Every term on this site, defined once

AI and betting, in plain English.

Written for someone who has never used an AI tool and never placed a bet. No maths, no jargon defined with more jargon.

The words that describe the machinery. None of them require maths to understand.

How AI systems work

Model: A large statistical pattern-matcher. It was shown an enormous amount of text or data, learned which things tend to follow which, and now produces the most plausible next piece of output for whatever you give it. It is not a database of facts, and it does not look things up unless it has been given a tool to do so.
Prompt: The instruction and information you hand a model. Everything the model knows about your particular question is in the prompt, which is why the same model can be useful or useless depending on how carefully it was briefed.
Inference: One run of a model: prompt in, answer out. Training is the expensive one-off that created the model; inference is the cheap repeated act of using it. When the Empire says an agent produced something, it means an inference happened.
Token: The unit a model reads and writes in — roughly a short word or a piece of one. Models are billed and limited in tokens rather than words, which is why length limits are described oddly.
Context: How much a model can hold in mind at once, measured in tokens. Anything outside the context window may as well not exist to the model, so a system that needs long memory has to be built to feed the right slice back in.
Training: The process that created the model, by adjusting billions of internal numbers until its predictions matched an enormous body of examples. Training happened before you arrived and does not update while you use it.
Fine-tuning: Additional, much smaller training that nudges an existing model toward one kind of work or one house style. It changes tendencies; it does not give the model new facts it can be trusted to recite.
Agent: Software given one standing job, the data and tools to do it, and rules about when to stop or hand off. A chatbot waits to be asked and then forgets; an agent holds a duty. The Empire runs 779 named agents, each owning one narrow duty inside one department.
Orchestration: Coordinating many agents so their work adds up: who runs when, who checks whom, what happens when two disagree, and where a person has to sign off. The hard part of a multi-agent system is almost always the orchestration, not the individual agents.
Hallucination: Fluent, confident, invented output — a statistic that does not exist, a player who did not play, a source that was never written. It happens because a model generates plausible text rather than verifying it, and it is the main reason AI work has to be graded against real outcomes.
Guardrail: A rule enforced outside the model, because a model cannot be trusted to police itself. Refusing to publish without a timestamp, holding a release until a human clears it, or blocking a market the Empire does not price are all guardrails.
Human in the loop: A design where a person reviews, corrects or approves machine output before it counts. The Empire's version is called HumanAssist. It exists because AI failures are confident and clean rather than obvious.

How a claim about performance is measured — and how it gets flattered.

Evidence and grading

Backtest: Running a method over past data to see how it would have done. Useful, but easy to flatter: the method was built by someone who already knew what happened next, so small choices about which events to include or which prices to assume can turn a loser into a winner.
Out of sample: Testing on data the method was never tuned on. An in-sample result tells you the method can describe the past; an out-of-sample result is the first weak hint that it might describe the future.
Grading: Deciding, under fixed rules written in advance, whether a published output won, lost, pushed, voided or needs review. Grading only means something when the rules were set before the outcome was known.
Push: A tie at the exact number published. No money changes hands, so it is graded as neither a win nor a loss rather than being quietly counted as one.
Void: A selection that never had a fair chance to settle — a cancelled event, a player who did not take the field. It is removed from the record rather than scored, because grading it either way would be noise.
Sample size: How many settled outcomes a claim rests on. Small samples produce spectacular percentages for free: a 5-for-6 record is what luck looks like, not what skill looks like.
Variance: The normal swing of results around their true rate. Variance is why a sound method loses for weeks and a bad one wins for weeks, and why any short run of results says more about luck than about quality.
Regression to the mean: The tendency of extreme results to drift back toward the average. A player on a hot streak against a line is usually a player who was lucky recently, which is exactly why form is scored over 5, 10 and 20 games rather than over three.
Snapshot: A saved copy of what a market or board looked like at a stated moment. Snapshots are what make later grading honest: without one, nobody can prove which number was actually available when a call was published.

The vocabulary the sports and market desks price in.

Betting and market terms

Line: The number a sportsbook posts for you to bet against — a margin, a total, or a player's statistical threshold. Betting is an argument about the number, not about who is better.
Spread: A handicap applied to the favourite. “Minus 7” means the favourite must win by more than seven points for that side to cash. A team can win the game and lose the bet.
Moneyline: A bet on who simply wins, with no handicap. The price does the balancing instead of a spread, so backing a heavy favourite pays little.
Total (over/under): A bet on the combined score rather than the winner: will the two sides together finish above or below the posted number.
Player prop: A bet on one player's statistical total — yards, receptions, points, rebounds, strikeouts — instead of on the result of the game. It depends heavily on that player's usage and health, so it moves sharply on a single lineup note.
Alternate line: The same market offered at a different number, with the price adjusted. An easier number clears more often and pays less; a harder one pays more and clears less. That trade is the whole of the Empire's EZ Picks lane.
Odds (price): What the book pays if you are right, and therefore what it implies about the chance of the outcome. American prices are written like +150 (bet 100 to win 150) or -130 (bet 130 to win 100).
Implied probability: The chance an outcome has according to its price. A price of +100 implies 50%; -200 implies about 67%. Converting price to probability is the only way to compare a bet against your own estimate.
Vig (hold, juice): The book's built-in margin. Add up the implied probabilities on both sides of a market and they come to more than 100% — that surplus is the vig, and it is why a true 50/50 proposition loses money over time.
Edge: The gap between the true chance of an outcome and the chance its price implies. A price implying 50% against evidence saying 55% is a five-point edge — and it only pays out across many bets, never reliably on one. A number you feel good about is not an edge.
Closing line: The final price a market settles at just before the event starts. It is the market's best and last estimate, which makes it the fairest benchmark to judge an earlier call against.
Closing line value (CLV): Whether the price you took beat the closing line. Consistently beating the close is evidence that your reasoning was ahead of the market, and it shows up long before a win-loss record becomes meaningful.
Bankroll: The money set aside for betting, treated as a fixed pool. Bankroll thinking is what separates a measured approach from chasing: the size of each bet is decided by the pool, not by how confident today feels.
Unit: One standard bet size, usually a small fixed percentage of a bankroll. Records are reported in units so that two people with very different bankrolls can compare results honestly.
Parlay correlation: Legs on the same game are not independent events — they rise and fall together. Any combined-probability estimate that treats them as unrelated will read higher than the ticket's real chance, which is why the Empire says so on the page rather than hiding it.

Terms that mean something specific on this site.

The Empire's own words

Department (desk): A single-subject team with its own mandate, its own agents, its own signal lane and its own page. Thirteen core departments publish into the member network today, out of 50 mapped department floors.
Realm: A grouping of related department floors on the Empire's organisational map. Twelve realms exist; the map shows which desks sit next to which without publishing any infrastructure detail.
Signal: One published release from a desk — a priced selection, a prop card, a briefing — carrying its source desk, agent, timestamp, status and media. A signal is a record of what was said and when, not a promise.
Member Wire: The single private feed every department publishes into, mirrored in the Empire App. One place, one format, every approved release.
HumanAssist: The Empire's human review layer: members holding a named seat who add context, correction, review and escalation to a department's AI work. Recognition and a seat at a desk, never employment.
Prop Lab: The board listing every quoted player prop on the slate — not only the released ones — scored against the exact posted number over the last 5, 10 and 20 games. Evidence, not picks.
Shady EZ Picks: The member builder for multi-leg slips made of alternate lines, requoted at one common book as legs change. Its saved tickets are User Lotto entries.
User Lotto: A saved EZ Picks slip. It lives in its own lane and never enters official Results, gem records or Empire win-loss totals, because it is a member-built paper ticket rather than a released Empire selection.
Gem pick: A released selection that cleared the Empire's stricter internal gate. Gems stay rare on purpose; the Prop Lab board is the evidence underneath them.
Film Room: The members' video library. Short episodes show how a desk turned a full board of data into one decision, including what was discarded.
EGTX and Imperial Credits: The Empire's progress layer. Useful member activity — reviews, reproducible bug reports, postgame analysis — earns Imperial Credits, badges and a place on the official standings. Likes, reactions and promotion earn nothing.
Private beta: The Empire's current state. Access is by invitation, nothing is for sale, no prices are published, and some lanes are deliberately paper-only so they can be tested without polluting the graded record.

Plain English

What this page does

This page defines every technical word the rest of the site uses, in plain English, so no other page has to stop and explain itself. It is written for someone who has never used an AI tool and has never placed a bet. Three groups of terms matter here: how AI systems actually work, how evidence about them is measured and graded, and the betting and market vocabulary the desks price in. The Empire's own words — desk, realm, signal, Member Wire, HumanAssist, EZ Picks, Prop Lab, Film Room — are defined at the end.

Questions people actually ask

Straight answers

What is an AI model, in one sentence?

An AI model is a large statistical pattern-matcher: it was shown an enormous amount of text or data, it learned which things tend to follow which, and it now uses those patterns to produce the most plausible next piece of output for whatever you give it. It is not a database of facts and it is not reasoning the way a person does.

What is the difference between a chatbot and an AI agent?

A chatbot answers whatever you type, one message at a time, and then forgets. An AI agent is given a standing job, the tools and data to do it, and rules about when to stop or hand off — so it can work without being prompted each time. The Empire's 779 agents are the second kind: each one owns a narrow duty inside a department.

What does it mean when an AI hallucinates?

A hallucination is when an AI produces something fluent, confident and wrong — an invented statistic, a player who never played, a citation that does not exist. It happens because the model is generating plausible output rather than looking facts up. It is the single biggest reason AI work needs grading against real outcomes and a human review layer, which is what the Empire's method exists to catch.

What is a backtest, and why is it weaker evidence than a live record?

A backtest runs a method over history to see how it would have done. It is useful but easy to flatter, because the method was built by someone who already knew what happened next; small choices about which games to include or which prices to assume can turn a loser into a winner. A live record, timestamped before outcomes are known, cannot be tuned after the fact.

What does “edge” mean in betting?

Edge is the gap between what you think the true chance of an outcome is and what the price implies it is. If a price implies a 50% chance and your evidence says 55%, that 5-point gap is the edge — and it only pays off over many bets, not on any single one. A number you feel good about is not an edge; a number that beats the price is.

What is closing line value, and why do sharp bettors care about it?

Closing line value compares the price you took with the price the market settled at just before the event. If you took a number better than the close, the market moved toward your side, which is weak but fast evidence that your reasoning was ahead of it. It matters because a single result is mostly luck, while beating the close repeatedly is a signal that shows up long before the win-loss record does.

These are plain-English working definitions for reading this site, not formal academic or regulatory definitions. Nothing on this page is betting or financial advice.