Artifact: `and-super-rev-w142.zip`, with `and-super-rev-w142.receipt.json` beside it. Predecessor: `W.1.4.1`, `sha256:b05b8922…8166`.
Release-evidence only. The runtime is not touched.
observe_once, multiplicity routing andRefusalLogare exactly as W.1.4.1 shipped them. Your §4 chase is right and I have not reopened anything.
You found that W.1.4.1's receipt recorded 113 text assets while the archive it binds contains 112, and you named the mechanism. Both are correct. I reproduced it, and then went looking for the rest of the class, which turned out to be larger than one file and pointed in both directions.
site/proof/measurements.jsonFigures are in the receipt; this document references it rather than restating it. Two numbers belong in prose because they are the subject of the round rather than a measure of its health:
check-source-hygiene, run against the W.1.4.1 artifact
clean extracted W.1.4.1 112
+ one previous and-super-rev-*.receipt.json 113
− it again 112That is your hypothesis, reproduced deterministically on this box. The receipt said 113 about an archive containing 112.
release.sh ran the hygiene gate at stage 3 and rm -f ./and-super-rev-* at stage 27, with packaging at stage 28. W.1.4 shipped immediately before W.1.4.1, so and-super-rev-w14.receipt.json was in the tree when the gate counted, and gone when the packager ran.
What makes it worth a round rather than a patch: nothing was transcribed. Both numbers came out of the same tool, on the same tree, within the same minute. The arc that produced emit-measurements.mjs was aimed at humans retyping figures, and this defect has no human in it. The figure was machine-generated, hash-bound, and describing different bytes than the ones it was bound to.
Your law is the right generalisation and I have written it into the source:
A release measurement must quantify the same artifact set the receipt binds. Measurement scope is part of artifact identity.
Your finding says one non-shipping file was counted. The inverse turned out to be true of more files, and one of them is not small.
ampd/erl_crash.dump shipped in every archive, and no gate ever read it5,066,337 bytes — 940,874 compressed, 40% of the W.1.4.1 archive. A BEAM crash dump from a runtime that died during boot on 2026-08-22 (Slogan: Runtime terminating during boot), 153,829 lines of scheduler and atom-table state.
It is named in the repo's own .gitignore, under a comment saying crash dumps are not history worth carrying. It was invisible to the hygiene gate because .dump matched no textual extension, and it was invisible to both packagers because neither skip-list mentioned it. Nothing had an opinion about it in either direction, which is how it rode along.
I checked it for disclosure before deciding severity, because "40% of a review artifact is an unread crash dump" deserves better than a guess. The scan hits are atom-table entries, not values: password, get_password, srp_user_secret_nif and srp_host_secret_nif are OTP crypto and :io function names, and every token hit is Elixir.String.Tokenizer. No home paths, no environment block. So this is weight and unread bytes, not a leak — but an archive that ships five megabytes no gate has read is making a claim about bytes it never opened.
| shipped in W.1.4.1 | why the gate could not see it |
|---|---|
ampd/c_src/ampd_fd_nif.c | .c was not in the textual extension list |
host/Cargo.lock | .lock was not either |
site/preview/preview-meta.json | the walker skipped every directory named preview |
The C file is the sharp one. release.sh ships it specifically to be compiled at the far end — its own comment says "Ship the twelve lines of C; let the compiler task build it where it will run" — and the gate whose stated purpose is no NUL in any text asset could not see it. This is the W.1.3.2 defect with the file changed: a parity gap that is currently harmless.
release.sh chose between zip and python on whether the zip binary happened to be installed. The zip path excludes neither .git nor .elixir_ls; the python path excludes both; neither excludes deps, which the hygiene walker did. The bytes a revision names depended on the toolchain of the box that packaged it — in the script that carries the paragraph titled A revision must name one byte history.
It never fired: super/ has no .git of its own, .elixir_ls and ampd/deps are absent, and this box has no zip. One mix deps.get before a release is all it would have taken.
`tools/release-scope.mjs` — new. One declaration of what a release contains. The hygiene gate walks it, the packager writes it, the replay re-derives it. Before this there were four disagreeing definitions — the hygiene walker's skip-list, the zip globs, the python tuple, and .gitignore — and no two described the same set.
Previous artifacts (*.zip, and-super-rev-*.receipt.json) are excluded from every walk. That is the class fix rather than the ordering fix, and it is where I deviated from your proposal — see §6.
`tools/package.mjs` — new. One packager, over shipped(), in sorted order. It refuses to package a file the declaration does not classify as either textual or opaque, which is what would have stopped 2.1: silence is no longer an available answer.
`tools/replay-artifact.mjs` — new. Your item 4, and the strongest single thing in the round. It extracts the archive that was just built and re-derives the figures from those bytes:
the archive is exactly shipped() — no more, no fewer
counted() over the archive equals the receipt's figure
no NUL in any of them, scanned again from the archive
the archive's own measurements.json is byte-identical to the one being bound
Each of those four failed somewhere in W.1.4.1.
`text_assets` → `source_text_assets`. Renamed because the old name did not say which set. The old noun is deliberately left unguarded: it now refers only to a superseded universe, which is why this document may quote what W.1.4.1 measured.
Receipt is `release-receipt@2` — adds files and artifact_replay: "reproduced". Reaching that line means the archive was extracted and its figures re-derived from it.
tools/sabotage-scope.sh — the new law, falsifiedYour item 3, widened. Seven cases, each naming the sentence the gate has to produce, run against a scratch tree the harness owns, exercising the shipped code paths:
a stray sibling receipt does not move the figure the W.1.4.1 defect itself
a NUL in the shipped C source invisible to the old gate
a NUL in the shipped preview JSON invisible to the old gate
a file the scope declaration does not classify refused at packaging
a file added between measurement and packaging replay refuses
a measurements.json edited after packaging replay refuses
a clean tree packages and replays a gate that refuses everything is not oneOne case was wrong on the first run and the tool was right. I originally added the late file after packaging and expected a refusal. It did not come, correctly: an archive built at T2 is not wrong about a file that appeared at T3, and the release is the archive. The real defect shape is a file arriving between the gate and the packager — in the archive, not in the figures — which is W.1.4.1 pointed the other way. That case now passes for the right reason.
And the hygiene gate caught me while I was widening it. Writing the file that forbids literal NULs, I put one in the comment that forbids them. The gate named the file, the line, the byte count and the fix on its first run. Unplanned, and better evidence than a probe.
Taking your first option. ampd/tools/sabotage.sh refuses in release.sh exactly as mix and cargo do; tools/preview-release.sh remains the degraded path that cannot package. Your argument is the one that decided it: this round is the proof that release.sh green ≠ falsifiers green, and the only evidence the falsifiers had run was a sentence in a brief — for the single property that separated the bad W.1.4 from the good W.1.4.1.
Putting it in the chain naively would have created a fresh drift defect. ampd/tools/sabotage.sh and tools/sabotage-host.sh printed byte-identical summary lines, and emit-measurements.mjs reads figures out of the release log by regex. The first battery to print would have been recorded under the other's name. bot, guard and count sabotage were already prefixed; these two were the only ones nobody had needed to tell apart. They are beam sabotage: and host sabotage: now, and beam_falsifiers is bound to the artifact like every other figure — which is your option 2, obtained for free by taking option 1.
You asked to move the `rm -f` above the measurement gates. I excluded the files from the scope instead, and left the `rm` late.
Two reasons. The ordering fix repairs the instance and keeps the class: a sibling receipt can arrive by other routes — a hand-run probe, an interrupted round, a copy made to compare two revisions — and would move the figure again. And a release that fails at stage 12 would have destroyed the last good artifact on its way to failing.
With the exclusion in place the ordering is irrelevant to every measurement, and case 1 of sabotage-scope.sh proves it: a stray receipt planted in the tree does not move the count. The rm stays where a failed release leaves the previous artifact intact.
On your preference for replay over more exclusions — agreed, and the replay is why the exclusions are safe. Two walkers reading one declaration probably enumerate the same files. That is exactly what was true of the old pair, right up until it wasn't.
`ampd/erl_crash.dump` is still in the working tree. It is excluded from the archive now and gitignored, so it cannot ship, but I have not deleted it — it is yours to remove, and nothing depends on it.
The archive is not byte-reproducible. ZIP entries carry mtimes, so packaging the same tree twice produces different bytes. The receipt binds what was built, which is sound, but "same tree → same hash" is not a property this pipeline has. Out of scope for this round; worth naming before someone assumes it.
**Neither the replay nor the packager is falsified against a corrupt archive** — a truncated or tampered ZIP. Every case in sabotage-scope.sh operates on well-formed archives.
Unchanged from yours, and this round is closed as you scoped it:
W.1.4.1 at-most-once semantics accepted, untouched
W.1.4.2 artifact/measurement closure this
W.2 Tauri LIVE LOCAL nextNo further runtime design round. §4 is closed by the shipped runtime and I did not reopen it.
The separate research brief in the bundle (MOTOR_MACHINE_RESEARCH_BRIEF_FOR_OPUS.md) is a design lane for after the desktop foundation, not a proposal to interrupt W.2. It ships in the archive because it is in the tree, not because it is part of this round.
bash tools/sabotage-scope.sh # the new law, falsified
bash tools/release.sh # every gate, now including the BEAM falsifiersTo reproduce the original defect against the predecessor artifact:
mkdir w141 && cd w141 && unzip -q ../and-super-rev-w141.zip
node tools/check-source-hygiene.mjs # the old gate, on the old bytes
cp ../and-super-rev-w141.receipt.json ./and-super-rev-w14.receipt.json
node tools/check-source-hygiene.mjs # one higher, and the archive has not changedTo watch the replay refuse, package and then add a text file to the tree before running it:
node tools/package.mjs and-super-rev-probe.zip
printf 'x\n' > docs/late.md
node tools/replay-artifact.mjs and-super-rev-probe.zipIt names the figure, both values, and which one the hash is bound to.