Appendix

20.1 Glossary of Key Terms

Acceptance Test
A binary statement that proves a change delivered its intended outcome (e.g., “Failing privacy gate blocks promotion in CI; link to gate log”). Acceptance tests turn plans into verifiable results and anchor Action Cards, gates, and funding tranches.

Action Card
The unit of execution in the roadmap: one owner, one mechanism, one acceptance test, shippable in ≤ two weeks, with dependencies, effort, and rollback plan. Action Cards appear on the “Now–Next–Later” board and feed status into weekly operating reviews.

Adoption (Activation, Weekly Active Ratio)
Behavioral usage metrics for a value case—who started, who returned, and whether work shifted to the new path. Paired with time‑to‑competence to verify enablement is working.

Anomaly (FinOps)
A material, unexpected variance in spend or unit cost (e.g., token surge, context bloat, cache miss). Must trigger investigation within 24 hours and a design change when warranted.

API (Application Programming Interface)
The programmatic interface between services or providers; in this Diagnostic it is a control surface for routing, caching, logging, and policy‑as‑code checks.

Back‑Out / Rollback Plan
A pre‑rehearsed method to safely revert a change (config toggle, router rule, deploy rollback). Required on Action Cards and experiments.

Baseline
The empirically measured “before” state for a KPI. All value claims and experiments compare against baselines with stated windows and confidence.

Benefits Realization (CFO‑Signed)
The method and evidence that convert improvements into financial outcomes (revenue, margin, cost, capacity redeployment). Must be signed by Finance within 90 days of scale.

BI (Business Intelligence) Embedded View
A web‑native dashboard (no downloads) that renders heat maps, scorecards, and panels. Must default to capped truth and link to evidence.

Cache / Cache Hit Rate
A performance and cost lever that reuses prior results. Tracked as a driver of unit cost and latency; adjusted via TTL and keying strategy.

Canary / Ramp
Controlled exposure patterns for changes (e.g., 1% → 10% → 50% → 100%), guarded by safety/latency/cost bands and rollback rules.

Cap (Score Cap)
A forced limit applied to a readiness score when a gate fails or evidence is insufficient. Caps ensure structural risks are visible in the heat map until removed.

Capped Truth (Default View)
Portfolio visualizations show capped scores by default, with hatch marks on capped cells and bold borders on scale‑blocked items. Prevents averages from masking risk.

Canonical ID
Stable identifiers (e.g., vc_id, sys_id, rel_id, corpus_id, ven_id, hashed person_id) that let you join telemetry, evaluator outputs, and logs consistently across tools.

Change Calendar
A web page listing dated releases, drills, QBR/JBRs, training, and office hours. All Action Cards above the cutline should appear here.

Change Freeze / Blackout
Time windows where promotions are paused (e.g., end of quarter). The roadmap and calendar must respect freezes.

CI/CD (Continuous Integration / Continuous Delivery)
Pipelines that build, test, and release software. In this Diagnostic, CI/CD enforces policy‑as‑code gates with failing‑gate examples.

CI Policy‑as‑Code
Executable checks in pipelines for privacy, safety, and model risk. Must have a failing‑gate example on record and a waiver register for exceptions.

Citation (RAG)
A visible link or reference that grounds an answer in retrieved content. Accuracy and completeness of citations are tracked as quality metrics.

Confidence (Metric)
A 0.0–1.0 signal of how much to trust a number, based on recency, sample size, triangulation, and data lineage health. Low confidence requires a dated plan to raise it.

Confidence Discount (Scoring/Prioritization)
A factor that reduces the Priority Index or Quick‑Win Index when evidence confidence is weak, curbing over‑commitment to uncertain estimates.

Cohort
A targeted group of users for staged adoption or analysis. Used in canaries, training assignments, and time‑to‑competence tracking.

Control (ERC)
A designed mechanism (gate, review, logging, reconstruction, DSAR purge) that prevents or mitigates risk. Controls must be demonstrated in operation, not just described in policy.

Corpus (RAG)
The set of documents or content used for retrieval‑augmented generation. Managed via a registry with licensing, classification, and re‑index cadence.

Curation (RAG)
Dedupe, canonicalization, tagging, and quality improvements to corpora, prioritized by query heat and failure analysis.

Cutline
The hard boundary on the roadmap that reflects real capacity for the next six weeks. Items below the cutline do not start.

Data Lineage
The provenance and transformation trail for data. Required for reproducibility, audit, and confidence scoring.

Data Residency / Retention
Where data is stored and how long it is kept. Retention windows must be visible on hub pages and enforced in logging.

Decision Log
A running record of approvals and conditions with owners and due dates. It is the audit trail from metrics to actions.

Dimension / Sub‑Dimension
The components of the Readiness Framework scored 1.0–5.0 (e.g., Data Foundation → Data Quality; Technology → Architecture & Observability).

Discovery / Pilot / Scale / Optimize / Sustain
Funding and delivery stages. Discovery proves feasibility; Pilot validates in a bounded cohort; Scale expands under guardrails; Optimize hits steady‑state economics; Sustain maintains and extends.

DORA Metrics
Delivery health measures (lead time, change failure rate, mean time to restore). Used alongside SLOs and error budgets.

DPA (Data Processing Agreement)
A contract governing data use by vendors. A non‑negotiable gate for any provider handling sensitive data.

DPIA / TRA (Privacy & Risk Assessments)
Documented impact assessments for data processing and threats. Status and mitigations must be visible in portal pages.

DSAR (Data Subject Access Request) Purge Drill
A tested path to locate and purge prompts, outputs, embeddings, and logs upon request. A stopwatch result is recorded.

Drill (Portability, Reconstruction, DSAR)
A scheduled test of a critical control. Drills are tracked on the calendar with time‑to‑complete and defects logged.

Ecosystem (Vendors/Partners)
The set of external providers and integrators you rely on. Managed via QBR/JBRs, price ladders, audit rights, and portability drills.

Egress (Cost/Ecosystem)
Fees or friction to move data out of a provider. A key input to exit economics and the Portability Risk Index.

Embedded Analytics (Web‑Only)
BI dashboards embedded in the portal with export disabled for viewers. Tooltips deep‑link to evidence.

Embedding / Vector Store
Numerical representations of text (embeddings) stored in a vector database for similarity search. Subject to portability drills and licensing constraints.

Enablement (Manager Kit, First‑Week Rituals)
The practical support that turns access into correct use: talk tracks, job aids, routines, in‑product guidance, and office hours.

Error Budget / Burn
The allowable unreliability over a period and how quickly you consume it. A burn breach triggers mitigation and can pause scale.

Evaluator (Offline / LLM‑as‑Judge)
An automated or human‑assisted tool that scores task quality, retrieval, grounding, or safety. Must be versioned and periodically calibrated with human adjudication.

Evidence Index
A master page of resolvable links to systems of record—logs, contracts, evaluator outputs, drills, dashboards. Enables the two‑click rule.

Exit Economics (Portability)
Time‑to‑switch, egress cost, and effort to move away from a vendor. Demonstrated via drills, not slideware.

Export Disabled (Web‑Only Standard)
A non‑negotiable setting for the portal and embedded dashboards—no CSV/PDF/PPT/DOC downloads for viewer roles.

Fairness (Slice Metrics)
Performance parity across protected cohorts on material tasks. Requires measurements, documented mitigations, and governance.

FinOps (AI)
Operational finance for model/platform usage: unit cost per task, anomaly response, commitment coverage, and price‑ladder posture.

First‑Week Rituals
Simple behaviors that accelerate correct adoption (e.g., “Run three tasks with citations, post one tip, log one issue”). Instrumented and measured.

Gate / Gating Ledger
A pass/fail control (e.g., logging & reconstruction < 60 minutes). The ledger lists all gates, their status, and acceptance tests; failed gates cap scores.

Gate Severity (Scale‑Blocking vs. Capped vs. Discretionary)
Scale‑blocking gates prevent external exposure; capped gates limit scores; discretionary checks are recommended but not gating.

Golden Dataset (Evaluation)
A labeled, versioned set of tasks for offline testing, including edge cases and fairness slices.

Golden Path / Paved Road
The supported standards (gates, templates, router defaults, evaluator configs) that make the right way the easy way. Updated when Learning Cards produce wins.

Grounded Answer (RAG)
An output where every factual claim is supported by citations. Measured via grounded‑answer rate and citation accuracy.

Heat Map (Capped by Default)
A portfolio visualization of readiness scores with cap hatching and scale‑blocked borders. Tooltips link to cap reasons and evidence.

HITL (Human‑in‑the‑Loop)
A reviewer workflow for edge cases or safety overrides. Must have SLAs and audit samples.

Hub Page (Change)
The single, web‑native source of truth for a rollout—what changes, why, what to do, support paths, disclosures, and links to training.

Incident / PIR (Post‑Incident Review)
A service impairment or safety issue, followed by a structured review that results in changed tests, gates, or runbooks within two sprints.

Inter‑Rater Reliability (IRR)
Agreement among human judges on labels or ratings; a quality control for evaluators and fairness studies.

Investment Envelope / Sequencing
The 90‑day funding amount and the staged release of tranches based on evidence and acceptance tests.

Jailbreak / Prompt Leakage (Safety)
Prompt patterns that bypass guardrails or reveal system instructions/PII. Tested in pre‑release packs and monitored in production.

Kill Ratio (Discovery/Pilot)
The share of experiments or pilots purposefully stopped. A healthy program kills ≥ 25% to avoid endless pilots.

Kill‑Switch (Router)
A config or policy that instantly stops sending traffic to a provider or model. Proven during portability drills.

KPI (North Star & Guardrails)
A compact set of metrics—one outcome KPI per value case plus guardrails for quality/safety, reliability/latency, unit cost, and compliance.

Latency p95 / p99
Time to serve 95% or 99% of requests. Guardrails must protect user experience during canaries and ramps.

Learning Card
A registered hypothesis, design, guardrails, and outcome. Successful cards update the paved road; unsuccessful ones are documented and retired.

License (Corpus/Data)
Legal rights to use content in retrieval or training. Corpus entries must include licensing status; unlicensed content fails a gate.

LLM (Large Language Model)
A generative model accessed via API or platform. Managed via routing, safety filters, and economics controls.

Logging & Reconstruction (< 60 Minutes)
The ability to recreate what the system did (prompts, outputs, data paths) quickly and lawfully. Proven via quarterly drills with a stopwatch result.

Manager Kit
A concise package (talk track, FAQs, rituals, escalation tree) that prepares managers to lead a change. Shipped before any broadcast.

Model Risk Management (MRM)
Governance for model behavior and use (bias, stability, misuse). Often paired with privacy to define gates and reviewer SLAs.

MTTR (Mean Time to Restore)
Median time to recover from an incident. Trends, paired with error‑budget burn, indicate operational maturity.

Narrative, Readiness
A one‑page story that answers “Are we ready to scale? What caps block us? What will we do by when?” Strictly evidence‑linked.

North Star KPI
The single metric that best captures the value case’s outcome (e.g., cycle‑time reduction). Always paired with guardrails.

On‑Call / Runbook
Staffing and instructions to handle alerts and incidents. Pages are tested via synthetic alerts or table‑tops.

Operating Model (AI)
Who owns what (product, platform, data/ML, risk, SRE, FinOps), how work flows (release trains, gates), and how decisions are made (cadence, escalation).

P0 Corpus / Dataset
Must‑have content or data for a value case. Registered, licensed, curated, and re‑indexed on a cadence.

Paved Road
See Golden Path.

Pilot
A bounded release to a target cohort. Must meet minimum viable thresholds and pass gates before broader scale.

Portability Drill
A scheduled, stopwatch‑timed switch to a qualified secondary (LLM, embeddings, vector store). Produces defect lists and fixes.

Portability Risk Index (PRI)
A composite score of data export friction, config portability, API openness, and substitutability. Lower is safer; drills validate claims.

Precision@K / Recall@K (Retrieval)
Quality metrics for RAG retrieval on a labeled set: how many retrieved documents are relevant and how many relevant documents were retrieved.

Price Ladder / Caps & Credits
Commercial terms by usage tiers, with protections against spikes. Managed alongside routing/context to control unit cost.

Priority Index (PI)
A ranking formula that combines Value at Risk, risk severity, time criticality, confidence discount, reversibility, and effort/delay. Sorts Action Cards by impact per unit of work.

Productivity Conversion
The mechanism to turn time saved into budgeted outcomes (hiring deferral, span changes, overtime reduction) so benefits are realized.

Prompt / Pattern Library
Reusable instructions and formats for tasks; curated and tested to reduce errors and speed up time‑to‑competence.

Quick Win / Quick‑Win Index (QWI)
A small, reversible change that moves a KPI in ≤ 30 days. QWI ranks candidates by immediate lift, confidence, reversibility, effort, and dependencies.

RAG (Retrieval‑Augmented Generation)
An approach that grounds generation in retrieved content. Governed via corpus licensing, retrieval/grounding thresholds, and citations.

Readiness Dimension / Maturity Level (1.0–5.0)
Scored aspects of readiness (e.g., Data, Technology, Talent). Levels range from ad‑hoc (1.0) to leading (5.0) per the rubric.

Readiness Narrative
See Narrative, Readiness.

Release Train
A predictable cadence for promoting changes with freeze windows, gates, and rollback rehearsals.

Reproducibility (Experiment/Metric)
The ability for a second person to recreate a result from raw artifacts within a stated time. A KPI for the learning system.

Reversibility Bonus
A multiplier in prioritization that favors safe‑to‑undo changes (router configs, caching, prompts) over hard‑to‑reverse moves.

Reviewer SLA
Time commitments for privacy, model risk, security, and legal reviews. Managed as visible lanes on the roadmap board.

Risk Severity Multiplier
A prioritization factor that boosts items clearing scale‑blocking or regulatory gates.

Router / Routing Policy
Logic that selects models/providers per request (by class or confidence), with kill‑switches and secondary fallbacks.

Safety Block / Override
A filter that prevents unsafe outputs and a reviewer pathway to override when justified. Override rates are audited.

Scale (Decision)
Authorization to expand a pilot broadly because gates are green, thresholds are met, and unit economics are under control.

Scale‑Blocking Gate
A gate that must be green before external exposure (e.g., logging & reconstruction, DPA executed). Marks heat‑map cells with bold borders.

Shadow Traffic
Mirrored requests used to exercise a candidate route or provider without impacting users.

SLA / SLO / SLI
Service contracts/targets/indicators. SLOs and error budgets govern reliability; SLAs bind responders and reviewers.

Snapshot Stamp
A banner (timestamp, thresholds version, weight set) on dashboards and pages that identifies the exact scoring snapshot behind the numbers.

Stage‑Gate Funding
Tranches released only when acceptance tests are passed (Discovery → Pilot → Scale → Optimize → Sustain).

Telemetries (Events, Traces, Logs)
Instrumented signals that feed the warehouse (with canonical IDs) and enable KPI calculations and reconstruction.

Threshold (Minimum Viable / Good)
The numeric lines a KPI must meet to scale, and the higher bar to exit optimization. Versioned quarterly.

Three Lanes (Portfolio Allocation)
Uncap & Control, Scale Winners, Explore & Enable. Fixed capacity allocations prevent starvation of foundations or diffusion into experiments.

Time‑to‑Competence (TTC)
Median time for a user to reach sustained correct use. A leading indicator of adoption quality.

Time Criticality Factor
A prioritization factor that boosts items with near‑term impact or expiring windows (e.g., queued releases, contracting deadlines).

Time‑to‑Switch (Portability)
Measured duration to move a workload to a secondary. Reported from drills with a stopwatch and defect log.

Token / Context Length
Primary cost and latency drivers in LLM usage (inputs/outputs, prompt size). Controlled via routing, trimming, chunking, and caching.

Two‑Click Rule
Any number in a dashboard or page must lead to its raw evidence in two clicks or fewer.

Unit Cost per Task
A normalized economics KPI: all variable costs (tokens in/out, vector queries, reranker calls, GPU time, egress, overhead allocations) divided by tasks completed.

Value at Risk (VaR, Diagnostic)
The 12‑month economic opportunity gated by caps or low maturity, weighted by dimension importance and potential lift to target.

Value Case
A specific, measurable business problem the AI transformation aims to improve (with a baseline, counterfactual, and benefits method).

Waiver Register
A public list of approved exceptions to gates with owners, mitigations, and expiry dates. Prevents “forever exceptions.”

Weight Set (Scoring)
The relative importance assigned to dimensions/sub‑dimensions. Versioned and published so score changes can be interpreted.

WIP Limit (Work‑in‑Process)
A ceiling on concurrent items (e.g., two Scale cases per BU, five foundation cards per platform team) to maintain flow and reviewer capacity.

Workplan (Diagnostic)
The timeline and dependencies for interviews, data collection, analytics, scoring, and readouts. It aligns stakeholder availability and reviewer SLAs.

This glossary is intentionally practical. If a definition does not help you decide, test, or ship, adapt it. The Diagnostic is designed to evolve; keep your glossary versioned and review it quarterly alongside thresholds, weights, and the paved road.