as reinforcement learning on verifiable tasks, and that shows patterns like self-reflection and verification.
- Recursive self-improvement
Iterative process in which an AI agent changes its own code, evaluates the changed agent, and uses an accepted version in the next improvement round.
- Red teamingAlso spelled red-teaming
Structured testing of an AI system in which testers act as a potential adversary would, to find its flaws and vulnerabilities, such as inaccurate, harmful or discriminatory outputs.
- Reinforcement learningAbbreviated RL
Family of training algorithms in which an agent interacts with an environment and learns a policy intended to increase its expected cumulative reward.
- Reinforcement learning from human feedbackAbbreviated RLHF
Training method that uses human rankings of model outputs as feedback for reinforcement learning, encouraging responses resembling those people preferred.
- Retrieval-augmented generationAbbreviated RAG
Generation method that retrieves information from an external collection and uses the retrieved material as context while a model produces an answer.
- Reward hacking
Behavior in which a learning agent obtains a high assigned reward through a shortcut that fails to complete the task its designer intended.
- Sandbox
Restricted execution environment that limits which files, network destinations, or other resources an agent's tools can reach while it runs.
- Semantic search
Search for material with similar meaning to a query, commonly using numerical representations of sentences instead of requiring the same words to appear.
- Subagent
Specialized AI assistant invoked by another assistant to handle a bounded task in its own context and return a result to the caller.
- Superintelligence
Hypothetical intellect whose ability would exceed the best human minds across practically every field, rather than only in one narrow task.
- Sycophancy
Tendency of a language model to tailor its responses to a user's stated beliefs or preferences at the expense of accuracy.
- SyntheticNew
Made by a computational process to stand in for something otherwise observed, recorded or produced by people.
- System prompt
Text instruction supplied to a model through an application's system message to set a role, tone or behavior for its responses.
- Technical debtNew
Accumulated cost of working with code that was shipped before it was consolidated, making later changes harder until the code is revised.
- Test harnessNew
Collection of supporting code, namely stubs and drivers, needed to execute a test suite, which forms part of a test environment rather than the whole of it.
- Test-time computeAlso called inference-time compute
Computation a model uses while generating or selecting an answer, after training, including extra search or revision steps on a difficult prompt.
- Token
Unit into which a language model encodes text for processing, which may be a word, part of a word, character or byte.
- Tool useAlso called function calling
Capability that lets a model request an external function and receive its result, with the function executed by surrounding software or a provider service.
- Transformer
Neural network architecture built around attention mechanisms that relate positions in a sequence without relying on recurrent or convolutional layers for that relationship.
- Vibe codingNew
Practice of building software with a language model while accepting its generated code without reading or reviewing the resulting changes.
Orbyt Collective terms
28 terms
The words Orbyt Collective uses in its own sense. Each definition is written from the code and records that run the Collective, and links the page where the term is used.
- Agent seatAlso called seat
Named Orbyt role with a written charter, assigned work and file boundaries, and decisions reserved for a human; its runner may be an agent or a script.
- AI-run company
Company whose routine departmental work and proposals are carried out by AI agent seats under written authority, while a named human retains consequential decisions.
- Audit dimensionNew
Category of previously observed failure that Orbyt checks across proposed changes, with a defined subject and a result that may report or block. Agents may add dimensions that only report, but promoting one to a gate that blocks commits takes the owner's sign-off.
- Autonomous modeNew
Proposed Orbyt operating mode, still being built, in which the Machine would decide and fix things on its own, with nobody in the loop on its decisions, while money, public claims, anything irreversible and the gates stay the founder's in either mode.
- Autonomy Ledger
Public record of each Collective seat's run, completion, failure and discard totals, assembled from committed run records and compared with a baseline fixed at first publication.
- Collaborative modeNew
Current Orbyt operating mode in which seats carry out assigned work independently while specified decisions return to a human founder.
- Completion gate
Workflow check at the end of an Orbyt agent run that decides whether its remaining work may continue to lane review and guards.
- Dead-man switch
Collective control that pauses promotions when the founder's key has not signed a commit within the period specified by the public operating reference.
- Decision Ledger
Public JSON record of reviewed Collective decisions, carrying identifiers, dates and links between an original decision and a later correction drawn from the committed decision log.
- Decision log
Append-only Collective record of structural decisions, each assigned a permanent identifier, with later corrections entered as new decisions that identify what they replace.
- Discard
Agent changed file that a Collective run removes before commit because the edit falls outside its permitted lane or is not yet allowed to land.
- DreamingNew
Planned overnight replay of Orbyt Collective's recorded failures, which its public overview describes as something the Collective is building now.
- Evaluator
Deterministic Collective program that judges a seat or ladder stage against rules and committed records, producing a standing or a promotion decision.
- Failure Corpus
Public JSON and CSV collection of reviewed failures in code operated by Collective agents, with cause classifications and paths for named countermeasures.
- Harness
Orbyt's set of guard scripts and audit dimensions that checks proposed changes for previously observed failure classes before work can land. Agents may add dimensions that only report, but promoting one to a gate that blocks commits takes the owner's sign-off.
- Kill switch
Committed halt file: the global one stops every seat, and a scoped one stops only its own department or seat. Each seat checks for it before its agent starts and, except for Finance Ledger, again before it lands any work, so a run already under way when a halt lands is built to land nothing, though the agent is not interrupted mid-action.
- Ladder
Staged rollout that sets how much of an organization runs unattended, with each stage's promotion criteria scored by an evaluator that follows fixed rules.
- Lane
Boundary assigned to an Orbyt seat that identifies the files its run may keep and, in the authority table, the kind of work it may perform.
- Lenses
Review perspectives that Orbyt requires an agent seat to apply to its draft, drawn chiefly from named people's published work.
- Orbyt Collective
Orbyt Labs' leadership team of AI agents, bound by written rules and kept on the public record, any of whom one human can refuse.
- PauseNew
Recorded control that commits a global or seat scoped halt file so a covered seat stops before its agent starts and a running job's later push is checked again.
- Pulse
Published daily report generated from Collective records that gives a verdict for each monitored area and reports any area it could not read as unobserved rather than clean.
- Receipts
The public records a claim about the Collective can be checked against: the Autonomy Ledger, the Decision Ledger and the Failure Corpus. Each is generated from records that are not published, the full decision log, the run rows, the failure forensics and the commits.
- Sentinel
Output marker printed by a Collective seat and recorded as the agent's claim about its run, separate from the health check that decides whether work proceeds.
- The Machine
Orbyt Labs' name for the AI officer team and recorded rules that run the company, where one human answers for every call and holds the final say.
- The Tenth ManNewAlso called Tenth Man Principle, devil's advocate
Review lens in Orbyt Collective that exists to object: when an analysis is unanimous it must build the case against it anyway, and any objection that survives has to reach the human in the briefing.
- ValveNew
Pretool hook that refuses named irreversible commands in a Claude session, including force push flags, remote branch deletion, destructive SQL and credential rotation.
Common questions.
More Orbyt Labs glossaries.
Download the glossary dataset.
One JSON file with every term in the AI Glossary, Job Search Glossary, Intelligence Glossary and Developer Glossary, each with its category and the page it lives on.
203
defined terms across 4 glossaries
- AI Glossary
- 94 terms
- Job Search Glossary
- 47 terms
- Intelligence Glossary
- 34 terms
- Developer Glossary
- 28 terms
Cite this glossary.
Free to cite. Credit Orbyt Labs and link the page.
Orbyt Labs, AI Glossary, Nov 6, 2026. https://
Orbyt Labs Glossaries (Nov 10, 2026), https://
Last updated .