[bitcoin-dev] A Calculus of Covenants

[bitcoin-dev] A Calculus of Covenants

[Jeremy Rubin]

[-- Attachment #1: Type: text/plain, Size: 6687 bytes --]

Sharing below a framework for thinking about covenants. It is most useful
for modeling local covenants, that is, covenants where only one coin must
be examined, and not multi-coin covenants whereby you could have issues
with protocol forking requiring a more powerful stateful prover. It's the
model I use in Sapio.

I define a covenant primitive as follows:

1) A set of sets of transaction intents (a *family)*, potentially recursive
or co-recursive (e.g., the types of state transitions that can be
generated). These intents can also be represented by a language that
generates the transactions, rather than the literal transactions
themselves. We do the family rather than just sets at this level because to
instantiate a covenant we must pick a member of the family to use.
2) A verifier generator function that generates a function that accepts an
intent that is any element of one member of the family of intents and a
proof for it and rejects others.
3) A prover generator function that generates a function that takes an
intent that is any element of one member of the family and some extra data
and returns either a new prover function, a finished proof, or a rejection
(if not a valid intent).
4) A set of proofs that the Prover, Verifier, and a set of intents are
"impedance matched", that is, all statements the prover can prove and all
statements the verifier can verify are one-to-one and onto (or something
similar), and that this also is one-to-one and onto with one element of the
intents (a set of transactions) and no other.
5) A set of assumptions under which the covenant is verified (e.g., a
multi-sig covenant with at least 1-n honesty, a multisig covenant with any
3-n honesty required, Sha256 collision resistance, DLog Hardness, a SGX
module being correct).

To instantiate a covenant, the user would pick a particular element of the
set of sets of transaction intents. For example, in TLUV payment pool, it
would be the set of all balance adjusting transactions and redemptions. *Note,
we can 'cleave' covenants into separate bits -- e.g. one TLUV + some extra
CTV paths can be 'composed', but the composition is not guaranteed to be
well formed.*

Once the user has a particular intent, they then must generate a verifier
which can receive any member of the set of intents and accept it, and
receive any transaction outside the intents and reject it.

With the verifier in hand (or at the same time), the user must then
generate a prover function that can make a proof for any intent that the
verifier will accept. This could be modeled as a continuation system (e.g.,
multisig requires multiple calls into the prover), or it could be
considered to be wrapped as an all-at-once function. The prover could be
done via a multi-sig in which case the assumptions are stronger, but it
still should be well formed such that the signers can clearly and
unambiguously sign all intents and reject all non intents, otherwise the
covenant is not well formed.

The proofs of validity of the first three parts and the assumptions for
them should be clear, but do not require generation for use. However,
covenants which do not easily permit proofs are less useful.

We now can analyze three covenants under this, plain CTV, 2-3 online
multisig, 3-3 presigned + deleted.

CTV:
1) Intent sets: the set of specific next transactions, with unbound inputs
into it that can be mutated (but once the parent is known, can be filled in
for all children).
2) Verifier: The transaction has the hash of the intent
3) Prover: The transaction itself and no other work
4) Proofs of impedance: trivial.
5) Assumptions: sha256
6) Composition: Any two CTVs can be OR'd together as separate leafs

2-3 Multisig:
1) Intent: All possible sets of transactions, one set selected per instance
2) Verifier: At least 2 signed the transition
3) Prover: Receive some 'state' in the form of business logic to enforce,
only sign if that is satisfied. Produce a signature.
4) Impedance: The business logic must cover the instance's Intent set and
must not be able to reach any other non-intent
5) Assumptions: at least 2 parties are 'honest' for both liveness and for
correctness, and the usual suspects (sha256, schnorr, etc)
6) Composition: Any two groups can be OR'd together, if the groups have
different signers, then the assumptions expand

3-3 Presigned:
Same as CTV except:
5) Assumptions: at least one party deletes their key after signing


 You can also think through other covenants like TLUV in this model.

One useful question is the 'cardinality' of an intent set. The useful
notion of this is both in magnitude but also contains. Obviously, many of
these are infinite sets, but if one set 'contains' another then it is
definitionally more powerful. Also, if a set of transitions is 'bigger'
(work to do on what that means?) than another it is potentially more
powerful.

Another question is around composition of different covenants inside of an
intent -- e.g., a TLUV that has a branch with a CTV or vice versa. We
consider this outside the model, analysis should be limited to "with only
these covenants what could you build". Obviously, one recursive primitive
makes all primitives recursive.

Another question is 'unrollability'. Can the intents, and the intents of
the outputs of the intents, be unrolled into a representation for a
specific instantiation? Or is that set of possible transactions infinite?
How infinite? CTV is, e.g., unrollable.


Last note on statefulness: The above has baked into it a notion of
'statelessness', but it's very possible and probably required that provers
maintain some external state in order to prove (whether multisig or not).
E.g., a multisig managing an account model covenant may need to track who
is owed what. This data can sometimes be put e.g. in an op return, an extra
tapleaf branch, or just considered exogenous to the covenant. But the idea
that a prover isn't just deciding on what to do based on purely local
information to an output descriptor is important.


For Sapio in particular, this framework is useful because if you can answer
the above questions on intents, and prover/verifier generators, then you
would be able to generate tooling that could integrate your covenant into
Sapio and have things work nicely. If you can't answer these questions (in
code?) then your covenant might not be 'well formed'. The efficiency of a
prover or verifier is out of scope of this framework, which focuses on the
engineering + design, but can also be analyzed.


Grateful for any and all feedback on this model and if there are examples
that cannot be described within it,

Jeremy




--
@JeremyRubin <https://twitter.com/JeremyRubin>

[-- Attachment #2: Type: text/html, Size: 12332 bytes --]

Re: [bitcoin-dev] A Calculus of Covenants

[Jeremy Rubin]

[-- Attachment #1: Type: text/plain, Size: 7757 bytes --]

note of clarification:

this is from the perspective of a developer trying to build infrastructure
for covenants. from the perspective of bitcoin consensus, a covenant
enforcing primitve would be something like OP_TLUV and less so it's use in
conjunction with other opcodes, e.g. OP_AMOUNT.

One must also analyze all the covenants that one *could* author using a
primitive, in some sense, to demonstrate that our understanding is
sufficient. As a trivial example, you could use
OP_DELETE_BITCOIN_ENTIRELY_IF_KNOWS_PREIMAGE_TO_X_OR_TLUV and just because
you could use it safely for TLUV would not mean we should add that opcode
if there's some way of using it negatively.

Cheers,

Jeremy
--
@JeremyRubin <https://twitter.com/JeremyRubin>


On Tue, Apr 12, 2022 at 10:33 AM Jeremy Rubin <jeremy.l.rubin@gmail.com>
wrote:

> Sharing below a framework for thinking about covenants. It is most useful
> for modeling local covenants, that is, covenants where only one coin must
> be examined, and not multi-coin covenants whereby you could have issues
> with protocol forking requiring a more powerful stateful prover. It's the
> model I use in Sapio.
>
> I define a covenant primitive as follows:
>
> 1) A set of sets of transaction intents (a *family)*, potentially
> recursive or co-recursive (e.g., the types of state transitions that can be
> generated). These intents can also be represented by a language that
> generates the transactions, rather than the literal transactions
> themselves. We do the family rather than just sets at this level because to
> instantiate a covenant we must pick a member of the family to use.
> 2) A verifier generator function that generates a function that accepts an
> intent that is any element of one member of the family of intents and a
> proof for it and rejects others.
> 3) A prover generator function that generates a function that takes an
> intent that is any element of one member of the family and some extra data
> and returns either a new prover function, a finished proof, or a rejection
> (if not a valid intent).
> 4) A set of proofs that the Prover, Verifier, and a set of intents are
> "impedance matched", that is, all statements the prover can prove and all
> statements the verifier can verify are one-to-one and onto (or something
> similar), and that this also is one-to-one and onto with one element of the
> intents (a set of transactions) and no other.
> 5) A set of assumptions under which the covenant is verified (e.g., a
> multi-sig covenant with at least 1-n honesty, a multisig covenant with any
> 3-n honesty required, Sha256 collision resistance, DLog Hardness, a SGX
> module being correct).
>
> To instantiate a covenant, the user would pick a particular element of the
> set of sets of transaction intents. For example, in TLUV payment pool, it
> would be the set of all balance adjusting transactions and redemptions. *Note,
> we can 'cleave' covenants into separate bits -- e.g. one TLUV + some extra
> CTV paths can be 'composed', but the composition is not guaranteed to be
> well formed.*
>
> Once the user has a particular intent, they then must generate a verifier
> which can receive any member of the set of intents and accept it, and
> receive any transaction outside the intents and reject it.
>
> With the verifier in hand (or at the same time), the user must then
> generate a prover function that can make a proof for any intent that the
> verifier will accept. This could be modeled as a continuation system (e.g.,
> multisig requires multiple calls into the prover), or it could be
> considered to be wrapped as an all-at-once function. The prover could be
> done via a multi-sig in which case the assumptions are stronger, but it
> still should be well formed such that the signers can clearly and
> unambiguously sign all intents and reject all non intents, otherwise the
> covenant is not well formed.
>
> The proofs of validity of the first three parts and the assumptions for
> them should be clear, but do not require generation for use. However,
> covenants which do not easily permit proofs are less useful.
>
> We now can analyze three covenants under this, plain CTV, 2-3 online
> multisig, 3-3 presigned + deleted.
>
> CTV:
> 1) Intent sets: the set of specific next transactions, with unbound inputs
> into it that can be mutated (but once the parent is known, can be filled in
> for all children).
> 2) Verifier: The transaction has the hash of the intent
> 3) Prover: The transaction itself and no other work
> 4) Proofs of impedance: trivial.
> 5) Assumptions: sha256
> 6) Composition: Any two CTVs can be OR'd together as separate leafs
>
> 2-3 Multisig:
> 1) Intent: All possible sets of transactions, one set selected per instance
> 2) Verifier: At least 2 signed the transition
> 3) Prover: Receive some 'state' in the form of business logic to enforce,
> only sign if that is satisfied. Produce a signature.
> 4) Impedance: The business logic must cover the instance's Intent set and
> must not be able to reach any other non-intent
> 5) Assumptions: at least 2 parties are 'honest' for both liveness and for
> correctness, and the usual suspects (sha256, schnorr, etc)
> 6) Composition: Any two groups can be OR'd together, if the groups have
> different signers, then the assumptions expand
>
> 3-3 Presigned:
> Same as CTV except:
> 5) Assumptions: at least one party deletes their key after signing
>
>
>  You can also think through other covenants like TLUV in this model.
>
> One useful question is the 'cardinality' of an intent set. The useful
> notion of this is both in magnitude but also contains. Obviously, many of
> these are infinite sets, but if one set 'contains' another then it is
> definitionally more powerful. Also, if a set of transitions is 'bigger'
> (work to do on what that means?) than another it is potentially more
> powerful.
>
> Another question is around composition of different covenants inside of an
> intent -- e.g., a TLUV that has a branch with a CTV or vice versa. We
> consider this outside the model, analysis should be limited to "with only
> these covenants what could you build". Obviously, one recursive primitive
> makes all primitives recursive.
>
> Another question is 'unrollability'. Can the intents, and the intents of
> the outputs of the intents, be unrolled into a representation for a
> specific instantiation? Or is that set of possible transactions infinite?
> How infinite? CTV is, e.g., unrollable.
>
>
> Last note on statefulness: The above has baked into it a notion of
> 'statelessness', but it's very possible and probably required that provers
> maintain some external state in order to prove (whether multisig or not).
> E.g., a multisig managing an account model covenant may need to track who
> is owed what. This data can sometimes be put e.g. in an op return, an extra
> tapleaf branch, or just considered exogenous to the covenant. But the idea
> that a prover isn't just deciding on what to do based on purely local
> information to an output descriptor is important.
>
>
> For Sapio in particular, this framework is useful because if you can
> answer the above questions on intents, and prover/verifier generators, then
> you would be able to generate tooling that could integrate your covenant
> into Sapio and have things work nicely. If you can't answer these questions
> (in code?) then your covenant might not be 'well formed'. The efficiency of
> a prover or verifier is out of scope of this framework, which focuses on
> the engineering + design, but can also be analyzed.
>
>
> Grateful for any and all feedback on this model and if there are examples
> that cannot be described within it,
>
> Jeremy
>
>
>
>
> --
> @JeremyRubin <https://twitter.com/JeremyRubin>
>

[-- Attachment #2: Type: text/html, Size: 14520 bytes --]

Re: [bitcoin-dev] A Calculus of Covenants

[Keagan McClelland]

[-- Attachment #1: Type: text/plain, Size: 10249 bytes --]

> One must also analyze all the covenants that one *could* author using a
primitive

So as I've been contemplating this more, I'm realizing that a calculus of
covenants themselves may not make as much sense as a broader calculus of
Bitcoin transactions as a whole. I think this comment that you made in your
followup solidified that position. If you have to analyze it in the context
of all of the other opcodes that could potentially interact with it, you
don't really have a closed algebra that you can really try to understand
and evaluate. I'm still ruminating on what such a calculus would be, but it
also makes me more convinced that Simplicity gets a lot right here. That
said, there is probably an opportunity for a significantly more domain
specific set of primitives than what simplicity offers that would allow you
similar practical use cases but with a much more high level analysis.

The way I think about this now is that most of the primitives in the
Bitcoin script VM right now are constraints on the witness, you have a
couple of opcodes that are constraints on the chain state, and then
covenants are really a constraint on the body of the transaction that
spends an input. I think most of the time we imagine covenants of output
constraints but you can also imagine a hypothetical covenant that says,
"this input may not be spent alongside any other inputs". This is still a
constraint on the spending transaction despite the fact that it mentions
nothing of the outputs, and I would still broadly think of this as a
covenant. I think depending on how you define "family" and "state
transition" it would tolerate this distinction. However, it definitely
complicates the question of things like unrollability. Is a covenant that
permits any output(s) but the input must be spent alone unrollable? Does
the concept unrollable even make any sense when you aren't constraining the
outputs?

These thoughts aren't completely baked but I figured I'd jot them down
while I was thinking about it.

Keagan

On Tue, Apr 12, 2022 at 9:04 AM Jeremy Rubin via bitcoin-dev <
bitcoin-dev@lists.linuxfoundation.org> wrote:

> note of clarification:
>
> this is from the perspective of a developer trying to build infrastructure
> for covenants. from the perspective of bitcoin consensus, a covenant
> enforcing primitve would be something like OP_TLUV and less so it's use in
> conjunction with other opcodes, e.g. OP_AMOUNT.
>
> One must also analyze all the covenants that one *could* author using a
> primitive, in some sense, to demonstrate that our understanding is
> sufficient. As a trivial example, you could use
> OP_DELETE_BITCOIN_ENTIRELY_IF_KNOWS_PREIMAGE_TO_X_OR_TLUV and just because
> you could use it safely for TLUV would not mean we should add that opcode
> if there's some way of using it negatively.
>
> Cheers,
>
> Jeremy
> --
> @JeremyRubin <https://twitter.com/JeremyRubin>
>
>
> On Tue, Apr 12, 2022 at 10:33 AM Jeremy Rubin <jeremy.l.rubin@gmail.com>
> wrote:
>
>> Sharing below a framework for thinking about covenants. It is most useful
>> for modeling local covenants, that is, covenants where only one coin must
>> be examined, and not multi-coin covenants whereby you could have issues
>> with protocol forking requiring a more powerful stateful prover. It's the
>> model I use in Sapio.
>>
>> I define a covenant primitive as follows:
>>
>> 1) A set of sets of transaction intents (a *family)*, potentially
>> recursive or co-recursive (e.g., the types of state transitions that can be
>> generated). These intents can also be represented by a language that
>> generates the transactions, rather than the literal transactions
>> themselves. We do the family rather than just sets at this level because to
>> instantiate a covenant we must pick a member of the family to use.
>> 2) A verifier generator function that generates a function that accepts
>> an intent that is any element of one member of the family of intents and a
>> proof for it and rejects others.
>> 3) A prover generator function that generates a function that takes an
>> intent that is any element of one member of the family and some extra data
>> and returns either a new prover function, a finished proof, or a rejection
>> (if not a valid intent).
>> 4) A set of proofs that the Prover, Verifier, and a set of intents are
>> "impedance matched", that is, all statements the prover can prove and all
>> statements the verifier can verify are one-to-one and onto (or something
>> similar), and that this also is one-to-one and onto with one element of the
>> intents (a set of transactions) and no other.
>> 5) A set of assumptions under which the covenant is verified (e.g., a
>> multi-sig covenant with at least 1-n honesty, a multisig covenant with any
>> 3-n honesty required, Sha256 collision resistance, DLog Hardness, a SGX
>> module being correct).
>>
>> To instantiate a covenant, the user would pick a particular element of
>> the set of sets of transaction intents. For example, in TLUV payment pool,
>> it would be the set of all balance adjusting transactions and redemptions. *Note,
>> we can 'cleave' covenants into separate bits -- e.g. one TLUV + some extra
>> CTV paths can be 'composed', but the composition is not guaranteed to be
>> well formed.*
>>
>> Once the user has a particular intent, they then must generate a verifier
>> which can receive any member of the set of intents and accept it, and
>> receive any transaction outside the intents and reject it.
>>
>> With the verifier in hand (or at the same time), the user must then
>> generate a prover function that can make a proof for any intent that the
>> verifier will accept. This could be modeled as a continuation system (e.g.,
>> multisig requires multiple calls into the prover), or it could be
>> considered to be wrapped as an all-at-once function. The prover could be
>> done via a multi-sig in which case the assumptions are stronger, but it
>> still should be well formed such that the signers can clearly and
>> unambiguously sign all intents and reject all non intents, otherwise the
>> covenant is not well formed.
>>
>> The proofs of validity of the first three parts and the assumptions for
>> them should be clear, but do not require generation for use. However,
>> covenants which do not easily permit proofs are less useful.
>>
>> We now can analyze three covenants under this, plain CTV, 2-3 online
>> multisig, 3-3 presigned + deleted.
>>
>> CTV:
>> 1) Intent sets: the set of specific next transactions, with unbound
>> inputs into it that can be mutated (but once the parent is known, can be
>> filled in for all children).
>> 2) Verifier: The transaction has the hash of the intent
>> 3) Prover: The transaction itself and no other work
>> 4) Proofs of impedance: trivial.
>> 5) Assumptions: sha256
>> 6) Composition: Any two CTVs can be OR'd together as separate leafs
>>
>> 2-3 Multisig:
>> 1) Intent: All possible sets of transactions, one set selected per
>> instance
>> 2) Verifier: At least 2 signed the transition
>> 3) Prover: Receive some 'state' in the form of business logic to enforce,
>> only sign if that is satisfied. Produce a signature.
>> 4) Impedance: The business logic must cover the instance's Intent set and
>> must not be able to reach any other non-intent
>> 5) Assumptions: at least 2 parties are 'honest' for both liveness and for
>> correctness, and the usual suspects (sha256, schnorr, etc)
>> 6) Composition: Any two groups can be OR'd together, if the groups have
>> different signers, then the assumptions expand
>>
>> 3-3 Presigned:
>> Same as CTV except:
>> 5) Assumptions: at least one party deletes their key after signing
>>
>>
>>  You can also think through other covenants like TLUV in this model.
>>
>> One useful question is the 'cardinality' of an intent set. The useful
>> notion of this is both in magnitude but also contains. Obviously, many of
>> these are infinite sets, but if one set 'contains' another then it is
>> definitionally more powerful. Also, if a set of transitions is 'bigger'
>> (work to do on what that means?) than another it is potentially more
>> powerful.
>>
>> Another question is around composition of different covenants inside of
>> an intent -- e.g., a TLUV that has a branch with a CTV or vice versa. We
>> consider this outside the model, analysis should be limited to "with only
>> these covenants what could you build". Obviously, one recursive primitive
>> makes all primitives recursive.
>>
>> Another question is 'unrollability'. Can the intents, and the intents of
>> the outputs of the intents, be unrolled into a representation for a
>> specific instantiation? Or is that set of possible transactions infinite?
>> How infinite? CTV is, e.g., unrollable.
>>
>>
>> Last note on statefulness: The above has baked into it a notion of
>> 'statelessness', but it's very possible and probably required that provers
>> maintain some external state in order to prove (whether multisig or not).
>> E.g., a multisig managing an account model covenant may need to track who
>> is owed what. This data can sometimes be put e.g. in an op return, an extra
>> tapleaf branch, or just considered exogenous to the covenant. But the idea
>> that a prover isn't just deciding on what to do based on purely local
>> information to an output descriptor is important.
>>
>>
>> For Sapio in particular, this framework is useful because if you can
>> answer the above questions on intents, and prover/verifier generators, then
>> you would be able to generate tooling that could integrate your covenant
>> into Sapio and have things work nicely. If you can't answer these questions
>> (in code?) then your covenant might not be 'well formed'. The efficiency of
>> a prover or verifier is out of scope of this framework, which focuses on
>> the engineering + design, but can also be analyzed.
>>
>>
>> Grateful for any and all feedback on this model and if there are examples
>> that cannot be described within it,
>>
>> Jeremy
>>
>>
>>
>>
>> --
>> @JeremyRubin <https://twitter.com/JeremyRubin>
>>
> _______________________________________________
> bitcoin-dev mailing list
> bitcoin-dev@lists.linuxfoundation.org
> https://lists.linuxfoundation.org/mailman/listinfo/bitcoin-dev
>

[-- Attachment #2: Type: text/html, Size: 18265 bytes --]