speckl/docs/whitepaper-v2.md

Speckl: Closing the Spec-Code Gap with Compilable Specifications

A whitepaper on SpeckDL, a domain-specific language for writing specifications that compile to runnable software with embedded provenance.

Abstract

Software specifications have a fundamental problem: they describe what code should do, but they cannot guarantee what code actually does. Formal methods (TLA+, Alloy, Coq) can verify properties of specifications, but their outputs remain disconnected from production systems — the spec and the code evolve independently until the spec becomes archaeology.

Speckl introduces a different approach: specifications that compile to software. The SpeckDL language lets you write state machine specifications with types, invariants, and actions — then compiles them into five auditable artifacts: PROV-O provenance graphs, CycloneDX SBOMs, SPDX license documents, TypeScript classes, and WebAssembly modules. The spec isn't documentation alongside the code — the spec IS the code, with a verifiable chain from intent to implementation.

This paper presents the SpeckDL language, its compiler architecture, and the provenance model that makes every generated artifact auditable back to its specification. We demonstrate SpeckDL on consensus protocols (Two-Phase Commit, Paxos, Raft) and a financial ledger, showing how the same specification produces both formal verification artifacts and production-runnable modules.

1. The Spec-Code Gap

Every software project has specifications. Requirements documents, architecture decision records, API contracts, RFCs — these artifacts describe intent. But they have a fatal flaw: they cannot be executed, and therefore cannot be verified against reality.

1.1 The Archaeology Problem

When a specification exists only as prose, it decays. Three months after writing, the spec describes what the system should have done, not what it does now. Engineers treat specs as orientation, not authority. The spec becomes archaeology — interesting for historical context, irrelevant for current behavior.

This decay compounds with AI-generated code. When an LLM writes code from a prompt, the reasoning chain is opaque. The generated code may satisfy the prompt, violate it in subtle ways, or introduce behaviors the prompt never specified. Without an executable connection between intent and output, verification is reduced to testing — which can only prove the presence of bugs, not their absence.

1.2 The Formal Methods Gap

Formal specification languages (TLA+, Alloy, Isabelle/HOL) solve the precision problem: they express intent unambiguously and can model-check properties of that intent. TLA+ in particular has been used to find critical design bugs in systems at AWS, Microsoft, and MongoDB.

But formal methods create their own gap: the verified spec and the running system are separate artifacts. After model-checking a TLA+ specification, someone must manually translate it into a programming language. That translation introduces the same spec-code gap the formal method was supposed to close. The implementation can diverge from the verified spec, and no tool detects it.

1.3 The SpeckDL Approach

SpeckDL sits between prose specifications and formal methods. It provides:

1. Executable semantics: Specifications compile to runnable code (TypeScript classes, WebAssembly modules). The spec doesn't describe the system — the spec IS the system, compiled directly.
2. Verifiable invariants: State invariants are checked at runtime through guard statements (require), providing living verification rather than one-time model checking.
3. Embedded provenance: Every compilation produces a PROV-O graph tracing each output artifact back to the specification elements that generated it. You can answer "why does this code exist?" by following the provenance chain.
4. Supply chain transparency: The compiler emits CycloneDX SBOMs and SPDX documents automatically, making AI-generated code auditable for compliance frameworks (NIST SA-11, SOC 2, FedRAMP).

1.4 A Concrete Example

Consider a ToggleSwitch specification in SpeckDL:

state ToggleSwitch {
    on: Bool
}

init ToggleSwitch {
    on = false
}

action TurnOn for ToggleSwitch {
    on = true
}

action TurnOff for ToggleSwitch {
    on = false
}

invariant NoDoubleToggle for ToggleSwitch {
    require TurnOn => !on  // Can't turn on if already on
    require TurnOff => on  // Can't turn off if already off
}

From these 14 lines, speckl compile produces:

• TypeScript: A ToggleSwitch class with typed TurnOn() and TurnOff() methods, guard checks, and state accessors
• WebAssembly: A .wasm module with exported turn_on and turn_off functions with guard-validated state transitions
• PROV-O: A provenance graph showing that the TurnOn method was derived from the action TurnOn spec block
• CycloneDX: An SBOM declaring the generated module with its hash and provenance
• SPDX: A license document for the generated artifact

The specification is 14 lines. The generated artifacts are hundreds of lines of production code. The provenance chain connects every line back to the 14-line intent. This is the spec-code gap, closed.

2. The SpeckDL Language

SpeckDL is a domain-specific language for defining state machines with types, invariants, and actions. It sits between formal specification languages (TLA+) and general-purpose programming languages — more structured than prose, more constrained than code, and designed from the ground up to be compilable.

2.1 Design Philosophy

SpeckDL follows four principles:

1. Spec as source of truth. The .speck file is the authoritative description of system behavior. Generated code is derived, not co-equal. If the spec and the code disagree, the spec wins — and the compiler tells you they disagree.

2. State machines, not procedures. Every SpeckDL specification models a state machine: typed state variables, guarded transitions (actions), and invariants that must hold across every transition. This is the same mental model as TLA+, but with a concrete compilation target.

3. Type-safe by construction. SpeckDL has a static type system with primitive types (Bool, Nat, Int, String), compound types (Set, Map, List, Records), and user-defined type aliases. Every expression is type-checked before code generation.

4. Emit for observability. Actions can emit typed events that become part of the provenance graph and generated event logs. Events make state machine execution observable without requiring external instrumentation.

2.2 Structure of a Speck

A SpeckDL file contains one or more speck blocks. A speck defines a named state machine:

speck Name {
    type ...          // optional type aliases
    state { ... }     // state variables
    init { ... }      // initial state
    invariant ...     // safety properties
    action ...        // state transitions
    next: ...         // action composition
}

Types

Type aliases are declared at the top of a speck using the type keyword. They name compound types for reuse across state declarations and action signatures:

type AccountId = Nat
type Transfer = {
    from: AccountId,
    to: AccountId,
    amount: Nat
}

Type aliases are inlined during compilation — they exist for readability, not runtime overhead.

State

The state block declares the typed variables that comprise the machine's state. Every state variable must have an explicit type:

state {
    counter: Nat
    users: Set(String)
    balances: Map(Nat, Int)
    history: List(Event)
    leader: { id: String, term: Nat }
}

State variables are the only mutable storage in SpeckDL. There are no global variables, no hidden state, no ambient context. Everything a speck can observe or modify is declared in its state block.

Init

The init block sets the initial values of all state variables. Init expressions use equality (==) rather than assignment (:=) — they describe what must be true at startup, not a sequence of operations:

init {
    counter == 0
    users == emptySet
    balances == emptyMap
    history == List.empty
    leader == { id: "", term: 0 }
}

If any state variable is missing from init, the compiler reports an error. Every variable must have an initial value.

2.3 Type System

SpeckDL supports the following types:

Type	Syntax	Description
Boolean	Bool	True or false
Natural number	Nat	Non-negative integer (≥ 0)
Integer	Int	Signed integer
String	String	UTF-8 text
Set	Set(T)	Unordered collection of distinct elements of type T
Map	Map(K, V)	Key-value mapping from type K to type V
List	List(T)	Ordered sequence of elements of type T
Record	{ field: T, ... }	Named fields with typed values
Optional	T | null	A value of type T, or null

Sets support union, intersection, \ (difference), subsetOf, in, notIn, and size(). The literal emptySet represents the empty set. Set literals use curly braces: {"a", "b", "c"}.

Maps support key access (m[k]), key assignment (m[k] := v), .keys, .values, and in for key membership. The literal emptyMap represents the empty map. Maps have no literal constructor — they are built incrementally through assignments.

Lists support indexed access (list[i], 1-based), length(), take(), concat(), and List.empty for the empty list.

Records are anonymous structural types. Two records are the same type if they have the same field names and types, regardless of declaration order:

type Point = { x: Nat, y: Nat }
// Equivalent to { y: Nat, x: Nat }

Record values are constructed with braces: { x: 3, y: 5 }. Field access uses dot notation: point.x.

Optional types use the | null syntax. A field typed String | null can hold either a string or the null value. Null checks use == null and != null in expressions.

2.4 Actions

Actions are the transitions of the state machine. Each action has a name, optional parameters, preconditions (require), a body of state mutations, and optional return/emit statements:

action Transfer(from: Nat, to: Nat, amount: Nat) {
    require from in accounts
    require to in accounts
    require balances[from] >= amount

    balances[from] := balances[from] - amount
    balances[to] := balances[to] + amount
    emit TransferCompleted { from: from, to: to, amount: amount }
    return balances[from]
}

Preconditions (require)

require statements are boolean expressions that must be true for the action to execute. If any require fails, the action aborts and no state is modified. Requires are the runtime enforcement of invariants — they're compiled to guard clauses in generated code.

In the generated TypeScript, a failed require throws a PreconditionError. In the generated WASM, it traps.

State Mutation (assignment)

State variables are modified with the := operator. Assignments can target:

• Simple variables: counter := counter + 1
• Map entries: balances[id] := 100
• Set membership: users := users union {newUser}
• Record fields (in expressions, not directly assigned): balances[pt.to] := balances[pt.to] + pt.amount

All assignments within an action are atomic — either all succeed (the action completes) or none apply (a require failed).

Let Bindings

let binds an expression result to a local name for use later in the action body. Let bindings are immutable:

action CommitTransfer(tid: Nat) {
    require tid in pendingTransfers.keys

    let pt := pendingTransfers[tid]
    balances[pt.to] := balances[pt.to] + pt.amount
    pendingTransfers := pendingTransfers \ {tid}
}

Let bindings reduce repetition and make action bodies more readable. They are scoped to the enclosing action.

If Statements

SpeckDL supports conditional logic within actions:

action GrantVote(voter: String, req: RequestVoteReq) {
    require req.term >= serverState[voter].currentTerm

    if req.term > serverState[voter].currentTerm {
        serverState[voter].currentTerm := req.term
        serverState[voter].role := "follower"
    }

    serverState[voter].votedFor := req.candidateId
}

If conditions must be boolean expressions. There are no else-if chains (nest if statements for multi-branch logic).

Emit

emit produces a typed event that becomes part of the provenance graph. Events are structured records:

emit TransferCreated { id: tid, from: from, to: to, amount: amount }

Emitted events serve three purposes:

1. They make state machine execution traceable.
2. They appear in generated PROV-O graphs as activities.
3. In generated TypeScript, they're emitted via an EventEmitter interface for integration with logging, metrics, and audit systems.

Return

return produces an optional output value from the action. Actions that don't call return implicitly return void. Return values are typed according to the action's signature or inferred from the return expression.

Forall Loops

Actions can iterate over sets with forall loops:

action TMCommit {
    require tmPrepared == rms

    tmState := "committed"
    forall rm in rms:
        rmState[rm] := "committed"
}

The loop variable (rm) is read-only. Forall loops iterate over the entire collection; there is no break or early termination.

2.5 Invariants

Invariants are properties that must hold in every reachable state of the machine. They are declared with the invariant keyword:

invariant NoNegativeBalances {
    forall id in accounts:
        balances[id] >= 0
}

Invariants serve two roles:

1. Documentation. They state the safety properties the system guarantees, in unambiguous, executable form.
2. Runtime verification. In generated TypeScript and WASM, invariants are checked after every action execution. A violated invariant throws a InvariantViolationError (TypeScript) or traps (WASM).

Quantified Invariants

SpeckDL supports universal quantification (forall) in invariants:

invariant SingleLeaderPerTerm {
    forall s1 in servers:
        forall s2 in servers:
            s1 != s2 and
            serverState[s1].role == "leader" and
            serverState[s2].role == "leader" implies
                serverState[s1].currentTerm != serverState[s2].currentTerm
}

Quantified invariants are compiled to nested loops in generated code. For finite state spaces (the common case), this provides tractable runtime verification. For large or infinite state spaces, invariants can be selectively disabled at compile time (--skip-invariants) and verified through external model checking.

Logical Operators

Invariants and require expressions use C-like logical operators:

Operator	Meaning
==, !=	Equality, inequality
<, <=, >, >=	Comparison
+, -, *, /	Arithmetic
and, or, not	Logical conjunction, disjunction, negation
implies	Logical implication (A implies B)
in, notIn	Set/map membership
subsetOf	Set subset relation
union, \	Set union, set difference
size()	Set cardinality
length()	List length

2.6 Next — Action Composition

The next declaration defines which actions can follow which states. It's the state machine's transition relation:

next: OpenAccount | CreateTransfer | CommitTransfer | VoidTransfer

This declares that at any point, any of these four actions may be invoked. SpeckDL does not currently support sequential action composition (action1 then action2) — the next operator | means "any of these, chosen non-deterministically." This matches the TLA+ model where the environment (user, scheduler, network) chooses the next action.

2.7 Complete Example: Two-Phase Commit

Here is a complete SpeckDL specification for the Two-Phase Commit protocol, demonstrating types, state, init, invariants, actions with require/emit/forall, and action composition:

speck TwoPhaseCommit {
    state {
        rmState: Map(String, String)
        tmState: String
        tmPrepared: Set(String)
        msgs: Set(Message)
        rms: Set(String)
    }

    init {
        let allRms := {"rm1", "rm2", "rm3"}

        rms == allRms
        rmState == mapKeys(allRms, "working")
        tmState == "init"
        tmPrepared == emptySet
        msgs == emptySet
    }

    invariant TPConsistency {
        tmState == "committed" implies
            forall rm in rms: rmState[rm] == "committed"
        tmState == "aborted" implies
            forall rm in rms: rmState[rm] == "aborted"
    }

    action RMPrepare(rm: String) {
        require rm in rms
        require rmState[rm] == "working"

        rmState[rm] := "prepared"
        msgs := msgs union {{ type: "Prepared", rm: rm }}
        emit Prepared { rm: rm }
    }

    action TMCommit {
        require tmState == "init"
        require tmPrepared == rms

        tmState := "committed"
        forall rm in rms:
            rmState[rm] := "committed"
        emit AllCommitted {}
    }

    action TMAbort {
        require tmState == "init"

        tmState := "aborted"
        forall rm in rms:
            if rmState[rm] != "committed":
                rmState[rm] := "aborted"
        emit Aborted {}
    }

    next: RMPrepare | TMCommit | TMAbort
}

This specification is 48 lines. The compiled output spans ~300 lines of TypeScript, ~200 lines of WAT (WebAssembly Text), and three machine-readable provenance documents. Every generated line traces back to these 48 lines of intent.

2.8 What SpeckDL Doesn't Have

It's instructive to note what SpeckDL deliberately omits:

No functions. There are no user-defined functions, no recursion, no higher-order abstractions. This is intentional — specifications should be flat and auditable. Complex logic belongs in the implementation, not the spec.

No loops (except forall). There are no while loops, no for loops with mutable state, no iterators with side effects. The only iteration construct is forall, which is side-effect-free in invariants and bounded in actions.

No concurrency primitives. SpeckDL models state machines sequentially. Concurrency is expressed through non-deterministic interleaving of actions (next: A | B | C), not through threads, locks, or channels. This is the same approach as TLA+.

No I/O. Actions can emit events and return values, but there is no file system access, no network calls, no database queries. The spec is pure computation over typed state.

No error handling. There are no try/catch blocks, no error values, no exception types. If a require fails, the action aborts. That's the only error path.

These omissions are the point. SpeckDL constrains the spec author to describe what the system does and what must be true — not how it does it, not how it handles failure, not how it optimizes. The implementation fills in those details.

3. Compiler Architecture

[Section 3: Parser → AST → Generators pipeline, multi-target compilation, test infrastructure]

4. Embedded Provenance

[Section 4: PROV-O generation, provenance graph structure, audit trail semantics]

5. Consensus Protocols in SpeckDL

[Section 5: Two-Phase Commit, Paxos, and Raft case studies with compilation results]

6. Compliance and Supply Chain

[Section 6: NIST SA-11 alignment, SBOM automation, regulatory use cases]

7. Related Work

[Section 7: Comparison with TLA+, Alloy, Coq, Dafny, and code generation approaches]

8. Future Work and Roadmap

[Section 8: Record types, temporal properties, IDE integration, ecosystem]

---

Whitepaper v2 draft. Sections 1 complete; 2-8 forthcoming.