
analyze whether a real procedure is fair using probability, design a fair procedure, prove its fairness and test it by simulation.
Instruction
A procedure is fair, in the probability sense, when every person it treats alike has the same chance of each outcome. Fairness can be checked with mathematics, and unfair procedures can often be repaired.
Checking a game. Two players roll two dice. Player A wins if the sum is 6, 7 or 8; otherwise B wins. From the 36-outcome grid, sums of 6, 7 and 8 occur in 5 + 6 + 5 = 16 outcomes, so P(A wins) = 16/36 = 4/9 and P(B wins) = 5/9. The game is not fair. Two repairs: change the rule (A wins on an even sum: 1 + 3 + 5 + 5 + 3 + 1 = 18 outcomes, exactly 1/2), or change the payoff (A scores 5 points per win and B scores 4, so each player’s long-run average is equal: 4/9 × 5 = 5/9 × 4 = 20/9).
A fair coin from an unfair one. Suppose a coin lands heads with an unknown probability p. Flip it twice. If the flips are HT, A wins; if TH, B wins; if HH or TT, flip twice again. Since the flips are independent, P(HT) = p(1 − p) = (1 − p)p = P(TH). The procedure is fair whatever p is. This trick is due to the mathematician John von Neumann.
Fair selection. Twenty-five students want 3 field-trip seats. A random draw of 3 from 25 gives each student probability 3/25 = 0.12 of a seat. You can prove this with counting: the draws that include a particular student number 24C2 = 276 of the 25C3 = 2,300 possible draws, and 276/2300 = 0.12. A “first come, first served” sign-up is not random: it favors students who see the notice first, and a two-way table of who signed up (for example, by bus riders versus walkers) can reveal that.
Your report. Choose a real decision that should be fair (seats, turns, parking, chores, team selection, a game). Then:
- Describe the current or proposed procedure and compute the probabilities it gives each person or group, using a sample space, a two-way table or counting.
- Decide whether it is fair, with reasons. If you use real data about people, keep it anonymous.
- Design an improved procedure and prove it is fair with a probability argument.
- Simulate your procedure at least 100 times (dice, cards, a spreadsheet’s random function or CODAP) and compare the results with your theoretical probabilities.
- Discuss one sense of “fair” that probability does not capture (for example, need or merit).
Presentation. Three minutes, live or recorded: the decision, the problem you found, your procedure and the proof. A written brief is an equal-credit alternative.
📖 Read for this lesson
The reading opens over the lesson; close it to come back to exactly where you were. The checks below draw on it.
Formative check
Work through these before moving on. They are not graded — they tell you whether the standard below has landed yet.
P(sum is 6, 7 or 8) with two dice
Tap to reveal16/36 = 4/9
Fair coin from a biased coin
Tap to revealFlip twice: HT means A, TH means B, otherwise repeat.
Chance a given student gets one of 3 seats among 25 in a random draw
Tap to reveal3/25 = 0.12
Test for a fair game with unequal chances
Tap to revealCompare each player’s long-run average payoff.
In a two-dice game, A wins on a sum of 6, 7 or 8 and B wins otherwise. Which change makes the game fair?
Stage 1: describe the decision you will analyze, who it affects, how it is made now, and which probability tools (sample space, two-way table, counting) you expect to use.
How confident are you that you can analyze and design fair procedures with probability and defend a recommendation?
Unit 4 checkpoint
Do this before you start the task. It covers the whole unit and tells you what to review first.
For two dice, P(sum = 6) = /36.
P(queen or spade) for one card is:
P(A or B) = P(A) + P(B) for any two events.
In a table, 45 of 150 students play an instrument, and 18 of those 45 are seniors. P(senior | instrument) is:
If P(A and B) = 0.12 and P(B) = 0.4, then P(A | B) = .
P(A) = 0.6, P(B) = 0.5 and P(A and B) = 0.3. A and B are:
After four tails in a row with a fair coin, heads is more likely than 1/2.
Mutually exclusive events with positive probabilities are independent.
Two cards are drawn without replacement. P(both kings) is:
The number of 3-person committees from 9 people is .
The number of ways to award gold, silver and bronze among 8 runners is .
A two-dice game: A wins on a sum of 6, 7 or 8. P(A wins) is:
A procedure that gives each of 25 students a 3/25 chance at one of 3 seats can be a random draw of 3 names.
Submit: Fair Decisions Report with simulation and three-minute presentation
Task: Analyze a real decision procedure for fairness with probability, design a fair procedure, prove that it is fair, test it with at least 100 simulated trials, and recommend it.
Submission workflow
- Stage 1 — Plan. The decision, who is affected, the current procedure and the probability tools you will use.
- Stage 2 — Peer-reviewed draft. Computations, fairness judgment, new procedure and proof; a partner re-computes one probability and tries to find a way the procedure is still unfair. Submit a revision note.
- Stage 3 — Revised final. Complete report with simulation results compared to theory and a discussion of what probability cannot decide.
- Stage 4 — Presentation. Three-minute live or recorded recommendation, or an equal-credit written brief.
How it is scored
| Criterion | What proficient looks like |
|---|---|
| Fairness analysis (A2.DS.A.6) | Probabilities for each person or group computed correctly and used to judge fairness. |
| Decision reasoning (A2.DS.A.6) | Improved procedure proved fair; trade-offs analyzed. |
| Tables and conditionals (G.CP.A.4, G.CP.A.3) | Two-way table or conditional probabilities used correctly where the question involves groups. |
| Counting (G.CP.A.8) | Permutations or combinations used correctly where needed. |
| Simulation (MP.4) | At least 100 trials; results compared with theory and differences explained. |
| Argument and presentation (MP.3) | Clear recommendation that addresses a limit of probabilistic fairness. |
A Real Lottery That Was Not Fair, and How It Was Fixed: USCIS H-1B Registrations, 2021 to 2026, With Two Scenarios to Analyze
USCIS, H-1B Electronic Registration Process, registration and selection counts by cap fiscal year 2021–2026 (page updated Sept. 21, 2026; retrieved Sept. 23, 2026); the field-trip and dice scenarios are SLS original · Public domain (U.S. Citizenship and Immigration Services figures), with SLS original worked examples and constructed scenarios
Why this reading. Your report analyzes a real decision that should be fair. Here is one at national scale, with the numbers the agency publishes. Each year the United States has a limited number of H-1B work visas, and when more employers register than there are visas, U.S. Citizenship and Immigration Services runs a random selection. For years the selection was by registration, and one person could be registered by several employers. The table below shows what that did, and what happened when the agency changed the rule. After the real case come two scenarios with complete data for practice, and the lesson’s own examples checked.
Table 1. H-1B registrations and selections, by cap fiscal year
Counts as published by USCIS. “Eligible” excludes duplicates, deletions, invalid passport information and failed payments. The two middle columns split the eligible registrations by whether the person named (the beneficiary) had other registrations too.
| Cap FY | Total registrations | Eligible registrations | Eligible, beneficiary had no other registration | Eligible, beneficiary had multiple registrations | Selected registrations |
|---|---|---|---|---|---|
| 2021 | 274,237 | 269,424 | 241,299 | 28,125 | 124,415 |
| 2022 | 308,613 | 301,447 | 211,304 | 90,143 | 131,924 |
| 2023 | 483,927 | 474,421 | 309,241 | 165,180 | 127,600 |
| 2024 | 780,884 | 758,994 | 350,103 | 408,891 | 188,400 |
| 2025 | 479,953 | 470,342 | 423,028 | 47,314 | 135,137 |
| 2026 | 358,737 | 343,981 | 336,153 | 7,828 | 120,141 |
USCIS also reports about 442,000 unique beneficiaries for FY 2025 and about 339,000 for FY 2026. Starting with the FY 2025 selection, the agency selected by unique beneficiary rather than by registration, so that a person named on several registrations was entered once; the page calls this the beneficiary-centric selection process.
Why selection by registration was unfair
Suppose every eligible registration had the same chance q of selection. In FY 2024, q ≈ 188,400 ÷ 758,994 = 0.248. A person with one registration had probability 0.248. A person with n registrations from n employers was selected if any of them was, and the chance that none was is (1 − q)n:
| Registrations for one person, n | 1 | 2 | 3 | 5 | 10 |
|---|---|---|---|---|---|
| P(selected) = 1 − (0.752)n | 0.248 | 0.435 | 0.575 | 0.760 | 0.942 |
Two people the rule should have treated alike had chances of 0.248 and 0.942. By the lesson’s definition, that is not fair. Now read the fourth column down the years: registrations for people with multiple registrations went from 28,125 to 408,891 in three years. Explain, with the table above, why the rule itself rewarded that growth.
How many people were actually competing in FY 2024? The page does not say, but the table bounds it. Every one of the 350,103 single-registration people counts once. The 408,891 multiple registrations belonged to people with at least 2 each, so they came from at most 408,891 ÷ 2 = 204,445 people. So there were at most 350,103 + 204,445 = 554,548 unique beneficiaries, and fewer if the average person in that column had more than two. Write the bound if the average was 3, and if it was 4.
After the change
In FY 2026, 336,153 of the 343,981 eligible registrations were for people with no other registration: 97.7%, against 350,103 ÷ 758,994 = 46.1% in FY 2024. With selection by person, the chance for each of the roughly 339,000 beneficiaries was about 120,141 ÷ 339,000 ≈ 0.354, and it was the same chance whether an employer filed once or five times. Compute the corresponding figure for FY 2025 using 442,000 beneficiaries. Then argue: did the change make the selection fair by the lesson’s definition? Is there any sense in which the FY 2024 rule was fair (to whom)?
The next argument. The same USCIS page now states, for a future selection: “If a random selection process is necessary, USCIS will conduct a weighted selection among properly submitted registrations for unique beneficiaries based generally on the highest OEWS wage level that the beneficiary’s proffered wage will equal or exceed.” Suppose, as a hypothetical to analyze rather than a description of the rule, that a registration at the highest wage level were entered four times and one at the lowest level once. Compute the two probabilities of selection in a pool of 340,000 beneficiaries and 120,000 selections if a quarter of the pool were at each of four levels. Is a weighted lottery fair? Your answer must say fair to whom, and that is the discussion the report’s last item asks for.
Scenario A (illustrative): the field-trip sign-up
Twenty-five students want 3 seats. Ten of them ride the bus, which arrives at 8:05; fifteen walk. The sign-up sheet was posted at 7:45 and full by 8:00. All 3 seats went to walkers.
| Got a seat | Did not | Total | |
|---|---|---|---|
| Bus riders | 0 | 10 | 10 |
| Walkers | 3 | 12 | 15 |
| Total | 3 | 22 | 25 |
Under a fair random draw, P(all three seats go to walkers) = 15C3 ÷ 25C3 = 455 ÷ 2,300 = 0.198. So the outcome alone is weak evidence: it would happen one time in five by chance. The argument against the procedure is the mechanism, not the outcome: a bus rider’s probability of a seat was 0 before the draw began. Design the fair procedure (the lesson’s random draw gives each student 3/25 = 0.12), prove it, and then propose a procedure that is fair in the probability sense and still rewards signing up early, if you think one exists.
Scenario B: the dice game, with a simulation record (illustrative)
Player A wins on a sum of 6, 7 or 8; B otherwise. Theory: P(A) = 16/36 = 0.444. A class simulated 120 rounds: A won 55, B won 65. The proportion 55/120 = 0.458 is close to 0.444 but not equal; state how far off it is and whether 120 rounds could reasonably produce it. Then simulate the repaired game (A wins on an even sum, P = 18/36) for your own 100 or more rounds and report the same comparison. The lesson’s payoff repair, 5 points to A and 4 to B, gives equal long-run averages: 4/9 × 5 = 5/9 × 4 = 20/9 per round. Show that this is fair in the average-payoff sense and unfair in the probability-of-winning sense, and say which sense a player would care about.
The lesson’s numbers, checked
Sums of 6, 7 and 8 occupy 5 + 6 + 5 = 16 of the 36 cells. Even sums occupy 1 + 3 + 5 + 5 + 3 + 1 = 18. Draws of 3 from 25 that include a particular student: 24C2 = 276 of 25C3 = 2,300, and 276/2,300 = 0.12 = 3/25. Von Neumann’s coin: P(HT) = p(1 − p) = P(TH) for every p, and the procedure stops with probability 2p(1 − p) on each pair of flips; compute how many pairs you should expect to need for p = 0.5 and for p = 0.9, and say what that costs in fairness (nothing) and in time (something).
Sources: USCIS, H-1B Electronic Registration Process (registration counts by fiscal year) (uscis.gov)