The scrutiny an Athenian official passed before taking office.
Not a report on what was done — the examination a thing submits to
before it is trusted.
Infrastructure for testing software that has to be right.
A suite that cannot fail proves nothing, and most of them cannot. The usual answer is to write more tests, which mostly adds more of the same confidence. Ours is to break every gate on purpose before it ships and require it to go red — on code whose own suite was already green, that habit has found a seat losing three failures in four under concurrency, a conformance reader ignoring whole case files, and a suite checking its answers with the library it was testing.
flowchart LR
S["your source"] --> E
E["eidos<br/>one symbol graph,<br/>seven languages out"] --> D
D["dokimi<br/>generates the suites"] --> B["benchmarks"]
D --> F["fuzz targets"]
D --> C["conformance"]
D --> M["model &<br/>linearizability"]
A["assert-*<br/>41 assertions,<br/>one meaning"] -.-> B
A -.-> F
A -.-> C
A -.-> M
B --> R["reporter"]
F --> R
C --> R
M --> R
R --> Z[("dashboard")]
classDef ours fill:#0d1117,stroke:#58a6ff,stroke-width:2px,color:#e6edf3
classDef asserts fill:#0d1117,stroke:#3fb950,stroke-width:2px,color:#e6edf3
classDef out fill:#0d1117,stroke:#d29922,stroke-width:2px,color:#e6edf3
class E,D ours
class A asserts
class R out
Writing that infrastructure by hand, once per language, is where the integrity goes. Six teams each implement what they believe the specification says, every suite passes, and nothing tests whether the six beliefs agree. So the definition is data rather than prose, because prose does not fail a build, and the harnesses around it are generated rather than typed out six times.
eidos parses source into a symbol graph belonging to no language and renders it into seven. dokimi turns a definition into the suites that exercise it. The assertion libraries are what those suites assert with, and the reporter carries what they found into one shape whatever produced it.
One set, defined as data, so the same test means the same thing in every
language that implements it. 41 assertions · 87 corpus cases · run
against both surfaces in every language.
| Repository | ||
|---|---|---|
| 📐 | assert-spec | The definition — meaning, names, and the cases every library runs |
| 🐹 | assert-go | Go |
| 🐍 | assert-python | Python |
| 🟦 | assert-typescript | TypeScript |
| ☕ | assert-java | Java · Kotlin |
| 🦀 | assert-rust | Rust |
Go's
cmpsays a nil slice equals an empty one. Python saysNonedoes not equal[]. Port a suite between them and both builds go green while testing different things.
Agreement is not a property any one library can have. A library can be correct on its own terms and still disagree with its five siblings, so the definition lives outside all of them and every one runs the same cases in its own CI.
Determinism first Same input, same bytes.
Mutate every check Break it, watch it go red, restore. Say so in the commit.
Never judge with the A runner that checks its answers using the library
subject under test passes having checked nothing.
Declare divergence Where a language cannot comply it says so, with a
reason someone can argue with.
A failure is a record Named values, not prose to search for a word.
| Greek | Meaning | |
|---|---|---|
| dokimasia | δοκιμασία | the scrutiny before office |
| dokimi | δοκιμή | the trial itself |
| eidos | εἶδος | the form, before any language gives it syntax |
Assertions 1.0.0, five languages, unpublished · eidos and dokimi unreleased
Built by ThesmOS · MIT