v1.6.0

Documentation and package metadata only. The code is the same as 1.5.0.

• The README opens with what Crucible decides, who uses it, a first run checked in a clean environment, a comparison with promptfoo, DeepEval, Inspect, OpenAI Evals and Ragas, and its limits.
• The licence section states the FSL-1.1-MIT terms, and README images use absolute URLs so the PyPI page shows them.
• The PyPI summary is rewritten, and the author entry gains zaindharper@gmail.com.

Install: pip install crucible-bench==1.6.0

v1.5.0

Adds judging modes for when a model or a panel does the judging, measured once with a 2B local
judge. Both-order pairwise lost to pointwise scoring for that judge; decomposition made the named
weakest claim more stable across reruns.

• Pairwise judging in both orders: crucible.pairwise asks a comparator with each answer first and
  returns a winner only when both orders agree; a split is ORDER_DISAGREE, never a preference.
  pairwise_report gives order-swap consistency and the first-slot rate with Wilson intervals, and
  PairwiseMeasure puts the check on the measure seam.
• Decomposed claims: crucible.decompose splits a claim into typed sub-questions (yes or no, number,
  choice), decides each with the same pure verdict function, combines them demote-only and names the
  single weakest sub-question. The parent claim's hash is unchanged.
• Minority views kept: crucible.views and report --views list the majority beside every minority
  position with its sources, and dissent_survival scores any summary on the minorities it kept.
• New CLI commands pairwise, decompose and views, and MCP tools crucible.pairwise,
  crucible.decompose and crucible.views (17 tools in all).
• Measured with a 2B local judge in docs/benchmarks/judging-v1: order-swap consistency 0.07, so
  both-order pairwise (0.07 accuracy) lost to pointwise scoring (0.87) for that judge; decomposition
  raised rerun agreement on the weakest claim from 0.20 to 0.70. Human-preference agreement is untested.
• MCP tool schemas move to crucible.mcp_schema; mcp_tools re-exports them. The status envelope
  now lists MCP tools from the schema instead of a second hand-kept list.
• Claude plugin listing: the Claude manifest in client-plugin/ gains a display name, keywords,
  homepage, documentation, support, privacy and terms links, and a 1024 px icon. The workspace,
  measurement-command consent and measurement command now come from userConfig, so Claude Code
  asks for them when you enable the plugin and the launch arguments carry no placeholder. The
  portable and Codex manifests are unchanged.
• The client README gains a data and network table, and the client privacy note names the two
  temporary files the optional measurement command uses.
• The client README and privacy note gain a "What this plugin runs and handles" section: hooks,
  the exact Claude launch command, network use, files written and the environment variables read.
• The environment section now lists what a traced run of the server read, split into Crucible's
  own reads and Python's standard library reads (argument parser and tempfile).
• The server code now ships inside the plugin folder at client-plugin/server/src/, so a plugin
  installed from that folder alone starts. python scripts/build_client_package.py --sync-vendored
  rewrites it from src/, and a test fails when the copy drifts. The launcher no longer falls back
  to the repository's src/; a missing copy exits with a one-line message.
• crucible.measurement_gate describes itself in plain terms, and the client privacy note says the
  measurement command gets two temporary files per run.

Crucible 1.4.0

Adds local client packages with explicit launch permissions. Client-specific installation
and marketplace acceptance remain separate qualification steps.

• Add a local MCP profile with an explicit readable workspace, typed permission refusals, and process/network denial at launch. The full CLI/MCP remains available separately.
• Add portable source plugins for Claude/Codex-compatible clients and Windows x64 ZIP/MCPB packages containing the runtime, licenses, source hashes and checksums.
• Check actual stdio behavior before packaging, reject untracked or credential-like release payloads, and attach checked client assets to the same product release.
• Add an explicit launch-time grant for one fixed measurement command. Reuse bounded subprocess measurement and witnessed assessment; tool arguments and ambient environment cannot grant execution. The approved child has OS user permissions, not sandboxed access.
• Keep network retrieval and persistent-state operations on the full CLI/MCP. The client profile provides local assessment and explicitly approved measurement commands.

crucible 1.3.0

Verdict-integrity fixes, a release that carries them, and the files a user needs after
pip install. Upgrade from 1.2.0 or earlier: the fixes below change verdicts that 1.2.0 accepted.

Security and verdict integrity

Items marked "1.2.0 and earlier" shipped in a release; a GitHub security advisory for 1.2.0
accompanies this release. Items marked "unreleased" were defects in code added after 1.2.0 and
fixed before any release carried them.

• HIGH, 1.2.0 and earlier. A measurement could widen the tolerance that decides MATCH or DRIFT and
  still re-derive MATCH: the deciding number was bound by no seal, a tolerance on a thesis claim
  was ignored, and verify_assessment and recheck_assessment recomputed from the stored
  tolerance. A claim may now seal its tolerance ("tolerance" on a thesis claim, or
  make_claim(..., tolerance=)), and verdict_for refuses a measurement carrying any other
  tolerance with UNVERIFIABLE. A hand-forged MATCH with recomputed seals fails
  verdicts_rederive. Reproduction: a claim that seals tolerance 0.5, a measured deviation of 8
  and a measurement tolerance of 10 returned MATCH on 1.2.0 and returns UNVERIFIABLE now
  (tests/fixtures/sealed-tolerance/). Claims without a sealed tolerance keep their hashes and the
  old behavior; seal the tolerance on every new claim.
• HIGH, 1.2.0 and earlier. doctor emitted MATCH for three checks it never ran, and the status and
  demo envelopes defaulted to MATCH. Informational envelopes now use the operational token OK,
  and doctor resolves each capability's entry point and reports available or absent.
• MEDIUM, 1.2.0 and earlier. verify_browser_evidence turned a packet's own carried MATCH into
  crucible's MATCH after a shape check. A well-formed packet carrying MATCH is now UNVERIFIABLE,
  with the packet's claim kept under carried_verdict; a carried DRIFT or UNVERIFIABLE still
  passes through.
• MEDIUM, 1.2.0 and earlier. CLI oracle replay packs must carry the template's top-level
  assessment binding. A missing, malformed or mismatched thesis ID, assessment seal or measurement
  seal now fails closed before replay instead of letting an unbound pack run.
• LOW, 1.2.0 and earlier. The Telos content check accepted content on a 32-bit FNV-1a match alone.
  When a descriptor carries a SHA-256 (content_sha256), that digest decides and a mismatch is
  refused; FNV stays for protocol parity and is documented as accident-evident only.
• Unreleased. refine graded correctness against the measurement's tolerance instead of the
  claim's sealed tolerance, so a widened tolerance read as "cohesively verified" while
  verdict_for refused it. The grader now honors the seal the same way.
• Unreleased. export, the bundle spec.json and the cleanroom review reconstruction dropped a
  sealed tolerance, so a thesis with a sealed tolerance failed its own review. The tolerance is now
  carried on all three surfaces when sealed; unsealed rows stay byte-identical.

Added

• ProofMeasure: a proof or type checker (Lean, Coq, any command) as a measurement oracle. An
  accepted proof is MATCH, a rejected one DRIFT, and an absent or failing checker UNVERIFIABLE. The
  measurement carries a recheck descriptor with the command and the artifact's SHA-256. A checker
  command must be a non-empty list of non-empty strings, or the claim is UNVERIFIABLE. Library
  only; the CLI and MCP server do not start it.
• MCP: crucible.recheck_template returns crucible.replay-template/1 over stdio, and
  crucible.recheck accepts template: true while keeping pack and template exclusive. Replay
  packs that recheck to ok: false are normal tool results, not transport errors.
• Replay interop: crucible.replay-template/1 outputs carry a privacy-bounded
  crucible.replay-set/1 binding over the descriptor-bearing replay contracts, without exporting
  descriptorless rows or their evidence. crucible.replay-pack/1 inputs preserve that binding;
  supplied bindings are checked exactly, while older bound packs may omit them. A present pack
  schema must be exactly crucible.replay-pack/1; schema-less legacy packs stay compatible.
• crucible examples --out DIR writes the quickstart inputs, which now ship inside the package, to
  a new folder, so the README's first run works after pip install without a clone. The command
  refuses a folder that already exists.
• verify_seals.py at the repository root: a stdlib-only verifier that re-derives an assessment
  artifact's three seals without importing crucible.
• Organ-bundle interop (crucible.interop) and crucible.interop.json, the manifest plexus reads
  to wire crucible into the tool mesh.

Changed

• License: FSL-1.1-MIT (LicenseRef-FSL-1.1-MIT) replaces the crucible Fair-Source License 1.0
  for this and later versions. Versions already published keep the license they shipped with.
• SECURITY.md names the private reporting route and the supported versions.
• One version, checked everywhere: scripts/check_version_sites.py fails when pyproject.toml,
  crucible.__version__, the README or the latest changelog heading disagree, and the release
  workflow runs it against the tag. status derives its version text from the package.
• Release workflow: every action is pinned by commit. A read-only build job checks the tag and
  version sites, builds, runs twine check, and smoke-tests the wheel in a fresh environment
  (version, sealed-tolerance probe, packaged examples, MCP server). Only the upload job can mint a
  publishing token, and its upload tolerates a re-run (skip-existing). A third job attaches the
  wheel, the sdist and SHA256SUMS.txt to the GitHub Release, downloaded from PyPI and checked
  against PyPI's digests and this run's build.
• CI runs the version guard and the wheel smoke test.
• README: the install section shows the packaged quickstart, and presentation surfaces carry the
  status, doctor, MCP and CI commands.

v1.2.0

What's Changed

• MATCH-provenance gate: witnessed vs asserted verdicts (--require-witnessed-match) by @HarperZ9 in #4
• Browser-evidence pipeline (parallel-session work, preserved + integrated) by @HarperZ9 in #6
• feat: explain missing evidence for UNVERIFIABLE and warn on ill-posed measurements by @HarperZ9 in #5
• feat(ci): native CI regression gate over registry verdicts by @HarperZ9 in #7
• Style: spectrum banner + feature-first README header by @HarperZ9 in #8
• Release v1.2.0 by @HarperZ9 in #9

Full Changelog: v1.1.0...v1.2.0

Crucible 1.1.0

Crucible checks claims against evidence and returns a closed verdict: MATCH, DRIFT, or UNVERIFIABLE. First release published to PyPI (crucible-bench) via OIDC Trusted Publishing.

• crucible run THESIS --registry DIR (--measurements FILE | --substrate FILE) runs the steelman -> measurement -> witnessed assessment -> disk-recheck path as one session; --json emits a complete run record and --report FILE a deterministic Markdown report.
• Creative measurement gate (crucible measurement-gate) verifies Telos measurement packets without exporting raw pixels/assets/prompts; keeps the decision outcome separate from the verification verdict.
• MCP parity across the status/doctor/assess/run/review/report/batch/registry/drift/refine/verdicts workflows through the same CLI contract.
• stdlib-only, zero runtime dependencies; installable console script crucible.

Install

pip install crucible-bench
crucible --version

MPL-2.0.