Migrating from earlier versions of Marlowe

This tutorial explains how the latest version of Marlowe differs from earlier versions of the language.

Remove Both

We do not include a Both construct in the latest version of Marlowe, which makes all contracts sequential.

Since in none of the versions of Marlowe was there communication among the branches of Both, the only extra functionality provided by Both in practice was the ability to wait for several money deposits at the same time.

We take care of this functionality by updating the When construct. Instead of having different branches wait for different inputs, we move to a completely sequential and synchronous model, where we can wait for one of several possible inputs at the same time (as in select).

The reason we removed this construct is that sequential programs are easier to analyse and easier to reason about, since there is no need for synchronisation and no opportunity for race conditions.

Include accounts

In earlier versions of Marlowe each commitment has its own timeout. This means that money deposited in a contract is not fungible, because it can be distinguished by its timeout. To achieve fungibility we have removed timeouts from individual constructs, and have a single timeout for all commitments. This contract lifetime is straightforward to infer from the timeouts in the contract.

We do, however, include accounts to organise the money deposited into the contract. This makes it more transparent how money flows within the contract, and in particular identifies to whom money is refunded when the contract terminates.

Each account is identified by a participant; the participant indicates who will get the money in the account by default when Close is reached.

The reason we chose to include accounts is that without them we found that we were essentially keeping track of the accounts manually. Additionally, in every leaf of the AST, we found ourselves calculating how much we must return to every participant, cluttering the tree with repetitious “boilerplate”. Thus, having money organised in accounts can make contracts easier to reason about and less prone to error.

Note that we can provide full fungibility by using a single account. We only need to write the appropriate Pay commands in the leaves of the contract. If all the money is paid to participants, then Close has no effect. 1

Discussion: Implicit vs Explicit Accounts

Many of the use cases for accounts – and all those that we can identify for ACTUS contracts – have one account per participant, and one participant per account (the “1-1 model”). This raises the question of whether we should give an implicit treatment of accounts, with each participant owning one account.

On the other hand, there are a variety of plausible scenarios for accounts which do not conform to the 1-1 model.

Examples where multiple participants use an account.

  • Alice owns an account to which she commits money for Bob to spend (think of Alice as Bob’s employer). Bob is able to spend up to the limit in the account, but after the commitment times out, Alice recovers anything remaining.

  • Alice owns an account to which she commits money for Bob and Carol to spend (think of Alice as Bob’s and Carol’s employer). They are able to spend (jointly) up to the limit in the account, but after the commitment times out, Alice recovers anything remaining.

  • On the other hand, they could each be given a separate account from which to spend: that would enforce individual limits as well as an aggregate limit.

  • If Bob [and Carol] want to spend money they can also add money to the account, but they should realise that anything unused will be refunded to Alice.

Examples of multiple accounts for one person:

  • Examples of underwriting would fit here. A person underwrites first-level risk, and second-level risk using different accounts. Only when the first level underwriting of all participants is spent will second level spending occur.

Close replaces Null / Pay

Since all contracts are sequential now, we can easily tell when a contract terminates, i.e: when only Null is left. We use this opportunity to close the contract and to refund any money remaining in the accounts; for this reason we have renamed Null to Close (aiding comprehensibility).

As noted earlier, there is no longer any explicit timeout in the accounts, since all contracts will eventually reduce to Close. In fact, we can statically and efficiently calculate an upper bound for when this will happen, making this aspect of Marlowe analysable.


Pay is now immediate, and it has a single continuation, and fewer parameters. 2 It allows payments from an account to a participant or to another account. We discarded PayAll, since it can be emulated as a finite series of Pay. In fact, we can define payAll as a Haskell function (see zeroCouponBond example).

It is a consequence of removing the Both construct, that it is now unequivocal which Pay goes first: they are all sequential, thus aiding analysis. With the Both construct we could potentially have Pays happen in any order (since both sides of Both are supposed to run concurrently).

Multi-clause When

We have modified When to include a set of possible actions that can be input while When waits. We call this approach “One of Many”, because it accepts one action out of potentially many allowed actions. When remains as follows:

When [Case] Timeout Contract

where When will either wait for Timeout and continue as Contract, or continue as specified in one of the Cases, whichever happens first. Case is defined as:

data Case = Case Action Contract

and Action as:

data Action = Deposit Party Party Token Value
            | Choice ChoiceId [Bound]
            | Notify Observation

A Case clause will be activated only if the corresponding Action is produced, and it will continue as Contract. In case of two Actions matching, the first one in the list will be executed.

Three kinds of actions are supported:

  • Deposit represents a deposit of money into an account; this was originally called Commit.

  • Choice represents a choice made by a participant from within a set of Integer values (specified by the list of Bounds).

  • Notify will wait for a Notify action issued when the Observation is true. We call it Notify in order to make it clear that we cannot just wait for Observations, but that someone must trigger the contract in a moment when an Observation is true.

We have discarded adding observations to Deposit and Choice since it would not be obvious whether the Observation would be evaluated before or after applying the action.

In addition to explicit cases in When, we must remember that the timeout branch is also a case, and it also needs to be triggered (similarly to Notify). 3 4

Observations and Values

We have discarded Observations and Values that can be expressed by combining others: like the general AvailableMoney (for the whole contract), or like DepositedMoneyBy, that remembers the amount of money deposited by a participant, since the contract can be restructured to observe that, and supporting would require additional information in the state (simplicity).

We have retained the ChoseSomething observation, even though, in the proposed semantics, every occurrence of ChoseSomething can be evaluated statically and efficiently by examining its context.

For example, in the following contract we can see that the first occurrence of ChoseSomething will evaluate to True, and the second one to False:

When [ Case (Choice (ChoiceId 1 Alice) [(1,1)])
            (If (ChoseSomething (ChoiceId 1 Alice))
     , Case (Choice (ChoiceId 2 Bob) [(2,2)])
            (If (ChoseSomething (ChoiceId 1 Alice))

Nevertheless, we have chosen to keep the construct for two reasons:

  • It allows for code reusability (convenience). For example, in the previous contract, we could define chosen1:

let chosen1 = If (ChoseSomething (ChoiceId 1 1))
When [ Case (Choice (ChoiceId 1 1) [(1,1)])
     , Case (Choice (ChoiceId 2 2) [(2,2)])

But this would not be possible if we did not have the construct ChoseSomething, since the value to which it reduces depends on the context.

  • It may no longer be the case that occurrences of the construct can be evaluated statically if we extend the When construct to support “many of many” inputs.

Inclusion of TimeIntervals

The EUTxO specification provides validation scripts with time-intervals instead of with slot numbers. This is to promote determinism in validation scripts. Nevertheless, we have kept the timeout of When (the only timeout) as a single timestamp even though it is now represented using POSIX time. The way we deal with time-intervals is by requiring that the interval of a transaction does not include any timeout over which the semantics has to make a choice. For example: if a timeout is noon of the 3rd of January, a transaction with interval 2nd of January to 4th of January will fail with AmbiguousTimeInterval. Participants would have to issue a transaction with interval from 2nd of January till one second before noon or one from noon on the 3rd of January to the 4th (or both).

Nevertheless, for Values, we provide the two constructs TimeIntervalStart and TimeIntervalEnd. An alternative to consider would be to modify the semantics so that Values are non-deterministic, that way we could include a CurrentTime construct and just invalidate transactions that are ambiguous, but this would complicate the semantics and make them less predictable.


We can potentially provide a way of statically analysing the contract to check whether there can possibly be any money left in any account when Close is reached.


This means that payments now obey a “push” model rather than a “pull” model.


Nevertheless, triggering the contract for processing timeouts is not urgent as it is with Notify, because while Observations can alternate between True and False, timeouts can only happen once and, independently of whether they have been observed by the contract or not, they cannot be reversed.


Indeed, an explicit Case can no longer be issued after the timeout, even if the timeout has not been observed by the contract, since the timeout is checked before the Inputs. However, a participant may want to trigger a timeout in cases where no other Inputs are needed, in order to trigger one or more payments, for example. In the current implementation of the semantics that would be done by issuing a transaction with an empty list of Inputs.