Appendix C — Executable Claim Contract
The book separates learner-visible code into an equation, a mechanism kernel, and build orchestration. Hashing paths, activating a vendored wheel, loading a JSON record, and checking its content digest are essential to publication but are not the semantic change a new chapter teaches.
The printed contract is therefore:
| Surface | Kept on the page |
|---|---|
| equation | the mathematical object, with shapes and reductions |
| mechanism kernel | normally 10–25 lines exposing the chapter’s new operation |
| critical assertions | two or three checks that protect the interpretation |
| provenance strip | claim ID, provenance class, seed, dtype, device, estimator, and artifact |
| full harness | linked source and collapsible HTML where useful; not repeatedly printed |
Chapter source still executes the full verification. It calls the centralized interfaces below from a hidden setup cell:
pin = activate_harness("ch-13", expected_wheel_sha256)
claim = verify_claim("c13-regime-001", expected_harness=pin)activate_harness verifies the manifest identity and wheel SHA-256 before placing the exact vendored wheel on the import path. verify_claim checks the claim registry, artifact identity, canonical payload digest, and wheel agreement. Removing those lines from repeated printed panels changes no assertion and no executed output.
Historical rolling-draft wheels are identified by their committed SHA-256; the book does not invent source tags after the fact. Beginning with harness-ch-14-r2-v1, each new wheel also records an annotated source tag and commit. The two cases are explicit in the manifest and checked by CI.
The complete implementation is available in book_support/claims.py. Claim records are browsable under artifacts/claims/, and historical harness pins under artifacts/harness/.
C.1 Reading a provenance strip
The strip is a locator, not evidence by itself:
- claim names the proposition being checked;
- class distinguishes laptop reproduction, CPU simulation, and a pinned external endpoint;
- seed names a stream but never substitutes for a seed panel;
- dtype/device scope the execution;
- estimator names the target, reduction, and denominator;
- artifact links the complete record and its environment sidecar.
The chapter remains responsible for the mathematical and empirical interpretation. A valid digest can prove that the intended artifact was read; it cannot make the artifact’s assumptions true.