Tags: VectorInstitute/backcheck

backcheck v0.2.0

Adds backcheck demo, makes the Stop hook advisory by default, and fixes
several false positives including backcheck flagging its own fixtures.
The README example is now generated from real output and guarded by a
test.

backcheck v0.1.8

'All checks pass' is now recognised and verified against whatever ran,
rather than being forced into the tests bucket. A silent tsc pass no
longer looks unproven.

backcheck v0.1.7

Adds --explain, which shows what was recognised and what ran that was
not. Reads npm lint scripts, shell loops and gh run list. Resolves
silent tools from && semantics rather than from a neighbour's output.

backcheck v0.1.6

Fixes a class of false 'supported' verdicts where one tool's output was
read as another's. Adds timeout-wrapped runs, gh run list/view, and
import-linter. Splits the runners module for contributors.

backcheck v0.1.5

Reads CI check results and pre-commit output as evidence. Nine further
verdict errors fixed, found by hand-adjudicating another dozen real
sessions.

backcheck v0.1.4

Eleven verdict errors fixed, found by hand-adjudicating six real
sessions. Chief among them: output from a chained command is now
attributed to the step that produced it, so one tool's result can no
longer be cited as evidence for another's.

backcheck v0.1.3

Inconclusive claims are no longer summarised as supported. Build output
from vite, Next, mkdocs, sphinx, python -m build and docker is now read
rather than reported as unreadable.

backcheck v0.1.2

First crates.io release. No behaviour changes since v0.1.1; this exists
so the publish runs under the hardened workflow, where the registry
token is reachable only from a v* tag and only after approval.

backcheck v0.1.1

Corrects the declared minimum Rust version to 1.85 and ships the CI,
clippy, logo, and documentation fixes made since v0.1.0. No behaviour
changes to the analysis itself.

backcheck v0.1.0

First release. Verifies a Claude Code agent's claims against the evidence
in its own session transcript, with zero model calls.

- claims: tests/typecheck/lint/build pass, committed, pushed, file written
- verdicts: supported, qualified, unsupported, contradicted, inconclusive
- test integrity: skip markers added, assertions weakened or deleted,
  test files removed
- runs as a Stop hook, a CLI, or in CI via --json
- validated against 78 real sessions (213 MB, 1.3 s)