styxx 7.50.0

styxx 7.50.0 — 2026-10-22

A minor release: the front door prints the sentence the lab's 2026-08-31 preregistration committed it to ("an agent's PR summary cannot lie about its diff" is not yet true of this instrument on real-world data; EXTERNAL-1, precision 0.23) and stops making the claim on every shipped surface, held there by a word lint in CI; the diff gate cuts a diff at git's own line ends and abstains where its two readers still read a line apart (#205), and a declared count is ASCII digits in both ports (#206); every diffgate record names the build and runtime that wrote it, and charon's verifier names its own; the research capsule, the horizon datasheet, the robustness preregistration and its correction, the real-model bridge's closure, the zoo screen and the sealed-ahead records; and what else merged on main after 7.49.0. Upgrade with pip install -U styxx==7.50.0. The PyPI summary changes with this release: the April positioning ("Cognitive observability", "HaluEval-QA AUC 0.998", a register-detection figure at a documented construct ceiling) is replaced by what the instruments do; earlier PyPI pages are superseded, not edited. The entries headed [Unreleased] until this cut are kept whole below, in the order this file carried them; entries written at this cut say so in their opening line.

What this release carries (17 entries, each with its receipts in CHANGELOG.md)

• conformance/sworn regenerated for 7.50.0: the version stamp moves every vector's id and no expected outcome
• the front door says, in EXTERNAL-1's committed words, that the gate claim is not yet true, and stops making it
• sealed-ahead: the build stopped on 2026-10-12; its battery, sizing week and stranger pool check are committed as records
• diffgate cuts a diff where git does and abstains where its two readers still read a line apart (#205); a declared count is ASCII digits in both ports (#206)
• the zoo D0 screen: 13 cached models on bridge v3's compare and recall at 16 tokens (pilot-class, licenses nothing)
• charon's verifier names its runtime: verifier scheme styxx.charon/verifier/v1, module shas over LF bytes and the runtime inside the digest (identity phase 3)
• a diffgate record names the build and runtime that wrote it: build and runtime after the eleven content keys (identity phase 2)
• the real-model bridge, attempt three: NOT STARTED, two pilots both NEITHER, the behavioural route closed
• PREREG: the permuted-MNIST headline under held-out permutations, new seeds and a 3x budget, sealed as a git blob before any run, with a CORRECTION sealed beside it before any data
• the research capsule, Phase A: one HTML file carries the chain of one preregistered experiment and prints three readings per link
• capsule v0.2 layer 2 compares a record's instrument identity apart from its content; styxx.instrument names the build and runtime a diffgate record is written and re-run under
• styxx.horizon: the horizon datasheet, an instrument from per-item bits to censored horizons, with the self-test that gates its printing
• capsule verify ends every hostile capsule in a named problem, never a traceback, and refuses the same file names on every system (#196)
• the GitLab job reports by default; blocking is opt-in with STYXX_SOFT_FAIL: "false"
• the leaderboard styxx leaderboard prints says what Baseline-019 measured, without the priority claim critique.py withdrew
• NOTE: the HANDOFF capsule reproduces under its minting build; a record that names no build cannot say so
• styxx 7.49.0 on Zenodo, as the next version of the software record

Integrity

artifact	sha256	bytes
wheel styxx-7.50.0-py3-none-any.whl	66744ee588eacfa413fe036b37fc4f7b171d15a86d2c0c31ceacf6f24a0560b3	8,208,487
sdist styxx-7.50.0.tar.gz	5f706fe0ba580e56155c37d4ca04f988ca14a87288189655ac6145871f1b7af5	9,041,414

The release assets and the PyPI files are the same bytes: the wheel PyPI serves hashes to the asset's digest (downloaded with pip and with gh on 2026-10-23), and PyPI's recorded sdist digest equals the asset's. A clean venv installs 7.50.0 with styxx.__version_mismatch__ None. Install: pip install -U styxx==7.50.0.

The claim "an agent's PR summary cannot lie about its diff" is not yet true of this instrument on real-world data (EXTERNAL-1, accusation precision 0.23 against a preregistered 0.95 floor; papers/closed-model-frontier/). The diff gate reports by default.

Archived on Zenodo as 10.5281/zenodo.23594959, the next version of the styxx software record (concept 10.5281/zenodo.19758618), with the source bundle of this tag, the wheel and the sdist.