Certified coverage for dead formats
The Codec Compiler synthesizes a byte-exact decoder per format; the Registry is the catalog of those decoders. Each entry is version-stamped and carries a certificate that says this codec was proven byte-exact on N held-out samples, and here is the SHA-256 you can recompute to check it. A codec earns “certified” only by round-tripping byte-exact on data it was never tuned on — training overlap and lossy codecs are rejected.
- byte-exact
- 'certified' = decode(encode(x)) == x on a held-out corpus, every sample
- reproducible
- each badge carries a cert_sha256 an auditor recomputes from codec + corpus
- held-out
- the badge is earned on samples the codec was never tuned on — overlap is rejected
- 3 founding
- WinBranch DBF · MT940 SWIFT · EBCDIC copybook — coverage grows over time
Certify it, then verify the badge
at1 codereg register --all # certified WinBranch-HST v1.0.0 — 8 held-out samples, # all byte-exact: True, ratio 3.27x # cert_sha256 b1ac47d0... holdout_root b3b9cbb8... # certified MT940 v1.0.0 — 7 held-out samples, byte-exact # certified EBCDIC-COPYBOOK v1.0.0 — 8 held-out, byte-exact at1 codereg verify EBCDIC-COPYBOOK # OK: recomputed cert_sha256 == stored (byte-exact: True)
Earned, not asserted
The registry doesn’t trust the badge — it runs the round-trip on a held-out corpus and mints the certificate. All samples byte-exact, or the status is REJECTED. A lossy codec never certifies.
Reproducible certificate
The cert_sha256is a pure function of the codec identity + the pinned held-out corpus. Anyone holding both recomputes the same hash — years later, with no surviving AT-1 software (pairs with Century self-decode).
Honest scope
“Certified” means byte-exact reconstruction on the pinned held-out set — not a claim about field meaning. Coverage starts at three founding formats and grows as each new dialect is onboarded with its own corpus.
A certificate proves reconstruction fidelity on the pinned samples, not interpretation of the fields. Coverage is a growing catalog; a decode failure in the field opens a new sample and a new codec version, never a silent pass.