Glossary
Each term this framework uses has one definition here. The list under From the data file is generated from framework.json, so those definitions cannot drift from the ones the tools and the pipeline use.
Terms
AgentA language model that runs tools on its own to finish a task, such as querying logs, writing a rule, or opening a pull request.
Agent runnerThe program that runs an agent: it supplies the instructions, decides which tools the agent may call, and keeps its credentials.
AGENTS.mdA file at the root of a repository that tells any agent runner how to work there. In the reference repository it is generated from framework.json.
Agreement floorThe share of alerts on which a playbook's outcome must match the analyst's, at validate and through propose mode. The default is 0.95, higher for a playbook that closes alerts on its own.
AI SOC productA product that triages or investigates alerts with a language model, separate from the SIEM or XDR that raised them.
AlertOne firing of a detection. It costs an analyst the minutes it takes to triage. Alert volume is what analysts feel.
Alert volume thresholdThe most alerts a detection may raise in its first 7 days live with alerting off: one and a half times the weekly count measured at research, and never under 3. Alerting switches on only when the count stays at or under it.
Alert-rate checkA scheduled query that counts a live detection's alerts per hour and switches its alerting off when the count passes the limit on its record, then opens a tuning pull request.
Atomic, micro, fullHow much an emulation run covers. Atomic is one technique and one signal question, micro chains two or three behaviors, and full runs end to end across planes and surfaces.
ATT&CKMITRE's public catalog of what attackers do, organized as tactics (the goal) and techniques (the method). Coverage here is counted per tactic and each detection names its technique.
BacklogThe list of detection candidates waiting to be built, in score order.
Binding ceilingWhichever of the three ceilings (triage, maintenance, or telemetry) the live count reaches first. It is the smallest of the three, and it sets how many live detections the program can carry.
Core logic lockA hash over the selection blocks that express a detection's behavior, taken when the rule is authored. A tuning change may add filters beneath those blocks and must leave the hash unchanged, so an agent tuning out noise cannot weaken the rule.
Definition of doneThe countable condition a detection must meet to leave a lifecycle stage, such as "N engine passes, N golden diffs, one approval". "It looks ready" cannot be counted.
DetectionThe whole artifact: the rule that matches the behavior, the record that explains it, the fixtures that prove it fires, and the score that justified building it.
Detection as codeKeeping detections as files in version control and shipping them through a build, so a change is reviewable, testable, and reversible.
Detection zeroThe first detection a program builds: an alert for each log source when its events stop arriving. Every other detection depends on it.
DetonationRunning the attack step of an emulation technique against a live sandbox, after its prerequisites exist. In Stratus Red Team, the step that writes the events a rule is written from.
EmulationRunning an adversary behavior on purpose, in a controlled environment, to produce the events a detection will be written and tested against.
Eventual consistencyA cloud control plane accepting a change before every service can see it. Poll for readiness, and let logs settle before calling telemetry missing.
Evidence lineThe one line under each scoring input that says where the number came from: an incident id, a report section, a query and its result, or a ticket. A reviewer argues with the evidence line, and CI rejects a record missing one.
False negativeAn attack that got through unflagged: a rule that did not fire on it, or a triage agent that closed its alert.
False positiveAn alert that was not the behavior it was written for. Its cost is the triage minutes it consumes.
False-close rateThe share of alerts later confirmed malicious that triage had closed as benign, by a person or an AI agent. Section 19's ai_false_close_rate is the same count for an agent alone.
Field dictionaryThe table and field names exported from the SIEM's own schema by a scheduled job, used to check every field a rule reads. It replaces documentation and memory, which is where invented field names come from.
FixtureA stored event, captured once, that a detection is replayed against in CI. A true-positive fixture must produce an alert; a false-positive fixture must not.
FootholdThe access an emulation starts from: stolen long-term credentials, an assumed role, or compromised compute. Starting from an admin session produces events no attacker would generate.
Golden fileThe approved copy of a converted query. CI diffs the freshly converted query against it, so a conversion change cannot ship unnoticed.
Inventory coverageA detection that is mapped to a technique and reports as coverage, but whose query never touches the telemetry that technique produces. The heatmap is green and the log is silent.
KPIKey performance indicator: a metric that says how well the program does its work. The team sets a target for it.
KRIKey risk indicator: a metric that says how exposed the program is. The team sets a tolerance for it, the level at which someone has to act.
Ladder tierHow much a playbook changes outside the alert: 1 enriches, 2 routes, 3 closes, 4 contains. Each tier needs more of a person.
MandateA regulatory or audit requirement attached to a candidate: the framework, the requirement number, and a due date. A mandate due within 90 days goes to the front of the backlog, ordered by date, and a mandated candidate with readiness 0 is still blocked.
Marginal coverageThe factor that scores a candidate on what it adds to the live set: 1.0 when no live detection covers its technique, 0.8 when its robustness is higher than the best one that does, and 0.5 when the technique is already covered as well. The scorer derives it from the live set, and CI fails a record whose value disagrees.
MCPModel Context Protocol, a standard way to give an agent a tool, such as a read-only query interface to a SIEM.
Miss codeThe one code a detection review gives an incident or exercise that no live detection caught first, naming where the path from log source to analyst broke.
Orphaned resourceSomething the emulation created that infrastructure as code does not know about: a backdoor identity, a key, a grant. Destroy these before the provisioned resources.
Pipeline pagerThe rotation that answers for the detection pipeline itself: a silent log source, a red build, an alert count past its limit, or a playbook action waiting for approval.
PlaybookThe automation that runs when a detection fires: it enriches the alert, decides, and acts or proposes an action. It has a record and moves through the same eight stages as a detection.
PrecisionOf a detection's alerts in a stated window that an analyst closed with an outcome, the share that were true positives. A benign true positive counts against it, and an undetermined close counts on neither side. The program reads it over 30 days, as precision_30d.
Precision floorThe lowest precision a detection can run at and still be worth its triage minutes: triage minutes divided by the program's cap on analyst minutes per true positive. A detection that stays under its floor for two 30-day periods in a row meets the first retirement trigger.
Prompt injectionText written to steer an agent that reads it. In detection work it arrives inside log content an attacker controls: command lines, user names, URLs, and file names.
Propose modeA playbook state in which it enriches, decides, and writes what it would do, and a person does it. Every playbook runs this way for 14 days before it acts.
ReplayLoading a detection's fixtures into a real instance of a query engine and running the converted query against them. A pass proves the query matches what it should. A conversion that only parses proves nothing about behavior.
RobustnessHow hard the detection is to evade, scored on the Summiting the Pyramid levels. A detection matching a filename is brittle; one matching the behavior itself lasts.
Run costWhat a detection costs every day it is live: alerts per day multiplied by the minutes to triage one.
Run tagA unique name and short id put into resource names, identity names, and session names for one emulation run, so its telemetry can be filtered and its leftovers found at teardown.
Shadow modeRunning an AI triage agent on live alerts with no power to act, so its verdicts can be compared with analysts' decisions before it closes anything.
SIEMThe system that stores the telemetry and runs the detections against it.
SigmaAn open format for writing a detection once in YAML and converting it to the query language of whichever SIEM you run.
Signed fixtureAn event captured by a collector the agent cannot write to, carrying the collector's signature. The build verifies the signature before it counts the event as proof that a rule fires.
Summiting the PyramidMITRE's model for scoring how hard a detection is for an attacker to evade, which this framework uses as the robustness input.
TargetOne query language the pipeline converts a Sigma rule into, listed in targets.yml with the backend that emits it. A replay target has an engine CI can run. A convert target has no public instance, so its query stays a draft until somebody runs it on a real one.
TelemetryThe events a detection reads: logs, EDR data, network records, whatever the SIEM receives.
Telemetry readinessWhether the data a detection needs is already in the SIEM and usable, scored 0 to 5. A readiness of 0 blocks the candidate, because nothing can be built on missing telemetry.
Threat-model intakeThe practice of filing every unmitigated threat in a service's threat model as an intake candidate, with the technique, the service's impact, and the log source it needs.
TriageAn analyst reading one alert and deciding whether it is real. Measured in minutes, because that is what runs out.
True positiveAn alert that was the behavior the detection was written for, whether or not it turned into an incident.
TuningChanging a live detection to cut its false positives without losing the behavior, by excluding a named source of noise.
Warm-upIn Stratus Red Team, creating a technique's prerequisites with Terraform before anything attacks, so the detonation has something to act on.
WIP limitWork in progress limit: how many detections a team allows in flight at once. Weekly engineering hours divided by the hours one detection takes.
From the data file
Scoring inputs, lifecycle stages, maturity levels, and the scaling ceilings, in their own words.
Threat relevance (T)How recently, and how close to you, has this technique been used?
Impact (I)What does the attacker reach if the technique succeeds?
Robustness (B)How hard is it for an attacker to slip past this detection using the data you have? Score it with the five Summiting the Pyramid levels: 1 matches something the attacker can change in seconds, 5 matches something the technique cannot work without.
Telemetry readiness (R)Is the data this detection needs already in the SIEM, and in good enough shape to use?
Build effort (E)How many hours of work from writing the hypothesis to passing validation?
Run cost (C)How many analyst minutes a day will this detection consume? Expected alerts per day multiplied by the typical minutes to triage one. For a detection that runs as scheduled SQL, add the warehouse compute it burns, converted the same way your finance team converts it.
Intake (stage)Write down a request for a new detection with enough context that it can be scored.
Hypothesis (stage)Write one sentence saying what the detection will catch, which log source it needs, and what it will miss.
Research (stage)Produce the events the rule will be written against, and confirm they exist in your logs. Then count how often they appear during normal business.
Build (stage)Write the detection in Sigma and the tests that prove it.
Validate (stage)Prove the detection fires on what it should catch, stays silent on what it should ignore, and still matches the query a reviewer approved.
Deploy (stage)Put the detection into production with alerting switched off, leave it running for 7 days, and switch alerting on only if the alert count stayed at or under its threshold.
Tune (stage)Keep the share of true alerts above the agreed floor as the environment changes.
Retire (stage)Remove detections whose alerts cost more analyst time than they are worth.
Ad hoc (maturity level)Detections exist. Nothing about them is written down in a place a second person can find.
Repeatable (maturity level)Every detection has a record and a defined way in. Two people would build the next one the same way.
Measured (maturity level)The program counts its own work: definitions of done carry numbers, the backlog carries scores, alerts carry outcomes.
Enforced (maturity level)Machines enforce the definitions of done. A skipped step fails the build.
Self-correcting (maturity level)Measurement drives change: detections retire themselves on evidence and coverage claims can be falsified.
Triage ceilingHow many live detections can the analysts absorb?
Maintenance ceilingHow many live detections can the engineers keep current?
Telemetry ceilingHow many detections can be built at all?