Reference
Glossary
Every term taught in the 30 days, in one list. Search as you study.
135 terms
10× stress
Recompute the envelope at ten times the growing axis (users, questions, hops, or input size). Success is the expensive case.
90-day plan
The first-quarter operating plan in the new seat. Interviewers ask; good hires have one.
ACL on chunks
Access control so retrieval cannot return documents the user is not allowed to see.
Agent
A model-in-a-loop with tools. Not a personality and not a license to skip controls.
Agent loop tax
The extra model calls created by think → tool → think cycles. Cap them.
Allowed gap
A missing skill the loop does not actually fail you on (usually deep training / leetcode).
Artifact
A probe-able document or working path (charter, prompt library, cost envelope, walkthrough). Not a badge.
Ask
The concrete decision you want in the room: people, access, money, or a no.
Baby eval
A small, honest scorecard (format, evidence, usefulness) you run by hand before you have a harness.
Bake-off
Running two or more models on the same eval set, same prompt, same path, and picking from the table.
Bound-auto
The model may act only inside a numeric envelope; outside it, confirm. Envelope unwritten means auto with a story.
Canary
Releasing a change to a small % of traffic before the rest.
Change control
Who may edit the prompt, and how you know production still matches the tested version.
Charter
The one document that funds a slice: scope, non-goals, data, eval, cost, RAID, sequence, ask.
Chunking
Splitting documents into retrieval units. A product decision with overlap and structure.
Citation theatre
Showing sources that do not actually support the generated sentence.
Context window
The maximum tokens a single request can see — instructions, history, documents, and output combined.
Data class
A handling category (public / internal / confidential / restricted) that decides legal path.
Data-ready
A corpus that has owner, class, ACL, retention, and sampled quality — a go-live gate.
Day 91
The first stretch after the vendor or the stand-up crew leaves. Ownership is real or it is not.
Death date
The calendar day the POC tenant dies unless a named exec signs a production design.
Demo-ware
A product that shines on the vendor's corpus and ACLs-off tenants, and fades on yours.
Deprecation
Vendors retire model names. Your evals and prompts must survive a swap.
Deprecation drill
The written project you run when a vendor retires a pinned model: eval, canary, rollback, comms.
dict
A key→value map. JSON objects become dicts in Python. The default shape of AI payloads.
DLP
Data loss prevention — scanning for secrets and PII before they leave a boundary.
Door
An enterprise platform (Azure OpenAI, Bedrock, Vertex) that sells a control plane and a data path, with model names behind it.
DPA
Data processing agreement. The contract that says what a vendor may do with prompts.
Embedding
A vector representation of text used for semantic similarity and retrieval.
Embedding model
A model that maps text to a vector for similarity, not for chatting.
Environment variable
A named value in the process environment. Standard place for secrets.
Error budget
How much miss you tolerate before freezing change.
Eval
A test set and a scoring method that tell you whether the AI system is good enough to ship — the AI equivalent of acceptance criteria.
Exit / export
Your ability to take prompts, evals, traces, and indexes with you when the vendor relationship ends.
Fallback
A designed degraded mode, not an accident.
Few-shot
Providing a handful of input→output examples in the prompt so the model copies the pattern.
Fine-tune
Updating model weights on your data. Rarely the first lever; expensive to maintain.
First token
The latency the user feels when streaming starts. Often the SLO that matters more than complete.
Forward Deployed Engineer (FDE)
An embedded builder-consultant who makes an AI / data product work in a specific customer's environment, not in a generic demo.
Freeze
Declaring a prompt version as the one that may run; further edits require a new version and a re-test.
Gateway / AI gateway
An enterprise control plane in front of one or more models: keys, logs, policy, routing.
Golden set
A versioned, labeled set of inputs and expected behaviors used as the test suite for an AI system.
Groundedness
Whether the answer's claims are supported by the retrieved sources.
Grounding
Supplying source material in context (or via tools) so the answer can be tied to evidence.
Grounding control
Putting the right sources in this request (or refusing) so claims can be tied to evidence. Not a polite sentence in the prompt.
Guardrail
Any control that constrains model behavior: policy, schema, filters, authz, caps, HITL.
Hallucination
Fluent, confident output that is not grounded in provided or true facts.
HIPPO
Highest paid person's opinion — a governance failure mode.
Hiring loop
The sequence of rounds. Different seats, different shapes — prepare the one you picked.
HITL
Human in the loop — a required review or submit step, default on writes and on low-confidence extraction.
HITL theatre
A human gate that cannot reasonably say no — no hours, no UI, no reasons, overflow into auto.
Honesty pass
Deleting claims you cannot sustain in a 10-minute probe.
Hosted API
A vendor runs the model; you send tokens over the network under a contract.
Human gate
A required confirmation before a write-tool executes.
Hybrid search
Combining vector similarity with keyword / filter retrieval.
Idempotency key
A client-supplied id so retries of the same action don't duplicate it.
Inference
Running a trained model to produce outputs. Distinct from training.
Ingest
The pipeline that turns raw documents into clean, chunked, metadata-tagged index records.
Input tokens
Tokens you send: system, history, retrieved docs, user. Usually cheaper per million than output.
Instruction hierarchy
A stated precedence when system rules, retrieved policy, and user requests conflict.
Instruction neglect
The model follows a conflicting user (or document) instruction instead of the standing spec.
Job OS
Your personal WIP-limited pipeline for conversations, applications, and pack upgrades.
Job-to-be-done
The user's actual task (find the decision, draft from SOP), not the technology (chat).
JSONL
A file with one JSON object per line. The usual format for eval sets and logs.
Kill criteria
Pre-agreed numeric conditions under which you pause or stop, published before you start.
Kill switch
A named, tested way to disable the feature or its write path within minutes.
Knowledge owner
The human accountable for whether a corpus is current, non-conflicting, and in-scope.
Leadership no
A refusal that includes a thinner path and a clear ask, not a lecture.
LLM-as-judge
Using a model to score another model's output against a rubric. Useful, biased, needs calibration.
Maker-checker
Classic control: one party proposes, another approves. HITL confirm is this pattern.
Multimodal
Models that take or produce more than text (images, audio). Same delivery problems, extra data types.
No-AI alternative
The process, search, or form you'd ship if models did not exist. A test of whether you understood the job.
Non-goal
A tempting expansion you explicitly refuse in v1 so the slice can ship.
Offline eval
Batch scoring on a fixed set, run on every change — the regression test.
Online eval
Live metrics: thumbs, traces, sampled human review, cost, latency.
Open weights
Model parameters you can download and run yourself or on a specialist host.
Operating note
The one-page runbook for the library: data rules, HITL, storage, owners, what to do when it invents.
Operating pack
The one-page table of a use case a steering group can actually decide on.
Output contract
The required shape of the answer (schema, headings, table). Makes evals and downstream use possible.
Output tokens
Tokens the model writes. Cost and latency both live here.
Pack
The single assembled portfolio you speak from and can send.
Permission-aware retrieval
Retrieval that only returns chunks the current user is allowed to see.
Pin
Locking a model version (and prompt version) so they cannot drift under you.
POC
Proof of concept. In AI, often a demo that never becomes a product because evals, data, and change were skipped.
POC charter
The written clock, scope, success bar, and death date of a proof of concept.
Positioning
The one-line story of who you are for, the problem you take, and the proof. Not a job title list.
Primary role
The seat you will write materials and stories for. One, not five.
Prompt cache
A vendor feature that discounts a repeated prefix of the prompt across calls.
Prompt injection
Hostile or accidental instructions in user input or retrieved docs that try to override the system rules.
Prompt library
A versioned set of prompts with owners, tests, and usage rules — like a process pack, not a chat history.
Prompt version
An id for the spec that produced the call. Required if you want to debug or eval.
Prompt versioning
Treating the prompt plus sampling plus schema as a named build with tests and a changelog.
Pushback
A predictable objection. You prepare it, you do not improvise it.
Qualify
Turning a slogan request into a job, a user, and a frequency before you design.
RAG
Retrieval-Augmented Generation: find relevant chunks, then generate an answer conditioned on them.
Rate limit (429)
The vendor is throttling you. Queue and back off; do not hammer.
Re-index SLA
The promised maximum lag between a source changing and answers reflecting it.
ReAct
A common pattern: reason + act (tool) + observe, repeated.
Recall@k
How often the pages a human marked as relevant appear in the top k retrieved chunks. A retrieve eval, not an answer eval.
Release bar
The numeric gates and named waiver path without which you do not ship.
requirements.txt
A list of Python packages a project needs. The bill of materials.
Reranker
A second-stage model that reorders a shortlist of chunks for a query.
Retention clock
How long you keep logs and indexes, and how long the vendor keeps prompts. Two numbers, both written.
Routing
Sending easy/high-volume calls to a small model and the low-confidence tail to a frontier model, with a written switch condition.
Sampling
Choosing the next token from the model's probability distribution (greedy, temperature, top-p).
Sanitization
Stripping names, IDs, and confidential facts before using a tool that is not approved for that data.
Shadow → assist → confirm
The default intensity ramp for AI features.
Shadow mode
The model runs in parallel with the old process; humans don't depend on it yet; you collect eval.
SLO
Service level objective — a numeric promise you manage to.
Steering-ready
A pack four readers can mark up without a tour: sponsor, engineer, security, finance.
Streaming
Sending tokens to the UI as they are generated to hide latency.
Structured output
Forcing the model to emit JSON / schema so downstream systems can consume it.
Study OS
The calendar holds, slip rule, redraw habit, and no-list that keep a 60-hour month from becoming a pile of tabs.
Subprocessor
A vendor's vendor who may see prompt text. Named in the DPA, or treated as unknown disclosure.
Sycophancy
The tendency to agree with the user's preferred answer instead of the evidence.
System of record
The authoritative business system (CRM, ERP, ITSM). AI that cannot write here cleanly stays a side chat.
System prompt
Standing instructions for the model: role, rules, output contract. Not the daily instance.
Take-home policy
Your pre-decided time cap and deliverable so you do not disappear for a weekend.
Talk track
The timed story of the artifact for an interview or steering session.
Temperature
A sampling knob. Low makes outputs drier and more deterministic; high increases variety and invention.
Thin slice
The smallest end-to-end path that creates value and can be measured — not a platform, not a chatbot of everything.
Title inflation
The gap between a posting's title and the actual first bullets.
Token
A chunk of text the model reads and writes. Pricing and context limits are in tokens, not words.
Tool calling
The model requests a named function with arguments; your code executes it and returns results.
Tool contract
Name, schema, authz, side effects, idempotency, and failure behavior of a function the model may call.
Top third
Headline, profile, and first bullets — the only part most screeners read.
Trace
The timeline of thoughts, tool calls, and results. Your debugger when it 'goes crazy.'
Unit economics
Cost (and latency) per successful task, not per demo.
Use-case scoring
A published rubric (value, feasibility, risk, change) used to fund or kill ideas.
Vector index
A store that can return nearest neighbors for a query vector (FAISS, pgvector, vendor DBs).
Virtualenv
An isolated Python package environment per project.
Waiver
A named, time-boxed permission to ship below the bar, with a rollback. Hope is not a waiver.
Walkthrough
A timed spoken tour of the pack, designed for questions, not a TED talk.
Weights
The learned parameters of the model. Frozen at inference unless you fine-tune.
WIP limit
A cap on how many opportunities you work at once so loops get quality.