Artifact: `and-super-rev-w143.zip`, with `and-super-rev-w143.receipt.json` beside it. Predecessor: `W.1.4.2`, `sha256:7dd1ab75…d1ff1`.
Release-evidence only, again. The runtime is not touched.
observe_once, multiplicity routing andRefusalLogare exactly as W.1.4.1 shipped them.
Both of your falsifiers reproduce, and I reproduced them before changing a line. You are right on every point, including the diagnosis of why the check was structurally incapable of firing.
site/proof/measurements.jsonFigures are in the receipt. The numbers that belong in prose are the subject of the round:
FALSIFIER A W.1.4.2 archive, hero-light-preview.png removed, valid ZIP
W.1.4.2 replay: exit 0
W.1.4.3 replay: REFUSED — names the missing path and both digests
FALSIFIER B same paths, same counts, no NUL, stabilityToken reverted to
`return viewClock;`
W.1.4.2 replay: exit 0
that archive's own authority battery, from its own bytes:
six assertions FAILED
W.1.4.3 replay: REFUSED — names site/app-prototype.html and both digestsB is the one that matters, and your framing of it is the correct one: a receipt could carry a browser-assertion figure with zero failures beside an artifact whose own battery fails six of them.
Your diagnosis is exact. The code was:
const walked = shipped(tmp); // tmp = the EXTRACTED ARCHIVE
const short = walked.filter(w => !inArchive.has(w));Both sides derive from the archive, so short was empty by construction. The branch that was supposed to name missing files could never name one. It read like a check and was dead code.
I should flag how this got written, because it is the more useful part. I considered comparing the archive against the build tree while writing W.1.4.2 and talked myself out of it — the reasoning was that post-package activity makes the live tree an unreliable comparand. That reasoning was sound about the live tree at replay time and I let it dispose of the whole idea, instead of reaching the obvious next step: capture the manifest before packaging and compare against that. A correct objection to one implementation retired the requirement.
Yours, adopted verbatim into the source:
Every release measurement must remain bound to the exact bytes whose behaviour produced it.
And the ladder, which is worth keeping written down because each rung looks like the previous one already handled it:
W.1.3.x a human copied the wrong measurement
W.1.4.1 a machine measured the wrong file SET
W.1.4.2 the set and the count matched and the BYTES differed`release-scope.mjs` gains contentManifest / serializeManifest / digestOf / contentDigest. The digest is over sorted (path, size, sha256(bytes)) — content, never ZIP metadata, so entry order, timestamps, compression method and permissions are excluded by construction. Sorting happens in exactly one function, so the tree side and the archive side cannot disagree by collating in two languages.
`package.mjs` writes the pre-package manifest — digest on line 1, one sha256 size path line per shipped file — to a path outside the tree. It has to be outside: a file describing the content digest would be part of the content it describes.
`replay-artifact.mjs` is rewritten. The expected side is that manifest; the actual side is hashed from the raw ZIP entries, decompressing each and hashing its bytes. Neither side is inferred from the other. It refuses on any missing path, extra path, or differing byte, and names which. It also still checks the receipt figure, NUL absence and measurements.json byte identity — byte equality implies those, but they tie the archive to the receipt rather than to the tree, which is a different edge.
It also no longer silently joins an absolute path onto the tree root, which turned "point this at an archive elsewhere" into a python stack trace.
Receipt is `release-receipt@3`:
"release_content_sha256": "…",
"artifact_replay": { "scope": "matched", "content": "byte-identical" }And the note wording is corrected, which was a fair hit. W.1.4.2 claimed every figure had been "re-derived FROM this archive". Four were; the behavioural figures are captured from the gated run. Claiming re-derivation the tooling does not perform is the same prose-over-evidence gap the proof battery exists to close, one level up — and it is exactly the kind of sentence this arc keeps finding. It now says measurements were produced by the gated run and that replay proved the shipped path set and bytes identical to that tree, so the behavioural figures are bound to the bytes whose execution produced them, not re-executed from the archive.
I agree with your architectural point and it is why I did not take the other route. Re-running the BEAM battery from the extracted archive would prove that archive passes. Proving
bytes the gates tested == bytes handed to the packager == bytes in the final ZIPkeeps every behavioural measurement attached to the exact code that produced it, for the cost of one hash instead of four and a half minutes.
tools/sabotage-scope.sh is nine cases now. The two new ones build well-formed archives — this is semantic tamper, not corruption:
a declared file removed from a valid archive your falsifier A
an archive edited to differ only in CONTENT your falsifier BEach names the sentence the gate must produce, so a gate that dies for an unrelated reason is not scored as a catch. I also ran your two originals — the real PNG removal and the full stabilityToken revert — against the new replay outside the harness, and both refuse while the untampered control passes.
ZIP timestamps. Identical content still yields a different archive hash, because entry mtimes vary. Normalising timestamps and permissions would make sha256 reproducible. The content digest already gives the "same code" property that actually matters, and the sibling receipt binds the real byte history, so this is ergonomics rather than evidence.
Corrupt or truncated archives fail noisily rather than being accepted. Inferior diagnostics, not a soundness hole — and, as you say, nowhere near the semantic-tamper acceptance.
`ampd/erl_crash.dump` is still in the working tree. Excluded from the archive and gitignored; yours to delete.
W.1.4.1 at-most-once semantics accepted, untouched
W.1.4.2 scope + measurement set concept accepted
W.1.4.3 exact-byte artifact binding this
W.2 Tauri LIVE LOCAL next, with nothing betweenbash tools/sabotage-scope.sh # nine cases, including both of yours
bash tools/release.sh # every gateTo reproduce falsifier B by hand against this artifact:
node tools/package.mjs probe.zip /tmp/man # manifest from the gated tree
python3 -c "import zipfile,re;
zin=zipfile.ZipFile('probe.zip')
zout=zipfile.ZipFile('tampered.zip','w',zipfile.ZIP_DEFLATED)
[zout.writestr(i, re.sub(rb'function stabilityToken\(b\)\{[^}]*\}',
b'function stabilityToken(b){ return viewClock; }', zin.read(i.filename))
if i.filename=='site/app-prototype.html' else zin.read(i.filename)) for i in zin.infolist()]
zout.close()"
node tools/replay-artifact.mjs tampered.zip /tmp/man # names the file and both digests