The finding

A lab session cannot tell personal guilt from partner-specific commitment. Comparing worlds with different background trust can.

In one minute

The question. Why keep a promise that no longer pays? Two reasons predict it. Personal guilt: it hurts to let down an expectation you created. Partner-specific commitment: it hurts to break your word to this person.
What a lab cannot separate. After a partner switch, both reasons predict that you roll less. The new partner knows nothing of the switch, so background trust is the same with and without it.
What separates them. Background trust: what people expect of someone before any promise. In a simulation, what personal guilt earns is low at both ends of it and high in the middle. What partner-specific commitment earns does not move.
What this is not. It does not say which reason people feel. It says which comparison could tell them apart. The model’s numbers were not calibrated against any experiment.

Already known: Vanberg (2008) switched partners to separate what the other expects from the word you gave. Promises were kept less often after a switch, and he read it as a preference for keeping one's word in itself.; Kawagoe and Narita (2014) proposed personal guilt, and derived that after a switch it is zero: the promise the new partner holds was not made by the one who decides.
What this work adds: Personal guilt and partner-specific commitment fall in the same pair of Vanberg's design, so his result fits both.; The variable that separates them: background trust. Personal guilt depends on it; partner-specific commitment does not.; A population model, in which reasons spread by what they earn, and an open simulation engine that measures it.; A negative result: who gets to speak does not change who survives (below).

Three worlds

In each world, the belief that a promise will be kept is the same, 76 out of 100. Only background trust changes.

A world where almost no one keeps their word
Background trust
5
What a promise adds
from 5 to 76
Guilt available
3.55
Above the threshold, 20/3?
No
Personal guilt rolls
No
Partner-specific commitment rolls
Yes
A world in between
Background trust
38
What a promise adds
from 38 to 76
Guilt available
14.44
Above the threshold, 20/3?
Yes
Personal guilt rolls
Yes
Partner-specific commitment rolls
Yes
A world where almost everyone does
Background trust
70
What a promise adds
from 70 to 76
Guilt available
4.20
Above the threshold, 20/3?
No
Personal guilt rolls
No
Partner-specific commitment rolls
Yes

Where almost no one keeps their word, little was expected to begin with. Guilt is the product of what was expected and what your promise added, and with almost nothing to start from, the product stays small.
Where almost everyone does, your promise added almost nothing. The product is small again.
In between, both amounts are large, and guilt is at its highest.
Partner-specific commitment looks at none of this: it rolls whenever the promise binds it.

What each reason earns

In the simulation, the other person sees which reason moves you before playing, and only joins if you will roll. If they join and you roll, you earn 10; if they stay out, each keeps 5. So rolling is what pays. Personal guilt earns 10 only inside its window. Partner-specific commitment and general guilt earn 10 everywhere.

Personal guilt rolls
Personal guilt doesn't roll
Threshold, 20/3

Guilt available, and whether personal guilt rolls, at each value of background trust
Background trust	Guilt available	Personal guilt rolls
0 of 100	0	No
5 of 100	3.55	No
10 of 100	6.60	No
15 of 100	9.15	Yes
20 of 100	11.20	Yes
25 of 100	12.75	Yes
30 of 100	13.80	Yes
35 of 100	14.35	Yes
38 of 100	14.44	Yes
40 of 100	14.40	Yes
45 of 100	13.95	Yes
50 of 100	13.00	Yes
55 of 100	11.55	Yes
60 of 100	9.60	Yes
65 of 100	7.15	Yes
70 of 100	4.20	No
75 of 100	0.75	No
76 of 100	0	No

Personal guilt
Partner-specific commitment
General guilt (control)
Personal guilt, with the cap (robustness)

Payoff of each reason at each value of background trust
Background trust	Personal guilt	Partner-specific commitment	General guilt (control)	Personal guilt, with the cap (robustness)
0 of 100	5	10	10	5
5 of 100	5	10	10	5
10 of 100	5	10	10	5
15 of 100	10	10	10	10
20 of 100	10	10	10	10
25 of 100	10	10	10	10
30 of 100	10	10	10	10
35 of 100	10	10	10	10
38 of 100	10	10	10	10
40 of 100	10	10	10	10
45 of 100	10	10	10	10
50 of 100	10	10	10	10
55 of 100	10	10	10	10
60 of 100	10	10	10	10
65 of 100	10	10	10	10
70 of 100	5	10	10	10
75 of 100	5	10	10	10
76 of 100	5	10	10	10

Top: guilt available at each row of the grid, against the threshold. Bottom: what each reason earns at the same rows. Each value of background trust is a separate world, held fixed: this is a comparison across worlds, not a population moving along the curve.

When the reasons compete

The figure above shows what each reason earns in a world. Here, the two reasons compete inside one. A population starts half personal guilt and half partner-specific commitment. In every generation people play, and the reason that earned more tends to be copied. Background trust is held fixed for the whole run, so each world stays the world it is.

Where each run ended, after 400 generations

A world where almost no one keeps their word · 5

A world in between · 38

A world where almost everyone does · 70

Where each run ended, after 400 generations
Background trust	Runs	Personal guilt gone	Only personal guilt	Still mixed
A world where almost no one keeps their word · 5	20	20	0	0
A world in between · 38	40	15	14	11
A world where almost everyone does · 70	20	20	0	0

One dot per run. At 5 and at 70, every run ends without personal guilt. At 38, the runs end spread out: some without it, some with only it, some still mixed.

For those who want the arithmetic

Four reasons, two pairs

Each reason of the main thread has a general version and a narrower one. Guilt can answer to anyone’s expectation, or only to one you created. Your word can bind you to anyone you promised, or only to this person. That makes four. With a partner switch, each one predicts:

Reason	What it asks of you	With a partner switch
General guilt	not to fall short of what anyone expects of you	you roll the same
General commitment	to keep your word, given to anyone	you roll the same
Personal guilt	not to let down an expectation you created	you roll less
Partner-specific commitment	to keep your word to this person	you roll less

Vanberg’s design separates the two pairs, but not what is inside each pair. That personal guilt is zero after a switch is not an assumption of this page: Kawagoe and Narita (2014) derive it themselves.

The formula

Call background trust a, and write every belief in hundredths, so that 76 is the belief that a promise will be kept: the mean second-order belief of dictators without a partner switch in Vanberg (2008, Table I). The guilt available to you is then a · (76 − a) / 100. Personal guilt rolls the die if θ · guilt > 4, the cost of rolling (14 − 10), with θ = 0.6, how much a unit of guilt weighs against a unit of money: that is, if the guilt is above 20/3, about 6.67.

Setting a · (76 − a) / 100 = 20/3 gives the analytic cut: personal guilt rolls between about 10.1 and about 65.9. The grid the curve measures only has rows at some values, so there it rolls from 15 to 65. The peak is at 38, where the guilt is 14.44.

Partner-specific commitment has a fixed cost, c = 5, for breaking your word. Since 5 > 4, it always rolls when the promise binds it. General guilt answers to the belief that a promise will be kept, which does not depend on background trust, so its choice does not either.

Two more results

Who speaks does not decide. In the lab, an agreement, where both people promise, is kept more often than a promise made by one side alone (Di Bartolomeo, Dufwenberg, Papa and Passarelli, 2023). In the model, letting both people speak does not reproduce that. If the one who decides promises just as often, a population where both can speak ends exactly like one where only the decider can: the same reasons survive, seed by seed, in 60 runs out of 60. This follows from how beliefs start: an agreement and a one-sided promise open with the same expectation, and no reason in the model treats them differently. It does not say that agreements do not matter to people. It says that, to produce that gap, a model needs more than these four reasons.

When a promise can stop binding. Let a promise stop binding this partner some share of the time, and let the other person see which reason moves you. Then the other joins someone who rolls only when bound if that share is below one half. Between zero and one half, personal guilt and partner-specific commitment survive. From one half on, general guilt and general commitment do. It is Vanberg’s split into two pairs, read as who wins. Inside each pair, the two reasons end up earning the same, so the share does not tell them apart, except in a narrow band just below one half that runs through the same belief channel (the paper, Section 6). What does tell them apart, across the whole range, is background trust.

These two results use other settings than the curve; the paper lists them (Section 6 and Table 1).

What would settle it

Another partner-switch session would not: background trust is the same in both of its conditions, so it sees a single point of the curve. In the model, personal guilt keeps its word only in the middle range of background trust, and partner-specific commitment keeps it everywhere. Telling them apart takes comparing sessions, or populations, whose background trust differs.

A robustness check

In one variant, the prior expectation can be worth no more than what the other gets without playing, 5. The right tail does not fall: from 70 on, personal guilt still earns 10. The variant was announced before the measurement was run.

Limits

The other person sees which reason moves you. If they could not, not rolling would pay more and the order of payoffs would flip; this was not varied.
This compares worlds with background trust held in place. In a normal run of the engine, background trust is rewritten every generation, and the population stays at a single point of the curve.
In the middle, where two reasons earn the same, which one remains is chance.
θ and c were not calibrated against the experiments.
This is an identification result, about what can be told apart, not a claim that people feel personal guilt.

Where the numbers come from

From a simulation engine I wrote in TypeScript. It is deterministic: every run has an explicit seed, and the same seed repeats the same run. Its tests pin the curve on this page, and one command regenerates every run in the paper. It is archived with a DOI. github.com/Montse2308/promises-to-whom-engine (opens in a new tab) (DOI 10.5281/zenodo.23222610 (opens in a new tab))

How to cite

The working paper

Hernández Gallegos, M. X. (2026). Promises to whom: Identifying personal guilt and partner-specific commitment across populations [Working paper]. SSRN. https://doi.org/10.2139/ssrn.7580218

@misc{hernandezgallegos2026promises,
  author = {Hern{\'a}ndez Gallegos, Montserrat Ximena},
  title  = {Promises to whom: Identifying personal guilt and partner-specific commitment across populations},
  year   = {2026},
  note   = {SSRN Working Paper},
  doi    = {10.2139/ssrn.7580218},
  url    = {https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7580218}
}

The engine

Hernández Gallegos, M. X. (2026). Promises to whom: Identifying personal guilt and partner-specific commitment across populations (Version 1.0.0) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.23222610

Working paper: Promises to whom: Identifying personal guilt and partner-specific commitment across populations (opens in a new tab) (SSRN).

← Back to “This is where I come in”