# What does it mean for a deal to be fair?

*Essay 001 · 15 August 2026 · 7 min — https://redeal.dev/blog/what-makes-a-deal-fair*

Fairness is not a feeling about an outcome. It is a set of properties a
procedure either has or does not have, and the difference matters more now
that machines make deals with each other.

## § 01 — You can't see fairness in the outcome

Two freelancers finish a £1,000 job and split the fee £612.50 / £387.50. Is
that fair?

You can't tell by looking. Those exact numbers could have come from a
published formula applied to both parties' sealed valuations, or from
whoever shouted loudest on the phone. Same outcome in both worlds.
Different fairness. Fairness lives in the procedure that produced the
numbers, not in the numbers themselves.

This sounds obvious and gets ignored constantly. People argue about
outcomes ("that's not fair") when the only thing that can actually be fair
or unfair is the process: who knew what and when, who had discretion, who
could change their answer after seeing yours, who can check the result
afterwards. A good outcome from a bad procedure is luck. A bad outcome from
a good procedure is a deal you can accept, because you accepted the rules
before you knew where you'd land.

## § 02 — The properties that make a procedure fair

Strip the idea down and a fair dealing procedure needs four properties.
Each names a way deals get rigged, and closes it.

- **No envy.** Every party ends up preferring their own share to anyone
  else's, by their own stated values. With two parties this is achievable
  exactly: the adjusted winner procedure is provably envy-free. [1] With
  more, the realistic target is proportionality, meaning everyone gets at
  least a 1/n share by their own valuation.
- **No peeking.** Going last is an advantage, unless every input is sealed
  before any is opened. A commitment scheme handles this with a hash. You
  publish a fingerprint of your input up front, and the fingerprint pins
  you to it. [2] Nobody gets to shade their number toward the field.
- **No discretion.** A referee with judgment calls can be lobbied, bribed,
  or simply wrong. A published deterministic algorithm makes no judgment
  calls: same inputs, same output, every time. Contribution splits get
  their fairness from a formula rather than a feeling. The Shapley value is
  the classic example. [3]
- **No trust required.** Anyone can re-run the computation and check the
  certificate: a party, a skeptic, another machine. Verification that needs
  the referee's cooperation is not verification. Random outcomes need the
  same discipline, with entropy drawn from somewhere no party controls. [4]

## § 03 — What fairness is not

**Fairness is not truth.** If both parties seal inflated numbers, a fair
procedure divides the fantasy fairly. The certificate attests that these
parties sealed these inputs and this algorithm ran. Whether the inputs were
honest is not something it can know. Attesting agreement rather than truth
is what keeps the referee neutral; a referee who judges inputs has taken a
side.

**Fairness is not enforcement.** A certificate moves no money and compels
no one. Enforcement belongs to other layers: escrow, contracts, reputation.
The dealing layer sits underneath them and produces the artifact those
layers can rely on.

**Fairness is not sameness.** Splitting everything 50/50 feels safe and is
often the least fair option on the table. Equal shares ignore that parties
value things differently and contribute unequally. A fair procedure
respects declared differences; a lazy one papers over them.

> A fair deal isn't one where you trust the referee. It's one where you
> don't have to.

## § 04 — Why this matters more for agents than for people

Humans have rough substitutes for fair procedure: ongoing relationships,
small-claims court, reputation, tone of voice. Slow and expensive, but they
exist. A software agent has none of them. It transacts with counterparties
it will never meet again. It cannot sue, cannot read the room, cannot rely
on working things out later. It can only check.

That changes what fairness has to be. Between people, fairness is mostly a
feeling to be managed. Between machines it has to be a computation that can
be verified: sealed inputs, a published algorithm, a signed receipt both
sides can confirm independently. The certificate becomes the smallest unit
of trust: not "I believe you" but "here is the proof, re-run it yourself."

The deals worth doing this way today are small: splitting an invoice,
sharing an API budget, picking who goes first. That's fine. Habits form at
low stakes, and the habit worth forming is demanding a procedure you can
check. When the deals get big, the procedure will already be there.

---

## Notes

1. Adjusted winner: Brams & Taylor, *The Win-Win Solution* (1999). Two
   parties, divisible goods, provably envy-free and equitable. It is the
   2-party core of Redeal's split.v1. https://redeal.dev/docs
2. Redeal commitments are `sha256(JCS(input) ++ salt)`; JCS is RFC 8785
   canonical JSON, so the hash is over bytes anyone can reproduce.
   https://redeal.dev/how-it-works
3. Shapley, "A Value for n-Person Games" (1953): each party's share is
   their average marginal contribution across every coalition they could
   join. Merit by construction, not negotiation. Redeal's shapley.v1.
4. random.v1 XORs every party's sealed nonce with the drand beacon at a
   round declared before any input was sealed. No party, and not Redeal,
   can steer the result.

---

*Redeal is a neutral fairness-computation API: sealed inputs, published
deterministic algorithms, identical signed certificates anyone can verify.
Try a split at https://redeal.dev/split, or read how it works at
https://redeal.dev/how-it-works.*