What distinguishes ALGAL

ALGAL's useful combination is a typed program format, bounded execution, content-addressed evidence, and portable process checkpoints. A manifest is JSON data: the host can inspect, hash, compose, generate, and admit it before execution. Model judgments are bounded cells inside that program. A process can wait for an external message and continue from recorded effects under a new CLI invocation.

None of these individual ideas establishes uniqueness. Durable execution, effect replay, checkpointing, and approval workflows have substantial prior art. ALGAL's claim is narrower: the same small program and evidence contracts join program evolution, execution, and offline verification across its TypeScript and Rust runtimes.

A concrete use

The process VM demonstration runs an evidence review, records a proposal, exits, waits for an approval, and resumes without invoking the recorded decision again. Approval authority stays in a host-admitted mailbox; the model can recommend but cannot approve. A denied actor completes without publishing. The report measures actual subprocess invocations, verifies all completed actor receipt generations, and can demonstrate both directions of TypeScript/Rust handoff.

The fixture makes two decision invocations across two actors, versus four when the same program restarts without a checkpoint. It demonstrates avoided work under that restart policy, not better intelligence or superiority over another durable runtime. Its local command adapter returns scripted data and incurs no provider charges.

How the pieces reinforce each other

Programs are values. A manifest digest identifies the typed graph and its declared bounds. Generated candidates use the same admission path as hand-written programs, so evaluation and promotion can retain exact source identities instead of informal prompt versions.

Effects are explicit. Model calls and tools have recorded requests and outcomes. Replay reconstructs orchestration using those outcomes, and a resumed process only invokes effects beyond the recorded prefix.

Authority is admitted by the host. Capability ports retain their exact classes. A manifest cannot mint a mailbox handle from a constant. The host decides which tools and executors are available and which capabilities enter the process.

Execution evidence is portable. Process records, manifests, and receipts have canonical representations. The local store can be handed between the two runtime implementations, and recorded generations can be checked without the original provider.

The outer process is bounded too. A named process has a generation budget. Waiting does not require a model polling loop. Dispatch uncertainty remains visible rather than triggering an automatic external-action retry.

These properties are useful together when a model-generated workflow needs admission, durable execution, and later inspection. They are unnecessary machinery for many single-call applications.

Neighboring systems

System or abstraction	Established capability	ALGAL's design choice
Temporal workflows	Workflow histories reconstruct state and reuse recorded activity results during replay.	Typed JSON manifests and canonical execution receipts are the shared program/evidence format. ALGAL's local supervisor is much smaller in operational scope.
LangGraph persistence	Checkpointers preserve graph state for continuity, interruptions, human review, and fault tolerance.	A data-only graph contract and independent TypeScript/Rust runtimes support the demonstrated local handoff and offline replay.
WebAssembly isolation	Runtime-enforced memory and control-flow isolation restrict untrusted machine code.	ALGAL interprets bounded graph/expression data but does not isolate its host tools or command executors. An OS or Wasm sandbox is a separate host concern.

The comparisons identify design choices, not missing capabilities in other products. They do not establish that an equivalent design could not be built on another workflow system. Sources were consulted on 2026-09-19.

Evidence and trust

A matching receipt replay establishes internal consistency under the admitted runtime, graph, and tool signatures. Content addressing detects changes relative to an expected digest. Neither property is a cryptographic proof of external execution, a provider attestation, an approval signature, or evidence that the model's answer is true.

Typed graphs and bounded expressions reduce the executable surface, but admitted functions, executors, storage, and capability custody are part of the trusted host. Do not infer that an arbitrary manifest is safe to execute with arbitrary tools. Ordered journals support bounded exact-intent recovery, and qualified durable operation adapters support explicit coding-job reconciliation. Unknown writes still require settlement evidence; arbitrary external effects do not acquire exactly-once guarantees. Distributed custody transfer and OS isolation remain outside the current process VM.