Part of HeepX. New here? Start at
workspace — what the organization is,
why it is shaped this way, and what needs doing. This repository answers only
what must a repository contain.
The rules every HeepX repository follows, and the tool that checks them.
HeepX repositories are capabilities: text and evaluation suites that make an AI system measurably better at one thing. Hundreds of them are only useful if they are consistent — same layout, same metadata, same evidence standard, so a consumer can discover, load, and trust any of them without reading each one first.
This repository defines that consistency as HXS-1, and ships
conform.py to enforce the mechanizable part of it. A standard
nothing checks is a blog post.
git clone https://github.com/HeepX/standard
python3 standard/conform.py path/to/your/repo
Output is violations by rule ID, exit 1 on any MUST:
HXS-1 conformance: warrant (type: capability)
rules evaluated: 57
MUST DOC-06 README does not link https://github.com/HeepX/standard
Then read SPEC.md — every rule, defined once, with a stable
identifier. Everything else in this repository explains or operationalizes it.
To start a new capability repository, copy templates/ and work
through CHECKLIST.md.
SPEC.md HXS-1 — every normative rule ← the single source
conform.py mechanizable enforcement, stdlib only, self-tested
CHECKLIST.md pre-publication checklist for authors
REVIEW.md what a human reviewer checks that the tool cannot
NAMING.md HXS-NAME repository names as public API
REPOSITORY_STRUCTURE.md HXS-LAYOUT the tree, and why each entry exists
METADATA.md HXS-META CAPABILITY.yaml field reference
VERSIONING.md HXS-VER semver, deprecation, compatibility
EVALUATION.md HXS-EVAL what makes a suite trustworthy
BENCHMARKS.md HXS-BENCH comparative claims, honestly made
QUALITY.md HXS-DOC / HXS-QUAL README standard, quality bar
CI.md HXS-CI required checks
SECURITY.md HXS-SEC threat model, incl. this repo's own
schema/capability.schema.json machine-readable CAPABILITY.yaml schema
templates/ starting files for a new repository
Documents never restate a rule. They reference it by ID (HXS-EVAL-03) and
explain why it exists. If a document and SPEC.md disagree, SPEC.md wins.
python3 conform.py --self-check # the checker's own fixtures
python3 conform.py . --type infrastructure # this repository against itself
conform.py builds fixture repositories in a temp directory and asserts that
each rule fires when violated and stays quiet when satisfied — including that
an empty directory produces findings rather than a traceback, since the tool
runs against half-built repositories by definition.
Both run in CI, along with a check that every rule ID cited anywhere in this
repository exists in SPEC.md. Cross-references cannot silently rot.
- It does not make a repository good.
conform.pycovers the mechanizable subset — roughly half the rules. Naming something well, scoping one capability, or writing an evaluation that measures the real thing are judgment, which is why conformance requires human sign-off againstREVIEW.mdand not just a green check. - It does not parse YAML.
CAPABILITY.yamlis read line-wise to stay dependency-free: required keys, value formats, and referenced paths. Full validation is available viaschema/capability.schema.jsonto anyone with a YAML loader. - It is calibrated against two repositories. HXS-1 was derived from
building
HeepX/warrantand generalized from it;HeepX/reasoningwas then built against the standard and conformed on its firstconform.pyrun without a rule change. That is weak evidence the rules generalize — two repositories by one author, with similar shape — and it is not evidence they generalize to capabilities that look nothing like these. Amendments driven by repositories of a different shape are expected; seeCONTRIBUTING.md. - It has not been tested at scale. No claim is made about how HXS-1 holds up across hundreds of repositories, because that has not happened yet.
| Repository | Type | Capability |
|---|---|---|
warrant |
capability | Naming a claim's warrant before asserting it |
reasoning |
capability | Naming a commitment's defeater before building on it |
MIT. Copy it, fork it, derive from it.