Skip to content

YuCe 0.2.0

Choose a tag to compare

@OrganeticSphere OrganeticSphere released this 12 Nov 19:05
· 52 commits to main since this release

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A source your agents quoted is withdrawn: the kernel lists the conclusions that stood on it, from the record and not from a model,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 (or audit_fallen on 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, diff and record read the same
    journal.
  • Quittances (yuce-proof/1). sbgk audit-proof cuts 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-verify checks 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, and audit-verify --checkpoint with a file of trusted keys.
  • The index beside a journal. sbgk audit-index builds a SQLite database any tool reads, rebuilt from the journal at any time;
    --check compares 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 with sbgk, 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-K and their grading control02-grading-fable; the first runs
    first02-run1-K and first02-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.