---
id: ADR-0001
title: "Stage 0 library stack: schemas, pure fold, and validators"
status: accepted
date: 2026-07-28
decided_by: Bernd
accepted_at: "2026-07-29"
workstream: TREV-WP-0002 (trust-service-foundation)
alternatives_considered: [TypeScript/Node, Go]
---

# ADR-0001 — Stage 0 Library Stack (Schemas, Pure Fold, Validators)

## Status

**Accepted 2026-07-29** by the maintainer (Bernd), confirming the stack
already implemented under `workplans/TREV-WP-0002-trust-service-foundation.md`
T02–T06: Python 3.11+, `jsonschema`, `pytest`, `hatchling` src-layout,
SHA-256 canonical-serialization hash chain, Ed25519 signing. This decision
locks implementation technology for Stage 0 schema/fold work only (§ Scope
below) — it does not extend to a future hosted Trust Service ADR.

## Scope

This decision covers **only** the Stage 0 deliverables of WP-0002:

- JSON Schema (or equivalent) definitions for Phase Manifest, Target Ledger
  Entry, Extension Contract, Transaction Allocation, Conversion Attestation;
- a pure, deterministic Outstanding Target fold;
- offline conformance validators;
- a golden Phase package and its test suite.

It explicitly does **not** cover: hosted Phase/Extension Registry services,
multi-tenant ledger APIs, a metrics product, or a federation protocol
(`SCOPE.md` §3–§4). Those require a separate, later ADR once a Trust Service
implementation workplan exists.

## Context

`specs/TechnicalSpecificationDocument.md` §10 is deliberately non-binding on
language, runtime, and storage. Something has to be chosen to write the
Stage 0 library, but the choice should be cheap to revisit later since the
hosted Trust Service is a different, larger decision.

Requirements pulled from TSD §3, §6 and `specs/OpenQuestions-WorkingDefaults.md`
Q14:

- deterministic, pure computation (no hidden state, no floating-point order
  sensitivity) for the Outstanding Target fold;
- JSON-Schema-shaped validation that a non-Python tool could later
  reimplement against the same `.schema.json` files;
- SHA-256 canonical-serialization hash chaining and Ed25519 signatures over
  the same canonical bytes;
- a test suite runnable offline, with no network services.

## Decision

**Python 3.11+, `jsonschema` for schema validation, dataclasses for typed
domain objects, `pytest` for the conformance suite, `hatchling` src-layout
packaging.**

Canonical serialization for hashing: JSON with sorted keys, no insignificant
whitespace, UTF-8 encoding (`json.dumps(..., sort_keys=True,
separators=(",", ":"))`), hashed with SHA-256. Signatures: Ed25519 via
`cryptography` or `PyNaCl`, over the same canonical bytes.

## Rationale

| What Stage 0 needs | Python fit |
|---|---|
| JSON Schema authoring and validation | `jsonschema` is a direct, widely-used implementation |
| Pure fold over ledger entries | Trivial with dataclasses + `functools.reduce`; no framework needed |
| Canonical serialization + hashing | Stdlib `json` + `hashlib` sufficient |
| Ed25519 signing for examples | `cryptography` covers this without custom crypto code |
| Offline conformance test suite | `pytest` is the de facto standard; matches this workspace's other Python repos (e.g. `shard-wiki`) |
| Low ceremony for a schema/fold library, not a service | Python's scripting ergonomics outweigh static-typing benefits at this scope |

### Why not TypeScript/Node

Would be a reasonable alternative if the primary consumer were browser-based
tooling (cf. `binect-js`), but Stage 0 has no UI deliverable — it is a
schema + pure-function library consumed by CLI/tests. Node adds packaging
overhead (npm registry, `package.json` versioning discipline) without a
corresponding benefit here.

### Why not Go

Go would fit well if this were becoming a long-lived, standalone,
performance-sensitive service (cf. `key-cape` ADR-0001's reasoning) — but
Stage 0 is explicitly **not** the hosted Trust Service. A statically-typed,
compiled toolchain is more ceremony than a schema/fold/test library needs at
this stage. Revisit for the later hosted-service ADR.

## Consequences

### Positive

- Fast path to a working, testable schema + fold library.
- `jsonschema` files are directly reusable by a future non-Python
  implementation (hosted service ADR is unconstrained by this choice).
- Matches existing Python conventions in this workspace (pytest, hatchling
  src-layout) for reviewer familiarity.

### Negative / risks

- Python's dynamic typing means domain-model discipline must be enforced by
  convention (dataclasses + docstrings + tests), not the compiler.
- Should not be read as a decision about the hosted Trust Service's stack —
  that is explicitly deferred and must not be assumed to inherit this choice
  without its own ADR.

### Compensating guardrails

1. Typed domain objects (`dataclasses`, `frozen=True` where the schema marks
   a field immutable) — no raw dicts crossing function boundaries in
   `fold.py` or `conversion.py`.
2. All schema validation goes through the `.schema.json` files under
   `schemas/` — no ad hoc field checks duplicated in Python without a
   corresponding schema entry.
3. The fold function must be pure: same `(manifest, entries)` input always
   produces the same Outstanding Target output, with no I/O.

## Revisit trigger

Reconsider this decision when a hosted Trust Service implementation
workplan is opened (`SCOPE.md` §3 "Production Trust Service"). That
decision should weigh multi-tenant hosting, storage, and operational
concerns not in scope here, and may reasonably choose a different stack
without invalidating this one.