A second reader written from the contract alone: agreement with the reference on every published vector
An internal, contract-only Python reader agreed with the Go reference on 435 of 435 version-0 cases, 11 of 11 network cases and 75,229 re-signed mutants; the reference agreed with 629 of 629 cases the reader's author wrote.
Prediction heldDoes it work?Re-checked
- Designed for
- To test whether the English contract is enough for an implementer who has not seen the Go reference or its vectors to accept and refuse the same bytes, with the same codes and findings, as the reference checker.
What was tested
Inside the project, a separately instructed AI agent wrote a Python reader of the layer’s signed records from the English contract alone, barred from the Go reference code and, until its own tests passed, from the reference vectors. Each side checked the other’s corpus, plus re-signed mutants.
What we observed
The reader passed 110 of its own tests before seeing any reference vector. It matched 435 of 435 version-0 cases, 11 of 11 network cases and 75,229 mutants over 5 seeds. The reference agreed with 629 of 629 reader-written cases. For version 1, the reader matched 33 of 33 vectors and the reference 130 of 130.
What this means for Heimdall
In this setup, an internal, contract-only second implementation agrees with the reference: the checker role can be built from the contract text, within these corpora.
Limits
One project and one agent workflow wrote both sides; isolation is self-reported. 397 of the 629 reverse cases mirror version-0 scenarios read beforehand. Agreement on the current 41-case version-1 corpus rests on CI runs, not a review record. Reviews are internal and not approved. Only the checker role has vectors.
What remains open
Executor conformance vectors and a reader written outside the project.