Glossary

|:--|:--| | Use when | You encounter an unfamiliar term, or need to explain a concept to stakeholders | | Time | Reference as needed | | Outcome | Understanding of every term in technical and executive language | | Related | Stakeholder Glossary ・ README |

---

Terms used in this manual, with definitions for both technical and executive audiences.

---

Failure Modes

Legibility Loss

Technical: The system produces outputs, but you cannot trace how a user action led to that output. Investigation requires manually reconstructing context.

Executive: We cannot explain why the AI did what it did. This is a compliance risk and an enterprise sales blocker.

Related: Legibility Loss

---

Control Surface Drift

Technical: When user behavior and system state interact in ways that multiply recompute, latency, or retries, often invisibly.

Executive: Our costs are rising because users are developing expensive habits that we did not anticipate.

Related: Control Surface Drift

---

Auditability Gap

Technical: You can show what the system output, but you cannot prove what it knew, why it acted, or who approved the behavior.

Executive: We cannot answer a regulator's question about why the AI made a specific decision. This is a compliance gap.

Related: Auditability Gap

---

Margin Fragility

Technical: Unit economics degrade as the system scales, often because hidden recompute grows faster than revenue.

Executive: Growth is making us less profitable because our AI costs do not scale linearly.

Related: Margin Fragility

---

Architecture Concepts

Decision Envelope

Technical: The complete context needed to explain and reproduce an AI output: inputs, policy versions, tool calls, and approvals.

Executive: A record that proves exactly how and why the AI made a decision.

Related: decision-envelope-schema.json

---

Speculative State

Technical: State produced under uncertainty: drafts, retries, partial generations. If treated as final state, it corrupts your data model.

Executive: AI outputs that have not been confirmed by a human. We track these separately from finalized decisions.

---

Reversibility Window

Technical: The period during which a technical decision can be changed cheaply. After the window closes, changes become expensive migrations.

Executive: How long we have before a technical choice becomes locked in. Some early decisions are now very expensive to change.

Related: Architecture Decisions

---

Control Surface

Technical: Any point where user behavior or system state can trigger recompute, latency changes, or cost.

Executive: Points in our product where user actions cause the system to do expensive work.

---

Interaction Contract

Technical: Explicit rules defining what users can request, when the system recomputes, and what state can be mutated.

Executive: The agreement between our product and users about what happens when they take actions.

Related: Interaction Contract

---

Control Plane Ownership

Technical: Whether you own or rent the systems that govern AI behavior: orchestration, logging, policy enforcement, and infrastructure.

Executive: Do we control the critical parts of our AI system, or are we dependent on vendors?

Related: Control Plane Ownership

---

State Model

Technical: The schema and semantics of how AI outputs are persisted, versioned, and proven.

Executive: How we store and track what the AI produces.

Related: State Model

---

Economics Concepts

Hidden Recompute

Technical: Compute work that happens without user awareness: retries, context invalidation, auto-saves triggering full pipelines.

Executive: AI processing that costs us money but does not create visible value for users.

Related: Hidden Recompute

---

Cost per Successful Outcome

Technical: Total cost (inference + orchestration + retries) divided by number of outputs that delivered value.

Executive: What it actually costs us to produce something useful, not just to run the AI.

Related: Cost Model, Metrics Reference

---

Trigger Type

Technical: Classification of what caused an inference call: user explicit action, edit, undo, auto-save, retry, or background job.

Executive: Why the AI ran. Some triggers are user-initiated, others are system-initiated.

---

Hidden Recompute Ratio

Technical: Percentage of compute triggered by non-explicit user actions (undo, auto-save, retry, background).

Executive: What fraction of our AI costs are invisible to users.

Related: Metrics Reference

---

Reliability Concepts

Circuit Breaker

Technical: A pattern that stops calling a failing service and returns a fallback, preventing cascade failures.

Executive: An automatic safety switch that prevents one broken component from taking down the whole system.

Related: orchestrator.py

---

Checkpoint

Technical: Saved state at a known-good point in a multi-step workflow, enabling resumption after failure.

Executive: A save point that lets us continue from where we left off instead of starting over.

---

SLO (Service Level Objective)

Technical: A target for system performance, typically latency percentiles or availability percentages.

Executive: Our performance target. What we promise users about speed and reliability.

Related: Latency and SLOs

---

SLO Coupling

Technical: The feedback loop where latency changes user behavior, and behavior changes system load.

Executive: When the system is slow, users retry, which makes the system slower.

Related: Latency and SLOs

---

Observability Concepts

Latency Percentiles (p50, p95, p99)

Technical:

• p50 (median): Typical user experience
• p95: Experience for 1 in 20 users
• p99: Worst-case for 1 in 100 users

Executive: How fast the system is for most users, and for the unlucky few.

Related: Metrics Reference

---

Trace ID

Technical: A unique identifier that follows a request through all system components, enabling end-to-end debugging.

Executive: A tracking number that lets us see exactly what happened for any user action.

---

Provenance

Technical: A verifiable record of how a piece of state was produced: inputs, tools, constraints, and approvals.

Executive: The evidence chain that proves how a decision was made.

---

Context Amnesia

Technical: Loss of the decision context that made an output valid at the time it was produced.

Executive: We can see what the AI produced, but we have lost the information about why it produced that specific output.

---

Quality Concepts

Golden Set

Technical: A fixed set of prompts and scenarios with known-good outputs, used for regression testing.

Executive: A collection of test cases that we use to make sure the AI still works correctly.

Related: Eval and Regression

---

LLM-as-Judge

Technical: Using a capable model to evaluate outputs for quality, safety, or accuracy.

Executive: Using AI to grade AI outputs at scale.

Related: llm-as-judge-prompts.md

---

Eval Gate

Technical: A CI/CD check that blocks deployment if evaluation scores fall below thresholds.

Executive: An automatic quality check that prevents shipping bad AI updates.

Related: eval-gate.yml

---

Drift Detection

Technical: Monitoring for semantic changes in AI output distributions over time, often using embedding distance.

Executive: Watching for the AI to start behaving differently than expected.

Related: Eval and Regression

---

Compliance Concepts

Sovereignty

Technical: The architecture that guarantees where data lives, who can access it, and under which jurisdiction it operates.

Executive: Our ability to prove to regulators and customers that data stays where it is supposed to stay.

Related: Sovereignty

---

Operational Independence

Technical: The ability to maintain critical functions without depending on vendor intervention.

Executive: Can we keep operating if our vendors have problems?

Related: Operational Independence

---

Data Residency

Technical: Guaranteeing that data is stored and processed only in specific geographic regions.

Executive: Proving to European customers that their data never leaves Europe.

Related: Sovereignty

---

Orchestration Concepts

Sequential Chain

Technical: Running agents or tools one after another, passing output forward.

Example: Input -> Research Agent -> Drafting Agent -> Review Agent -> Output

---

Router Pattern

Technical: A decision node that routes to different agents based on input classification.

Example: Query -> Classifier -> (Simple Agent OR Complex Agent)

---

Parallel Fan-Out

Technical: Running multiple agents or tools simultaneously, then aggregating results.

Example: Query -> [Agent A, Agent B, Agent C] -> Aggregator -> Output

---

Hierarchical (Supervisor)

Technical: A supervisor agent that coordinates specialized sub-agents.

Example: Supervisor -> (Research Agent, Analyst Agent, Writer Agent) -> Supervisor -> Output

---

State Machine

Technical: An orchestration pattern with explicit states and transitions, useful for complex workflows with branching.

Related: Orchestration

---

Quick Reference

Term	One-Sentence Definition
Legibility Loss	Cannot explain why outputs happened
Control Surface	Where user actions trigger AI work
Control Surface Drift	Costs rising from user habits
Auditability Gap	Cannot prove why AI decided
Margin Fragility	Growth erodes profitability
Decision Envelope	Proof of how AI decided
Speculative State	Unconfirmed AI outputs
Reversibility Window	Time before changes get expensive
Context Amnesia	Lost the "why" behind outputs
Hidden Recompute	Invisible compute waste
Cost per Successful Outcome	True cost to produce value
Interaction Contract	Rules for user-triggered actions
Provenance	Evidence chain for decisions
Control Plane Ownership	Who controls the AI system
Sovereignty	Where data lives and who controls it
Operational Independence	Running without vendor dependency
Circuit Breaker	Automatic failure isolation
Checkpoint	Save point for resumption
SLO	Performance target
Trace ID	Request tracking identifier
Golden Set	Regression test cases
LLM-as-Judge	AI grading AI outputs
Eval Gate	Quality check blocking deployment
p50/p95/p99	Latency percentiles
Hidden Recompute Ratio	Fraction of invisible AI costs

---

For Executive Presentations

Opening Statement

AI products have a unique failure mode: they can work perfectly and still become impossible to explain. When that happens, we cannot defend our decisions to customers, regulators, or partners.

The Four Risks

1. We cannot explain why the system did what it did (legibility)
2. User behavior is driving up costs invisibly (control surfaces)
3. We can show what happened but not why (auditability)
4. Growth is making us less profitable (margin fragility)

Related

See: Board Explainer for the full presentation template. See: Stakeholder Glossary for non-technical translations.