Repository navigation
Releases
YuCe 0.2.2
YuCe 0.2.2: one agent on its own folder from one line of configuration, a preview of what a withdrawal would make fall, and the journal as W3C PROV-O.
The journal format and the rules of the kernel do not change: journals of 0.2.0 and 0.2.1 replay and verify as before.
New
-
yuce mcp: one agent on its own folder from one line of an MCP client's configuration.yuce mcp --docs <folder> --ask <question>is one MCP server over stdio that the client starts itself.- It starts or resumes the run
audit-serve --profile solowould, with the journal in<folder>/.yuce/run.wal. - It keeps the run's two tokens to itself and answers as both subjects: the reading agent's tools, and the owner's reading tools.
- A long answer comes in parts for clients that cap the size of one tool result;
audit_moregives the rest.
-
A preview of a withdrawal.
audit_withdraw_previewtakes the arguments ofaudit_revokeoraudit_withdraw_textand answers with the report of what would fall:- the records that would fall;
- the quotes that would go;
- the questions to answer again.
Nothing is written. The withdrawal is staged the way a real one is, so a refusal is the withdrawal's own. The Python client has
preview_withdrawal. -
Provenance export: the journal as W3C PROV-O.
sbgk audit-prov --wal run.walwrites one JSON-LD file that any RDF tool reads, rebuilt from the journal at any time:- who wrote each record and what it rests on;
- on which bytes of which document version an observation stands;
- each period in which a record stood, and the frame that ended it.
The same journal always gives the same bytes, and the file names the SHA-256 of its content. A sample is in
examples/prov-export.
Fixed
- The server's reading tools write nothing. Before, a reading tool could add the server's pending entry to the journal, such as an agent's failure found during a long poll.
- Now the reading tools, and
audit_statuswithtelemetry, write nothing. What is pending is written by the next call that coordinates or writes. audit_statuswithouttelemetrymoves the run on as before, and its description says so.
- Now the reading tools, and
Other
- NOTICE and the README name the product names YuCe and Organetic Sphere. The license is Apache-2.0 as before.
The full list is in CHANGELOG.md.
YuCe 0.2.1
YuCe 0.2.1: a table of what each part of the documents came to, strict signed notes, and answers that list each record once.
The reviews of 0.2.0 and a second corpus for the gap pass found what this release fixes. The journal format and the rules of the
kernel do not change: journals of 0.2.0 replay and verify as before.
Fixed
- An answer lists each of its records once. After a repair in several attempts, an answer of 0.2.0 could name a record twice
(control run 1 of 0.2: 11 of 54), and the report of what fell counted its standing grounds too high. Journals of 0.2.0 replay as
written; LIMITATIONS.md says how to count them. audit-proofandaudit-signnever write over what they read. In 0.2.0audit-proof --outpointed at the journal wrote the
quittance over the run. Both now refuse such a path in any spelling and write a new file before renaming it into place.- Signatures are checked strictly, and a weak verifier key is refused. Under the check used before, a "signature" made for a key
of small order verified any text. Keys made bysbgk keygenare never weak; no signature written so far changes. - Signed checkpoints hold to the reference implementation of C2SP signed-note, with vectors written by the code of its authors.
Five places where the kernel read a note more leniently than the reference now read it as the reference does. - The journal of a running server is read by other processes, on Windows too. Beside a live server,
audit-verify,
audit-read,audit-indexandaudit-prooffailed on Windows (os error 33); they now read up to the last complete entry and
say so. One writer still, on every system. - A repair after a withdrawal is told what the withdrawal took: only the quoted bytes it names, not their document or revision;
a point the records no longer establish is named in the done text as a gap. The kernel checks form, not meaning, and no run with a
model has measured this. - A question about what changed takes the claims that say something was removed or renamed. Measured by replaying journals
without a model, not by a run with one. - The coverage of the gap pass is complete only when no second reading is owed, and the library writes no seal right after
another.
New
- A coverage table drawn from the journal:
sbgk audit-read coverage --wal <run.wal>, andaudit_coverageon the live server,
give a row for every list item, paragraph and table row of every document with what it came to: quoted and by which records,
answered empty, left by the reading budget, withdrawn, or unread. The index isyuce-index/2; build an index of
yuce-index/1again withsbgk audit-index --rebuild. audit_seal: the owner seals the journal on request, for example right before a withdrawal;seal()in the Python client.verifier/: the code ofyuce-verify, an independent verifier of 0.2 journals: a standalone implementation in TypeScript from
the open specification of the format, sharing no code and no dependencies with the kernel, 391 tests. It passed both control runs
of 0.2.scripts/verify_published_runs.pychecks the runs of 0.2 too.
A second corpus for the gap pass
benchmarks/PREREGISTRATION_0.2.1_A2A.md
fixed it before any run: the A2A specification of v0.2.5, v0.3.0 and v1.0.1 with its changelog, an oracle of 14 facts that ended by
v1.0.1, two runs with the gap pass and two without, graded blind by Gemini 4 Argon.
- By the pre-registered rule this counts against the gap pass: both runs with it stated more ended facts as current (8 and 6 at
stage 1) than both runs without it (2 and 1). - The journals show a defect in the kernel, in the claims it chose for a question about what changed. 0.2.1 fixes it (above).
The comparison does not separate the defect from the pass, and no run with a model has measured the fix. - The gap pass stays an option, off by default. "Nothing skipped in silence" is now carried by the coverage table, which shows
every part left unread, row by row.
The four runs are summarized in BENCHMARKS.md; their journals are not attached to this release.
Results and caveats: BENCHMARKS.md.
Full list: CHANGELOG.md. What it does not do: LIMITATIONS.md.
Assets
- Binaries:
yuce-v0.2.1-<target>for Linux x86_64, Windows x86_64 and macOS (arm64, x86_64), each withsbgk, its alias
yuce, LICENSE, NOTICE, THIRD_PARTY_LICENSES.md and README;SHA256SUMS-<platform>.txt. - The Python client: organetic-yuce on PyPI.
- The recorded runs stay in v0.2.0,
v0.1.7, v0.1.6
and v0.1.0.
Check a download with sha256sum -c SHA256SUMS-<platform>.txt.
YuCe 0.2.0
YuCe 0.2.0: what fell after a withdrawal, a quittance for one conclusion, and a journal anyone can check.
A source your agents quoted is withdrawn: the kernel lists the conclusions that stood on it, from the record and not from a model,
and the others stay standing. 0.2 makes that record checkable by anyone, piece by piece.
New
- What fell, and why.
sbgk audit-read fallen --wal run.wal(oraudit_fallenon the live server) names each record that lost a
ground, the quotes that went (by their bytes, not their text) and why, the quotes under it that still stand, claims on other
sources that say much the same, and what to reread.status,dependents,active-claims,diffandrecordread the same
journal. - Quittances (
yuce-proof/1).sbgk audit-proofcuts one file with a record and everything it rests on, down to its quotes,
under the signed checkpoint of a seal: 7 to 24 KB in the control runs.sbgk audit-proof-verifychecks it alone, with no journal
and no model. - Sealed frames and signed checkpoints. The journal (
sbgk-session-log/3) is written in frames with a header and a body; the
server seals it every 256 frames, at each phase of a circle and at the close of a run. Seals are signed as C2SP checkpoints
(Ed25519):sbgk keygen,audit-serve --signing-key,audit-sign, andaudit-verify --checkpointwith a file of trusted keys. - The index beside a journal.
sbgk audit-indexbuilds a SQLite database any tool reads, rebuilt from the journal at any time;
--checkcompares it with the journal row by row. - The gap pass (
--gap-pass, off by default). The parts of read sections that no quote covers go back to the agents as narrow
tasks: record what a part states with an exact quote, or answer empty. - A repair after a withdrawal shows the live support of what fell, so a conclusion that another page still carries is not lost.
- A settings file (
audit-serve --settings) and one agent on its own folder in one command (--profile solo --docs <folder>).
A new run writes the 0.2 journal; audit-verify checks journals of both formats, and the published journals of 0.1 verify as before.
The server no longer resumes a 0.1 journal: finish it with a 0.1 release.
The control runs
benchmarks/PREREGISTRATION_0.2.md fixed them
before they ran: side K alone, from this tree before the tag, with the binary of its release workflow. Run 1 had three DeepSeek
readers with the gap pass; run 2 the four readers of four makers of 0.1.7, so that against 0.1.7 one thing changed, the kernel. Their
first versions found a defect of the repair, which was fixed and both runs were made again; the first runs are published too.
- After each withdrawal no claim the kernel reports as grounded rests on the withdrawn text: 0 of 940 and 936, and of 815 and 816.
- Every check of what 0.2 adds passed on both journals: the journal and every seal, 46 and 34 signed checkpoints, the report of what
fell, quittances for all 93 and 91 standing answer rows (none for the fallen ones), the index, and an independent verifier
written in a clean room from the specification of the format (yuce-verify; its code follows in a later commit). - Graded blind by Claude Fable 5.2 together with the K run of 0.1.7 and T and L of 0.1.6: the four makers scored 98.9% of correct
statements at every stage, against 86.5 to 88.8% for 0.1.7 with the same readers, and no earlier state written as current, against
8 to 12. Both withdrawals were kept in both label orders. With three DeepSeek readers the second withdrawal was kept partly: the
model read the withdrawal of one changelog line as the withdrawal of the whole revision. The grader found all four lines planted to
test it. One run per side: this is not a measure of answer quality.
One defect of the kernel was found after the grading, while counting repeats: after a repair in several attempts, the list of
records of one answer named 11 of its 54 records twice, and the report of what fell can count the standing grounds of such an
answer too high. It changes no answer's text and no result; the tag stays on the tree of the runs, and 0.2.1 is to list each
record of an answer once (LIMITATIONS.md).
audit-proof --out does not refuse the journal's own path: pointed at the journal, it writes the quittance over it. Keep the
default folder or name another file; 0.2.1 is to refuse it.
Results and caveats: BENCHMARKS.md.
Full list: CHANGELOG.md. What it does not do: LIMITATIONS.md.
Assets
- Binaries:
yuce-v0.2.0-<target>for Linux x86_64, Windows x86_64 and macOS (arm64, x86_64), each withsbgk, its alias
yuce, LICENSE, NOTICE, THIRD_PARTY_LICENSES.md and README;SHA256SUMS-<platform>.txt. - The Python client: organetic-yuce on PyPI.
- The control runs:
control02-run1-K,control02-run2-Kand their gradingcontrol02-grading-fable; the first runs
first02-run1-Kandfirst02-run2-K, whose four graded packets are in the grading archive. Each with a manifest, and
SHA256SUMS-runs.txt. A run archive holds its journal, the signed checkpoints of its seals and the verifier key that checks them;
the agents' tokens and the signing key are left out. Earlier runs stay in v0.1.7,
v0.1.6 and v0.1.0.
Check a download with sha256sum -c SHA256SUMS-<platform>.txt.
YuCe 0.1.7
YuCe 0.1.7: a repair names what fell, and the audit server takes your own text files.
The blind grading of the second run on four makers found the gap: after a withdrawal the kernel kept a conclusion on a page that still
stands, but its request for a repair did not say which row had fallen, and the model wrote the whole answer again without it.
Fixed
- A repair after a withdrawal names what fell. It names each fallen row by id, with the quotes under it that stand, and keeps the
row's own text out, as it may state what was withdrawn. A repair takes a few lines, and the rows that stand stay as they are. - Copies of the kept rows stay out of a repaired answer. A line that says what a kept row says, on the same grounds, is left out
and counted (dropped_copies). The same words with another date or on another quote are another claim. - A new quote looks for withdrawn text near itself, not through the whole document, with the same decision.
yuce --versionsaysyuce.
New
sbgk audit-serve --docs <folder>runs on your own text files: contracts, policies, logs, code, any UTF-8 text, with no manifest
and no repository. A document is named by its path inside the folder and bound by its bytes. Hidden files, service folders, links and
the server's own files stay out, and the server prints what it took and what it left. A file that is not text stops the start, by
name.--selectnames the files instead of walking the folder.audit-verifychecks the file set of such a journal (sources), and says whether the documents are the set the run started on.examples/findings-gate/puts a security review's findings through the kernel: a made-up finding is refused by its bytes, a lost
reader does not hang the work, and a file that changes takes the findings on it until each is checked again.
The journal format does not change: the release before verifies a journal of this one to the same roots.
The control run
benchmarks/PREREGISTRATION_0.1.7.md fixed it before it
ran: side K alone, from this tree before the tag, with the four readers of run 2. The repair took one new claim for each answer, no copy,
and the conclusion of run 2 stood. The grader marked the second withdrawal kept in one label order and partly in the other, for an
earlier state written as current that the answer held before the withdrawals. Results, with T and L of run 2 graded again:
BENCHMARKS.md.
Full list: CHANGELOG.md. What it does not do: LIMITATIONS.md.
Assets
- Binaries:
yuce-v0.1.7-<target>for Linux x86_64, Windows x86_64 and macOS (arm64, x86_64), each withsbgk, its alias
yuce, LICENSE, NOTICE, THIRD_PARTY_LICENSES.md and README;SHA256SUMS-<platform>.txt. - The Python client: organetic-yuce on PyPI.
- The control run:
control017-Kandcontrol017-grading-opus, with a manifest each andSHA256SUMS-runs.txt. The runs on four
makers stay in the v0.1.6 release, the earlier ones in
v0.1.0.
Check a download with sha256sum -c SHA256SUMS-<platform>.txt.
YuCe 0.1.6
YuCe 0.1.6: a withdrawal closes text of a document.
The blind grading of the first run on four makers found the gap: after a withdrawal a reader went back to the page and stored the
withdrawn sentence with its list number, three bytes wider, and the answer of the kernel's side cited it.
Fixed
- A wider quote no longer brings withdrawn text back. The kernel refused only the exact span, so any quote that contained it was a new
record with its support. A new quote that overlaps the withdrawn bytes is now refused withEVIDENCE_REVOKED, which names the span. The
text stays closed wherever those bytes stand in the document, in any version. - A withdrawal takes every stored quote that overlaps it. It reached one record: the others were found only by the benchmark's runner.
The answer now says how many other records went with it (also_revoked). A quote that went with another can still be withdrawn itself,
which closes the rest of its text. LIMITATIONS.mdsaid too much. "No claim stays grounded on a withdrawn quote" held for the record, not for its text. It is corrected,
andBENCHMARKS.mdnames the one wider quote of the published runs, which no answer cited.
New
audit_withdraw_textwithdraws text nobody has quoted yet, so a withdrawal no longer depends on a stored record. The Python client has
it aswithdraw_text.
Old journals
New journals record the rule in their genesis (r2r-audit-genesis/2). A journal of an earlier release replays under the rule it was written
with, so the published runs verify with the same roots. An earlier binary refuses a new journal.
Pre-registration
benchmarks/PREREGISTRATION_0.1.6.md fixes the repeat
of the run from this tag. That run and the run of 0.1.5 are in BENCHMARKS.md, both as they came out.
Full list: CHANGELOG.md.
Assets
- Binaries:
yuce-v0.1.6-<target>for Linux x86_64, Windows x86_64 and macOS (arm64, x86_64), each withsbgk, its alias
yuce, LICENSE, NOTICE, THIRD_PARTY_LICENSES.md and README;SHA256SUMS-<platform>.txt. - The Python client: organetic-yuce on PyPI.
- The two runs on four makers:
makers1-*andmakers2-*, one archive per run and per grading, with a manifest each andSHA256SUMS-runs.txt. - The earlier recorded runs of BENCHMARKS section 1 stay in the v0.1.0 release.
Check a download with sha256sum -c SHA256SUMS-<platform>.txt.
YuCe 0.1.5
YuCe 0.1.5: a Python client, and a runner that takes a team of several makers on every side. The kernel did not change.
Its code is that of 0.1.4, and journals verify as they did.
New
- A Python client of the audit server, in clients/python: one
connection per agent, the server's tools as methods, standard library only, Python 3.10 or newer. A refusal of the kernel is an ordinary
answer; an exception means the call itself failed. Install it withpip install organetic-yuceand import it asyuce: the nameyuceon PyPI belongs to an unrelated project.
Fixed
- The side on Temporal starts with a Codex or an Antigravity seat. Neither declared its longest call, which bounds a Temporal step, so
that side stopped at its start. - The money brake runs beside a seat without a price per call. With an Antigravity seat the runner refused the brake even when other
seats were paid. It now counts the paid seats. A team with no paid seat is refused as before. benchmarks/revision_outcomes.pycounts what it says. A quote that partly overlaps the withdrawn text is touched, found in the page's
bytes, and a reader's stored quotes come from the journal, apart from its other records.benchmarks/run_limits.pyfails a campaign whose journal does not verify. It marked it complete and exited 0.
Pre-registration
benchmarks/PREREGISTRATION_0.1.5.md replaces the
one of 0.1.4 for the run, which had not started: four readers of four makers, and the differences are listed in the file. That run is not
part of this release; its results will be added when it has been done.
Full list: CHANGELOG.md.
Assets
- Binaries:
yuce-v0.1.5-<target>for Linux x86_64, Windows x86_64 and macOS (arm64, x86_64), each withsbgk, its alias
yuce, LICENSE, NOTICE, THIRD_PARTY_LICENSES.md and README;SHA256SUMS-<platform>.txt. - The Python client: organetic-yuce on PyPI.
- The recorded runs of BENCHMARKS section 1 stay in the v0.1.0 release.
Check a download with sha256sum -c SHA256SUMS-<platform>.txt.
YuCe 0.1.4
YuCe 0.1.4: a journal checks about five times faster, the limits campaign is in the open, and the runner is ready for a run from this tag.
Roots, journal format and checks are unchanged: journals of earlier versions verify as they did.
Faster
- A journal replays about five times faster. On the recorded runs of about 3,000 records,
audit-verifywent from about 11 s to
2 s, and reopening a journal gains the same. Every entry's outcome and roots are still computed and compared, so a journal that
misreports one is refused as before. The whole state is still hashed at every entry, so the time grows close to the square of the
journal's length.
Fixed
- A refused start of
audit-serveno longer writes to a journal or leaves a new one. Every flag that needs no journal is checked
before the journal is opened, and a new run whose start fails is taken away again. One thing stays: opening a journal cuts a torn
tail, so a start refused by a check against that journal has already cut it. It says so at once and keeps the bytes. audit-verify --at-end. With--expect, the receipt's state must be the journal's last: a journal that went on after a terminal
receipt is a mismatch, and so are bytes after the last complete entry.conerefuses an entity no claim names (exit 2), and a container it cannot read. It printed an empty cone and exited 0.- Tables of records take a key per process, so an agent cannot search for record ids that fall in one bucket and slow every pass.
Measured and repeatable
benchmarks/run_limits.pyrepeats section 4 of BENCHMARKS.md
and says how busy the machine was. The section is taken again on this kernel code, and its memory column is now the peak of the whole
process.- The runner names the tag, the trees of
cratesandbenchmarks, the binary's version and the hash of every runner source in a
run's provenance, and refuses a binary of another version than its tag.benchmarks/revision_outcomes.pycounts what a withdrawal
cost and what it took with it, on every side. benchmarks/PREREGISTRATION_0.1.4.mdfixes what a run from this tag will measure, before it runs: a team of models from different
makers, the withdrawals, the grading rules and the outcomes. That run is not part of this release; its results will be added when it
has been done.
Changed in the texts
README now says what the audit server does: a claim rests on quotes, all of them required, and CURRENT_UNCERTAIN falls on a claim that
rested on an earlier and a later report. LIMITATIONS adds what a withdrawal reaches: in the journal, no claim stays grounded on a withdrawn
quote; the text an agent writes can still repeat it in other words, and that is measured, not prevented. BENCHMARKS section 3 keeps the
check of quotes alone: the comparison of answers with knowledge-graph systems, which have another purpose, is out until the graph work
of a later version.
Full list: CHANGELOG.md.
Assets
- Binaries:
yuce-v0.1.4-<target>for Linux x86_64, Windows x86_64 and macOS (arm64, x86_64), each withsbgk, its alias
yuce, LICENSE, NOTICE, THIRD_PARTY_LICENSES.md and README;SHA256SUMS-<platform>.txt. - The recorded runs of BENCHMARKS section 1 stay in the v0.1.0 release.
Check a download with sha256sum -c SHA256SUMS-<platform>.txt.
YuCe 0.1.3
YuCe 0.1.3: fixes from an outside review of 0.1.2, and the results stated as far as they go.
Two independent outside checks rebuilt the project and recomputed the published tables from the recorded runs: the numbers match.
They also found defects in the command line and in how the results were presented. This release closes them.
Fixed
- An unknown flag or a stray argument is refused. A typo used to switch a check or a limit off in silence:
audit-verify --expct receipt.jsonprintedVERIFIEDwithout reading the receipt. Each command now names its flags and how many
other arguments it takes; anything else exits 2 before the command runs. applyexits 1 when a step is refused.apply && verifyused to take a refused run for a sound one.- A journal whose end is no complete entry is no longer a clean pass. Without a receipt,
audit-verifyreportsTORN_TAIL
(exit 1). A resuming server still recovers, as it must after a crash, but says so at once and keeps the bytes beside the journal. audit-verifyon a file it cannot read exits 2, not 1.audit-servebinds its port before it creates the run.
update-docrefuses a document the container does not have. The inspector page refuses a malformed report.
Changed in the texts
README and BENCHMARKS no longer present a count of comparisons won or the model calls after a withdrawal as a result. The numbers
stand, but the samples are small, most of those calls had nothing to do with the withdrawal, and the kernel's side makes more
calls. We claim no gain in answer quality. See "How far these numbers go" in BENCHMARKS.md.
LIMITATIONS now says plainly what a journal cannot show without a receipt.
Full list: CHANGELOG.md.
Assets
- Binaries:
yuce-v0.1.3-<target>for Linux x86_64, Windows x86_64 and macOS (arm64, x86_64), each withsbgk, its alias
yuce, LICENSE, NOTICE, THIRD_PARTY_LICENSES.md and README;SHA256SUMS-<platform>.txt. - The recorded runs of BENCHMARKS section 1 stay in the v0.1.0 release.
Check a download with sha256sum -c SHA256SUMS-<platform>.txt.
YuCe 0.1.2
Use 0.1.3: it fixes silent flag typos,
applyexit codes and a damaged journal end, and states the results as far as they go.
YuCe 0.1.2: the benchmark suite reads sbgk verify's verdict again when a check fails.
0.1.1 made sbgk verify exit 1 when a check fails. The CLI wrappers of the benchmark suite took every non-zero exit for a
crash, so the tamper-attack experiment (python benchmarks/run_sbgk_suite.py --only 6), which verifies hundreds of tampered
containers and counts the ones caught, stopped at the first one. From 0.1.2 the wrappers take exit 1 from verify, with its
verdict on stdout, as the answer; every other non-zero exit is still a failure. A new test runs that experiment and checks
that every attack is caught.
The binaries behave as in 0.1.1 (verify: 0 verified, 1 a check failed, 2 usage error); the kernel, the audit server, the
demo and the benchmark results are unchanged. See CHANGELOG.md.
Assets
- Binaries:
yuce-v0.1.2-<target>for Linux x86_64, Windows x86_64 and macOS (arm64, x86_64), each withsbgk, its
aliasyuce, LICENSE, NOTICE, THIRD_PARTY_LICENSES.md and README;SHA256SUMS-<platform>.txt. - The recorded runs of BENCHMARKS section 1 (journals, gradings, manifests) stay in the
v0.1.0 release: they were recorded with 0.1.0, and neither
fix changes them.
Check a download with sha256sum -c SHA256SUMS-<platform>.txt.
YuCe 0.1.1
Use 0.1.2. With this release's exit code, the tamper-attack experiment of the benchmark suite (
run_sbgk_suite.py --only 6) stops at its first tampered container; 0.1.2 fixes the suite. The binaries are the same in behaviour.
YuCe 0.1.1: sbgk verify now fails with exit code 1 when a check fails.
In 0.1.0, sbgk verify <container> printed "ok": false for a tampered container but still exited 0, so a script or a CI
step that reads only the exit code would take a tampered container for a sound one. From 0.1.1 it exits 1 when a check
fails, as sbgk audit-verify already did. The exit codes are now: 0 verified, 1 a check failed, 2 usage error. A new test
covers all three: the quickstart container, the same container with one bit flipped, and no container.
Nothing else changed: the kernel, the audit server, the demo and the benchmark results are the same as in 0.1.0. See
CHANGELOG.md.
Assets
- Binaries:
yuce-v0.1.1-<target>for Linux x86_64, Windows x86_64 and macOS (arm64, x86_64), each withsbgk, its aliasyuce, LICENSE, NOTICE, THIRD_PARTY_LICENSES.md and
README;SHA256SUMS-<platform>.txt. - The recorded runs of BENCHMARKS section 1 (journals, gradings, manifests) stay in the
v0.1.0 release: they were recorded with 0.1.0, and this
fix does not change them.
Check a download with sha256sum -c SHA256SUMS-<platform>.txt.