v0.2.0

Added

• check --judge-url <url> --judge-model <name> (or SNIFFTEST_JUDGE_URL and
  SNIFFTEST_JUDGE_MODEL) answers the judgment rules with a model you run
  yourself, through any Chat Completions server such as Ollama or LM Studio.
  On localhost nothing leaves the machine, no key is needed and no send
  question is asked; any other host is a named destination with its own
  consent. Jev stays the default. Its answers are cached under the local
  model's name, never Jev's. (#6)

Changed

• The check usage line identifies Jev and carries the input-token rate beside
  the estimated cost, marked as an unsourced price. --quiet hides that line
  without hiding flags or diagnostics.

v0.1.0

Added

• snifftest check reads Markdown and plain text and prints one line per flag:
  file, line, rule, probability, message. --format json carries every reading
  for a machine, including the ones under the threshold.
• A YAML ruleset, rules/default.yaml, with fifteen rules. Five countable
  rules run as regular expressions and never touch the network. Ten judgment
  rules are answered by a hosted judgment model, one paragraph per request,
  one probability per rule. Five of the judgment rules are tagged marketing
  and sit out an ordinary run until --only marketing asks for them.
• A .snifftest.yaml in the project extends or replaces the shipped ruleset,
  merged rule by rule on id. Countable rules take a built-in or a pattern of
  your own; judgment rules take a description, its near misses, examples and
  two criteria.
• --dry-run runs the countable rules only and sends nothing. The first time
  a judgment rule would send text, the CLI prints what leaves the machine and
  where, and asks. --yes answers and remembers; SNIFFTEST_SEND answers for
  one run in CI and names the destinations it covers.
• A no-judgment band. A probability between 0.4 and 0.6 is reported as no
  judgment and never counts as a flag or a pass. A run whose judgment arm
  answered nothing exits 2 rather than 0.
• Partial results. A paragraph that fails is skipped and the run carries on;
  three transient failures in a row stop the judgment arm; a rejected key
  stops it at once. Every answer already received stays in the verdict.
• An on-disk answer cache, so a rerun after an outage does not pay for every
  paragraph again. Entries expire after fourteen days and never hold the
  paragraph's text. --no-cache and SNIFFTEST_CACHE_DIR=off turn it off.
• snifftest rules prints the ruleset in force and where each rule came from.
• snifftest eval plants one known fault per rule into your own clean text,
  from an independent seed bank, runs the corpus three ways and reports what
  each arm caught, what it flagged on clean paragraphs, what it cost and how
  long it took. --twins measures what a sentence addressed to the checker
  does to the readings around it.
• snifftest bench asks a panel of general models the same questions over the
  same corpus and puts cost, speed and accuracy side by side, with consent
  asked per destination.
• A bench row can be called on OpenAI's own API with OPENAI_API_KEY, so an
  OpenAI model can be measured without an OpenRouter account and its latency is
  the provider's rather than a proxy's. bench/panel-direct.yaml is a panel of
  nothing but direct rows, and SNIFFTEST_SEND=openai answers the send question
  for that destination and no other.
• snifftest serve opens a local page where you type and the nose reacts,
  live or from a recorded replay that needs no key. serve --record writes
  that replay from a live run, with the measured time and cost of each answer.
• The judgment arm carries a budget, twenty seconds a paragraph with a floor
  of a minute, and a run that gets no answer at all exits non-zero instead of
  reporting a pass. An unknown tag on --only or --skip is a usage error.
• The tarball is checked against an allowlist in both release workflows, and
  the hook, the Action and the plugin install with --ignore-scripts.
• A git pre-commit hook, two entries for the pre-commit framework, and a
  composite GitHub Action. All of them run the countable rules only unless you
  opt in, and a pull request from a fork never sends anything.
• A Claude Code plugin and skill, /snifftest, that runs the free pass by
  default and never answers the sending question on the user's behalf.
• An example corpus of ten documents, a structure set, three adversarial twins
  and a committed eval run under bench/results/2026-09-17/, with the direct
  panel's bench run beside it: Claude Haiku 4.5, Sonnet 5 and Opus 5 and
  gpt-5.6-sol on their own APIs, cost and latency measured, one paragraph per
  request. The README's numbers table reads from that run.
• The nose: one hand-drawn SVG with six expressions, and a favicon cut.