A lossless byte↔pixel-grid codec: decode(encode(x)) === x, byte-exact,
for any input — not a visual summary, not a vibe. Every frame carries a CRC-16 checksum, so a single
altered pixel is caught and named, not silently returned as different-but-valid bytes. This
page runs the real kernel below — nothing here is a mockup.
The core promise is about arbitrary bytes, not text. Drop any file — an image, a PDF, anything — and this compares the decoded bytes to the original file byte-for-byte, not just "looks the same."
Encodes the text above, then flips one channel of one pixel and tries to decode the corrupted grid — proving the checksum actually refuses bad data instead of silently returning wrong bytes.
node --test kernel.test.mjs — 51 tests, clause-isolating, plus a 500-buffer fuzz pass and a
CRC-16/CCITT-FALSE standard reference-vector check (0x29B1 for "123456789").
node tools/witness.mjs mutate kernel.mjs --timeout 15000 --cap 400 --test node --test kernel.test.mjs
— 50/57 mutants killed directly, 7 argued equivalent-mutant exemptions (each with a written mathematical
reason in witness.baseline.json, none hand-waved), zero unexplained survivors.
CI re-proves both, pinned to witness v0.6, on every push.