the second hundred · metaphor 141
A result came from a team. Now split the credit. The one who arrived last and closed the deal feels decisive — but they only looked decisive because everyone before them had already built the thing up. Who really deserves what?
Every shared achievement drags this question behind it. The startup's exit, the paper's citations, the family that raised the child, the platform that connects a buyer to a seller — each was made by several hands, and none of them can be cleanly pulled apart. Ask any member "what did you add?" and you get an honest but useless answer: it depends who else was there. Alone, they'd have made little. In the right company, they were the difference between nothing and everything.
The trouble is that our instinct anchors on the order things happened in. The last hire before the breakthrough, the closer, the parent who was home the year it mattered — they catch the light because they showed up at the decisive moment. But the moment was only decisive because of who came first. There is a way to strip the accident of order out entirely: don't credit one story of how the team came together — credit the average over every story it could have been.
| order they joined | A’s marginal add | B’s marginal add | C’s marginal add |
|---|
The idea
Start with the only credit you can measure without fantasy: a person's marginal contribution — how much the group's worth jumps the moment they join. But that jump depends entirely on who's already inside. Join an empty room and your marginal add is whatever you're worth alone. Join a room already primed to use you, and your marginal add can be huge. Join last, after the value is already made, and it can be nothing.
So a single marginal number is not fair — it's a hostage to arrival order. Lloyd Shapley's move is to refuse to privilege any one order. With three players there are 3! = 6 orders the team could have formed. Compute each person's marginal contribution in all six, then take the plain average. That average is the Shapley value — and it is the unique way to split the total that treats identical members identically, gives nothing to a member who adds nothing anywhere, and always divides the whole pie exactly, no more and no less.
That last property — the shares summing to precisely the grand total — is called efficiency, and the instrument shows it holding on every setting. Nothing is invented, nothing is left on the table; the value the team actually created is what gets distributed.
What to try
Press Average the 6 orders and watch the table light up one row at a time. Each row is a different story of how the team came together — A→B→C, then A→C→B, and so on — and each cell is the marginal jump that player caused in that story. The same person can be worth a lot in one order and little in another. The bars at the bottom hold the running average, and by the sixth order they settle onto the fair shares.
Now edit the team. The sliders set what each group is worth together: what a player earns alone, what a pair unlocks by combining, what all three unlock as a whole. Try the presets. In The glue, one member is nearly worthless alone yet makes everyone else productive — and the Shapley value hands them the largest share, because across the orders they keep being the piece that turns nothing into something. In Free rider, one member never changes any group's worth, and their share collapses to zero automatically. You never assign credit; the averaging does.
The mapping
The Shapley value is what you reach for whenever a joint result must be split and the parts don't separate. Author order and credit on a paper that no single person could have written. Bonuses across a sales team where the closer and the lead-generator each think they were decisive. The cost of a shared road among towns that use different stretches of it. Attribution in advertising, where a customer touched five channels before buying — which one gets the sale? The honest answer to all of them is the same: not the one who happened to be last, but each party's average difference across every order they could have entered.
What the metaphor gives ordinary life is a discipline against the recency of credit. We over-reward the visible closing move and under-reward the invisible setup, because we watched the story unfold in one particular order and mistook that order for the causes. The Shapley value says: run the tape from every starting point. The person who is consistently the difference-maker — whatever the order — is the one who earned the share, even if they were never the one in the room when the champagne opened.
Read as life lessons
Whoever arrives at the decisive moment looks decisive. But their marginal add is large only because others built the setup. Average over orders and the glory redistributes to whoever kept being the difference.
A member who earns nothing solo can still take the biggest share — if they're the synergy that makes others productive. Fair credit tracks what you unlock in company, not what you'd make on a desert island.
If your presence never changes any group's outcome, the average of your marginals is zero, and the formula hands you nothing — no politics required. It also never invents value: the shares sum to exactly the whole.
In the wild
SHAP values — Shapley values over a model's features — are now the standard way to explain a prediction: each feature's fair share of the pushed-up or pulled-down score, averaged over the orders features could enter.
Airports split runway costs among aircraft of different sizes, and firms divide profits from joint ventures, using Shapley allocations — the one split everyone can be argued into as uniquely fair.
Marketing teams credit conversions across the channels a customer touched, moving past "last click" toward Shapley attribution that rewards the setup, not just the closing tap.
The mapping, exactly
| Mathematics | Life |
|---|---|
| a player | Anyone with a stake in a shared result — a co-founder, co-author, teammate, channel, contributor. |
| coalition value v(S) | What a given subset of the team would be worth together — the thing you can measure without pretending to separate people. |
| marginal contribution | How much the outcome jumps the moment you join a group already assembled — your worth given who's there. |
| the 6 orderings | The many stories of how the team could have come together — no single one of which deserves to define your credit. |
| averaging over orders | The refusal to reward the accident of arrival time; the fairness of running the tape from every start. |
| efficiency (shares sum to v) | Dividing exactly the value that was made — inventing none, leaving none behind. |
The honest model
The sliders define a value for each of the eight subsets of three players: zero for the empty team, a solo worth for each individual, and for each pair and the trio, the solo worths plus whatever synergy combining unlocks (synergy can be negative — people who get in each other's way). Nothing else is stored.
From that table the script enumerates all 6 orderings. For each order it walks the team in one by one and records every player's marginal jump, v(joined-so-far + you) − v(joined-so-far). A player's Shapley value is the plain mean of their six marginals: φᵢ = (1/6) Σ marginalᵢ. Every number on the page — each cell, each bar, the grand total, the sum of shares — is computed from those enumerated marginals live; the "sum of shares equals the grand total" check isn't asserted, it's recomputed and displayed, and it holds because the marginals in any single order already telescope to the whole.
Where the metaphor tears
The Shapley value assumes you know what every subgroup would have been worth — including the ones that never happened. In real teams those counterfactuals are unknowable: what would the paper have been worth without the third author? We can't rerun history. The math is exact; its inputs, in human life, are usually guesses dressed as data.
Averaging over orderings with equal weight is a fairness axiom, not a fact of the world. If some collaborations were always going to form first — a founder before a late hire — treating every arrival order as equiprobable can misplace credit. The neutrality that makes it feel fair is also a thumb on the scale you chose not to see.
Shapley splits by marginal value created, full stop. But we also allocate by need, by effort, by risk borne, by promise made. A caregiver who contributed most may deserve rest, not the largest bonus. The value answers "who made the pie?" — a question that is not always the same as "who should get the biggest slice?"