Glossary
The glossary explains the words that Auga uses, and the words that it avoids.
It is for internal use while the words settle. Choose a group to filter the list.
Words
| Term | Plain meaning | More exactly | Do not use | Seen on |
|---|---|---|---|---|
| Agent registerPlaces | The list of every agent, who holds each role for it now, and its history. | The menu calls it Agents. Only register sources create agents. Uploaded files and bundles never do. | inventory | Agent register |
| Agent HealthPlaces | A report on one time window of runs in one workflow: what changed, how sure we are, and what it may cost. | The report grades seven dimensions. The total status is the weakest dimension, read at the low end of its likely range. It covers the whole workflow, so each agent's profile shows its own part under Health. | Agents Health, Health, when the whole report is meant | Agent HealthLoan orchestrator |
| Agent profilePlaces | The page for one agent: how it works, who holds each role, how it is doing, where Auga sees it, and its history. | Its portrait shows real runs of the agent's workflow in 3D. Its Health section shows the agent's part of the latest Agent Health report, and leads to the report. | – | Agent registerLoan orchestrator |
| Bundle AnalysisPlaces | A place to practise with Auga on a simulated dataset where the root cause is hidden. | It appears only when the uploaded file is a bundle. The answer stays sealed until you open it. | Proving Ground | credit-nov |
| CoachPlaces | The place where a person reviews the decisions of agents. | The wide menu says Coach your agents. The page title is Decision review. Coach is a working name. | Review | Decision review |
| CompliancePlaces | The place that shows the evidence for each article of a law. | The first law is the EU AI Act. The page title keeps the name EU AI Act, because several laws will follow. | AI Act, as a menu name | EU AI Act |
| ConnectionsPlaces | How Auga is connected, how you can use it, and which systems receive evidence from Auga. | – | – | Connections |
| Data boundaryPlaces | What left your network, and what stayed inside it. | The menu says Boundary. | – | Data boundary |
| EvidencePlaces | All runs, the findings that explain them, and the cause chains. | – | – | Evidence: all runsEvidence: graphEvidence: one run at a time |
| GlossaryPlaces | This page. It explains the words that Auga uses and the words it avoids. | The source is glossary.json. It is for internal use while the words settle. | – | Glossary |
| ImprovePlaces | The changes to your agents, whether each change helped, and what Auga proposes to change next. | – | – | Improve |
| IncidentsPlaces | What has gone wrong with your agents, how sure Auga is, and who handles each incident. | – | – | IncidentsINC-1: Denials rose after bureau feed 3.3 |
| Intelligence ExchangePlaces | Sharing what your agents learned with other companies, without sharing the data. | The menu says Exchange. Only graded findings are shared, never rows of data. | – | Intelligence ExchangeIntelligence Exchange, sharing findings |
| LedgerPlaces | The signed record of decisions, reviews and actions. Each entry is linked to the one before, so any change can be seen. | Hash-chained and signed with the guard's key. It is not a register of agents. | journal | The ledger |
| MandatesPlaces | What each agent may do in each field of work, and what it needs for the next stage. | – | – | Mandates |
| Preview a stagePlaces | A view of Mandates that shows what a stage would have done to real traffic, before an agent gets that stage. It sets nothing. | Earlier called Mandate Simulator, which is also the name of the separate prototype it comes from. | Simulator | Preview a stage |
| ReportsPlaces | Copies of the evidence that you can download, and the review of findings. | The wide menu says Claim review & reports. | – | Claim review & reports |
| RisksPlaces | What could go wrong with an agent again, what it would cost, and what your company decided. | – | – | RisksRISK-1: The bureau feed can again make scores stale |
| What-IfPlaces | A calculator for the value and the losses of all your agents, with a report for the board. | – | – | What-IfBoard report |
| ObservedStages and grades | The agent only watches. Its proposals are logged, and nothing runs. | The first of five stages. | – | MandatesAgent register |
| SupervisedStages and grades | The agent acts, and a person reviews every action. | – | – | MandatesAgent register |
| ConstrainedStages and grades | The agent acts within limits. Actions that cannot be undone go to a person. | – | – | MandatesAgent register |
| TrustedStages and grades | The agent acts within limits, also for actions that cannot be undone. | – | – | Mandates |
| MandatedStages and grades | The agent acts on its own, because an experiment showed that it works. | The dashboard says proven by experiment. The word proven is listed for the team to decide. | – | Mandates |
| Field of workStages and grades | One area where an agent can act. Each agent has a stage in each field of work. | Tool use, Data handling, External communication, Payments, Escalation and Output accuracy. | field, when it is unclear | MandatesAgent register |
| GradeStages and grades | How strong the evidence for a finding is, from F to A+. | F: seen by one method. D: two methods agree. C: effect measured. B: survived the challenge tests. A: proven by experiment. A+: predicted and held. | grade, for a status | Evidence: graphClaim review & reports |
| StatusStages and grades | HEALTHY, WATCH or UNHEALTHY for a part of Agent Health. A status is not a grade. | – | – | Agent Health |
| UpgradedStages and grades | The stage of an agent went up in a field of work. | An agent is upgraded one stage at a time, when every criterion of the next stage is met. The word comes from credit ratings. Sources: Fitch Ratings: Rating Definitions, Moody's Rating Terms. | raised, moved up, promotion | Agent registerMandates |
| DowngradedStages and grades | The stage of an agent went down in a field of work. | An agent is downgraded at once when a criterion is not met. The word comes from credit ratings. Sources: Fitch Ratings: Rating Definitions, Levels of Autonomy for AI Agents. | lowered, moved down, demotion | Agent registerMandatesEU AI Act |
| UnchangedStages and grades | The stage stayed the same in the period. | Not affirmed: affirmed would say that a review took place. | – | Agent register |
| Held lowerStages and grades | A person keeps the stage below what the evidence allows. | People may hold a stage down, but never raise it above the evidence. | – | Agent registerMandates |
| Moved, for the values of a signalStatistics | How much the values moved compared with the first period. | PSI and KS for numbers, Jensen–Shannon for categories. | – | Agent Health |
| The moment the pattern changedStatistics | The time when the data started to behave differently. | Change point. | – | Agent Health |
| How strong the sign isStatistics | Weak, clear or strong. | Gain in nats at the change point. | – | Agent Health |
| Likely rangeStatistics | The range where the true value probably is. | A 95% interval, often a Wilson interval. | 95% sure | Agent HealthImprove IMP-2 |
| Far from the normal valuesStatistics | A value that is very different from the usual values. | Robust z above 10. | – | Agent Health |
| How much the outcome depends on a signalStatistics | How much knowing the signal tells you about the outcome. It does not show a cause. | Mutual information. | causes | Agent Health |
| Clearly differentStatistics | The likely ranges of two groups do not overlap. | Separated. | – | Improve IMP-2Agent Health |
| We cannot tell the groups apartStatistics | The likely ranges of two groups overlap. | Not separated. | no difference | Improve IMP-1Improve IMP-3 |
| Runs that stayed on the old versionStatistics | Runs in the same hours that did not get the change. They show what would have happened without it. | Concurrent control. | – | Improve IMP-2Agent Health |
| Some periods laterStatistics | A link where one thing follows another after a number of periods. | Lag n. | lag | Agent Healthcredit-nov |
| 95% were fasterStatistics | 95% of runs or steps were faster than this time. | p95. Half were faster than p50. | – | Agent HealthA run in credit-nov |
| 1 year in 10 is worseStatistics | In the simulation, 1 year in 10 is worse than this figure. | P10. | worst case | What-IfBoard report |
| Fair shareStatistics | Each agent's share of a total, when all the agents worked together. | Shapley value. | cause | Board report |
| PeriodStatistics | A slice of time that the data is divided into. | Bucket. | – | Agent Health |
| Percentage pointsStatistics | The difference between two percentages, for example from 16% to 100% is 84 points. | pp. | – | Agent HealthImprove IMP-2 |
| Difference: none, small, clear, largeStatistics | How big a difference is, in four steps. | Use these words only with their own thresholds, never for other measures. | effect size, d | Evidence: graph |
| Accountable personGovernance and law | The person who answers for what an agent does, and who decides about its risks and incidents. | – | responsible, when accountable is meant | Agent registerLoan orchestrator |
| Technical ownerGovernance and law | The person who keeps an agent running and makes the changes. | – | – | Agent registerLoan orchestrator |
| Human oversightGovernance and law | The people who review the decisions that the policy sends to a person. | EU AI Act, Article 14. | – | Agent registerEU AI ActDecision review |
| DeployerGovernance and law | The company that uses an AI system. The EU AI Act sets duties for it. | Compliance is the deployer's conclusion. Auga shows evidence, not a verdict. | – | Loan orchestratorEU AI Act |
| Model providerGovernance and law | The company that makes the AI model that an agent uses. | Not the same as provider in the EU AI Act, which means the maker of the whole AI system. | – | Loan orchestratorConnections |
| High-risk AI systemGovernance and law | An AI system where mistakes can seriously affect people, as defined in the EU AI Act. | Annex III lists the areas. Point 5(b) covers the creditworthiness of people. Your company classifies the system. | – | Loan orchestratorPlans |
| Serious incidentGovernance and law | An incident with serious harm, as defined in the EU AI Act. | Whether an incident is serious is your company's decision. Auga records who decided, and when. | – | IncidentsINC-1: Denials rose after bureau feed 3.3 |
| Incident levelsGovernance and law | Something changed, Harm suspected, Harm confirmed. | Something changed: the data moved. Harm suspected: a graded finding links the change to a harm. Harm confirmed: a person or the real outcome confirms it. | – | IncidentsINC-1: Denials rose after bureau feed 3.3 |
| Risk acceptanceGovernance and law | A signed decision by a person to accept a risk for an agent and a field of work, until a date. | An accepted risk never raises the stage of an agent. | – | MandatesRisks |
| AttestationGovernance and law | A signed statement of the stage that an agent has, which other systems can check. | – | in force | MandatesEU AI Act |
| Evidence packageGovernance and law | The signed evidence about an incident or a risk, which a receiver can check without Auga. | – | – | INC-1: Denials rose after bureau feed 3.3RISK-1: The bureau feed can again make scores staleConnections |
| GRC systemGovernance and law | GRC is short for governance, risk and compliance. A GRC system is the software a company uses to manage these. | Examples are ServiceNow, Archer, AuditBoard, LogicGate Risk Cloud and OneTrust. Auga has not been tried with any of them. | – | Connections |
| Register sourceGovernance and law | An OpenTelemetry Collector that your organisation uses in production or staging. Only these create agents in the register. | – | – | Agent registerLoan orchestrator |
| Analysis dataGovernance and law | A file or bundle that someone uploads. It never creates an agent in the register. | – | – | Agent registercredit-nov |
| Declared, seenGovernance and law | Declared means that a person registered the agent. Seen means that Auga sees it in a register source. | Seen, not declared can be an agent that nobody governs. | – | Agent register |
| CompliantGovernance and law | Auga never uses this word. Compliance is the deployer's conclusion. | – | compliant | EU AI ActClaim review & reports |
| RunEveryday words | One execution of an agent's work, for example one loan application. | The tab Traces & Spans and technical attribute names keep the word trace. | trace, case, episode | Evidence: all runstraces-nov |
| StepEveryday words | One part of a run, for example one tool call. | Span, in technical views. | – | A run in credit-nov |
| Failed runEveryday words | A run with the outcome that you count as bad, for example a denial. | A denial counts as a failed run, but a denial can be correct. | bad, bad outcome, bad-rate | Agent HealthINC-1: Denials rose after bureau feed 3.3 |
| FindingEveryday words | Something that Auga found in the data, with a grade for how strong the evidence is. | – | claim, edge, hypothesis | Evidence: graphClaim review & reports |
| SignalEveryday words | A value that Auga records for each run, such as income. | – | variable, column, attribute | Agent HealthEvidence: all runs |
| LinkEveryday words | A connection between two signals in a graph. | Auga edge is the name of a product part, not a link. | edge, for a connection in a graph | Evidence: graphcredit-nov |
| Cause chainEveryday words | The steps from a cause to an outcome, each with its finding. | – | chain | INC-1: Denials rose after bureau feed 3.3RISK-1: The bureau feed can again make scores stale |
| ChangeEveryday words | A difference between two versions, for example of a model or a feed. | – | switch | ImproveAgent Health |
| RolloutEveryday words | A controlled release of a change to more and more runs. | – | – | ImproveImprove IMP-5 |
| LossEveryday words | Money lost because of something that went wrong. | – | damage, lost | What-IfAgent Health |
| VerdictEveryday words | A person's decision on one agent decision: correct, wrong or harmful. | – | verdict, for a report or a change | Decision reviewThe ledger |
| ResultEveryday words | What a report, a change or a check came to. | – | – | Agent HealthImprove |
| StaleEveryday words | Out of date. A stale score was not updated when it should have been. | Listed for the team to decide, because some readers may not know the word. | – | INC-1: Denials rose after bureau feed 3.3RISK-1: The bureau feed can again make scores stale |
The source of this page is glossary.json. Change a word there, then run node scripts/glossary.mjs.