The version tag could not be recorded where the ruling said, because the
stored value is validated to a bare sha256 shape by a pinned dependency in
another repository. Hands implemented an equivalent implicit version read,
flagged the divergence, and asked whether proceeding was right. It was: the
ruling named a mechanism, the hazard was live, and halting would have left the
fleet exposed.
Five tests prove it by failing without the fallback, which is the point, since
the whole finding was that nothing had ever exercised the stored path.
The frozen-source digest turned out to have the same exposure through a
narrower window, and this cut's own sixth and seventh batches crossed it
twice. Same answer applied, with the prior shapes reproduced from git history
rather than from memory.
Hands also spotted a third instance and declined to chase it: the systemd
actuator compares a raw content-only digest read through procfs against a
stored value now carrying mode, gid and xattr terms. Two different algorithms
compared to each other, which is the same defect a third time. Flagged for its
own look rather than guessed at.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The bypass is a compile error now, proven by builds that fail on Soul's exact
spellings rather than by reading the lint config. The repository had no CI
surface at all before this, so there was nothing for "denied in CI" to attach
to until now.
Hands flagged that the single-file artifact digest format changed and reasoned
nothing compares against literals. That holds for fixtures and not for
receipts: two call sites in host.rs compare a freshly computed digest against
a stored value. So the first upgrade past this commit makes every already
installed workload fail "installed executable differs from Expected", and
every running one fail "the installed executable changed under the running
process". Idunn is the deployment and daemon-survival authority, and this
would land silently at restart across the fleet.
The ruling versions the digest and verifies in the version the receipt names,
rather than re-sealing admitted state to make new code agree with itself. It
is not closed until a test seals under the old format and verifies under the
new, because the whole finding is that nothing exercised the stored path.
Two bugs in this batch were found by running the suite rather than trusting
it, including a test that set an attribute on a root whose own metadata is
never hashed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The port can be walked around. Two spellings of the original bug restore it
with all 193 tests green, and the proof it leaks is already in the file: one
function still calls fs::read_link directly, and a probe measured sixteen
counted calls beside three uncounted syscalls. My record claimed the count was
total rather than conventional. It is still conventional; the convention moved
from remembering the counter to remembering the port. That is the second time
this record has overclaimed what a pin does.
The artifact digest's single-file path never got the ruling, and it is the
path the installed executable actually takes. Setuid, mode and a file
capability are all invisible there, immediately before launch. Every test for
it wraps the file in a directory so the other path runs, and one of their doc
comments claims to pin exactly the scenario it does not reach.
The xattr framing is sound and nothing proves it. The suite stays green with
values dropped entirely, so a capability edited from cap_setuid to
cap_sys_admin is covered by no test.
Soul named one cause with three faces: every pin here is a convention a later
edit steps around, and the fixtures use the convenient shape rather than the
one production uses. So the eighth batch makes the bypass a compile error and
moves the tests onto the production shape, instead of adding a ninth round of
point fixes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
193 tests, and the mutation that passed in 0.39 seconds on pass 7 now fails.
Putting canonicalize and symlink_metadata behind a port that charges the
counter inside the same method as the syscall is what turned an honour system
into a measurement; the hand-placed call sites are gone.
The validator mask tests are precise rather than merely red: narrowing the
file mask fails exactly the setgid test and narrowing the directory mask fails
exactly the setuid one, with the other five staying green. That is the
evidence they pin single bits.
The unreachable test is deleted rather than repaired, with a note saying why
the defence it claimed never existed. The xattr tests close Soul's file
capability finding and state in their own doc comments that the probe runs as
root and proves nothing about an unprivileged attacker.
Hands shipped a compile bug in the first commit and its own Yggdrasil run
caught it before the report. That is the order these should happen in.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Soul restored the exact bug this line of rulings exists to prevent, without
instrumenting it, and the new call-count pin passed in 0.39 s. The counter
only counts calls someone remembered to annotate, so it is an honour system
rather than a measurement. The code said so in a doc comment four lines below
the assertion; my record claimed the opposite. That record is corrected here.
The fixture I recorded as deleted for being unable to fail is in the tree,
still unable to fail. Hands' reasoning about why the collision class is
unreachable was sound and Soul retested it; what landed on top of that
reasoning was a test that cannot go red.
A file capability is invisible to both the validator and the digest, which is
the grant setuid used to be with no mode bit to see it. no-new-privileges
prevents it on our own deployment path, but the digest is the detection and
Cut 4's verdict claims the tree is the one we froze. The artifact digest also
got a single executable boolean where the ruling said full mode and gid, so a
post-install setuid re-verifies clean.
Six rulings for the seventh batch. The first moves filesystem access behind a
narrow injected port so the count is total rather than conventional, which is
the shape doctrine already asks for.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
183 tests green. The timing pin is now a call counter asserting an exact
formula, which is what it should have been before a clamp made a 1.63x
separation look like 33x.
I ruled the bind mount should carry nosuid. Docker exposes no such
per-bind-mount flag through either syntax, and Hands verified that against
the real CLI rather than arguing from memory. The container already runs with
no-new-privileges, which is process-wide and stronger than what I asked for.
The ruling is withdrawn: the prevention was already there and I ordered a
second one without looking.
Two fixtures in this batch would have lied and were caught by mutating rather
than reading. The container's chown clears a freshly-set setuid bit, so a
fixture that sets then chowns tests nothing. And the first observe_frozen test
tampered with the recipe file, which a different check inside observe_frozen
catches, so it stayed green with the digest check deleted.
The artifact digest's symlink terminator is deliberately left unpinned:
symlink(2) cannot put a NUL in a target, so no honest test exists. Hands built
one, confirmed it did not go red, and deleted it instead of keeping it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Five of seven rulings hold. The containment argument, where I asked Soul to be
most suspicious after S5-8, survived twenty escape shapes it built itself.
R-I1's pin is not a ratio test. Its clamp fires on every run, so it reads as
"under 200 ms", and clean code has already measured 198.9 ms under the
three-slot load the operator just authorised. The reported 33x separation was
an artifact of that clamp; the real separation is 1.63x, which a slightly
faster host erases. It fails in both directions, and under Cut 4 a false red
is a false verify verdict on unattended work.
R-I3 is not met. The validator masks the mode to 0o777 and checks uid but not
gid, so setuid, setgid, sticky and gid are all permitted to vary and none are
hashed. A frozen binary made setuid-root re-observes clean against its
receipt, and the tree is bind-mounted into runner containers with no nosuid.
The ruling adds the mount flag as well as the digest terms: the digest detects
and the flag prevents, and the flag costs a mount option.
Soul also found that the artifact digest still has the blindness the frozen
source digest just lost, and that observe_frozen, which is the tamper check
itself, has no test at all.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
All seven rulings landed, 175 tests green, and every new test dies under its
own mutation. The two-link cycle fixture hangs rather than asserting, which is
the correct falsification: nothing but the budget bounds a cycle, so a timeout
is what a broken budget produces.
The batch's real finding runs the other way. Soul called three containment
checks redundant and I ruled delete them. Hands deleted two, hand-traced the
third, and found that with all three gone a symlink whose whole target is ".."
at the root resolves to the parent of the root and returns Ok, because no
later Normal step ever runs to trip the other arm. That is a containment
escape, the exact class this resolver exists to prevent. Soul was wrong, my
ruling repeated it, and a hand-trace caught it.
Soul's HTTPS rig had also evaporated: no session, no work directory, no
scratch trace. Hands rebuilt it from this map's description, which worked only
because the description was detailed. A probe that is going to be committed
has to be handed over while it still exists.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Soul could not falsify any claim the fourth batch made. What it found instead
is that the S4-1 fix reached the exponent and not the per-component cost, and
that the line responsible is dead code: the value is already canonical, so
deleting four lines removes a cubic term and leaves the suite green.
The other half of the bill is coverage. The traversal budget is the only thing
that terminates a symlink cycle and nothing tests it; cargo-mutants missed two
mutants on that single line. The digest that serves as the tamper check is
blind to symlink targets and to the executable bit, both of which the
validator permits to vary. S4-6 was reported landed and has no test at all.
Seven rulings for the fifth batch. R-I5 departs narrowly from my earlier
"keep S6's behaviour exactly" and says so rather than letting it pass as
drift: the 40-traversal ceiling matches the kernel's own ELOOP limit and is
kept, but it must be declared and the error must name the link.
FU-Heimdall-Gitlink cleared upstream on its own. Heimdall froze end to end,
945 entries, zero mismatches. The ops follow-up closes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Operator ruling 2026-09-22 after Starfire froze under agent build load. A
verify transaction takes an exact revision and declared steps, runs them in
capped runners in its own worker, returns a typed verdict, seals nothing and
sits outside the deployment brake. Imagination's map; ten operator questions
open, five of them blocking Cut 2.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
