chore(release): 1.4.2

CHANGELOG.md

@@ -7,6 +7,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [1.4.2] - 2026-10-12
+
 ### Fixed
 
 - A `semantic` evaluator set to `blocking: true` no longer fails runs whose every row

DOCS.md

@@ -12,7 +12,7 @@ hosted (opt-in)  run history, diffs, PR gates
 
 The suite is the crux, so the capture SDK is the recommended way to build one: it records real production runs to disk and `evalshift capture sync` promotes them into golden suites. Hand-written suites are fully supported — see [The golden suite](#the-golden-suite).
 
-- **Package name:** `evalshift` · **CLI entry point:** `evalshift` · **version:** 1.4.1
+- **Package name:** `evalshift` · **CLI entry point:** `evalshift` · **version:** 1.4.2
 - **Python:** >= 3.11 · **License:** Apache-2.0 · **Status:** stable
 - **Local-first.** Runs, scores, stats, and reports all happen on your machine under `.evalshift/`. The only network calls are the model API calls you asked for — and, if you opt in, pushes to the hosted service and reads from a capture bucket you own (`captures.store`).
 - **Four pieces:** CLI (this doc), SDK, GitHub Action, hosted server — each with its own machine-readable reference for AI tools. See [Ecosystem and AI-tool references](#ecosystem-and-ai-tool-references).

docs/evaluators.md

@@ -65,6 +65,16 @@ part of the answer and kept; an empty fenced block counts as an empty
 output. Whether an output is valid, unfenced JSON is a job for
 `structural.json_schema` or an `llm_judge` criterion.
 
+**Drift inside `min_similarity` is no change.** Because the source is
+pinned at 1.0, any rewording is a small negative delta. A row that stays
+within `min_similarity` therefore counts as *equivalent* in the policy
+budgets and enters the paired significance test as a delta of 0 — nine
+rows at 0.90–0.98 used to test as a `critical` regression and fail a run
+whose every budget passed. Only flagged rows (`SEMANTIC_REGRESSION`) carry
+their full drop into severity. `delta_avg_score` still shows the raw
+drift, and the comparison notes how many rows were within the threshold.
+See [methodology](methodology.md#drift-inside-min_similarity-is-tested-as-no-change).
+
 Use when:
 * You don't have a clean structural check.
 * You want to detect "wandered off" outputs that still look fine
@@ -205,7 +215,10 @@ refusal changes scored separately.
 You rarely wire the first two by hand: `evalshift capture sync` writes
 `tool_selection` and `tool_arguments` into each suite's own `suites:` entry
 based on what that suite's captures contain, so a tool-free suite gets no
-tool evaluators at all. See
+tool evaluators at all. A tool evaluator it inherits from the top level —
+`trace_invariants` included — skips every pair that was offered no tools:
+no row, no coverage entry, no effect on the verdict, and `evaluate` (or
+`compare`, under its verdict) warns that it checked nothing. See
 [Agent migrations](agents.md#configuration) and
 [Configuration → per-suite evaluators](configuration.md#per-suite-evaluators).
 

llms-full.txt

@@ -1,7 +1,7 @@
 # evalshift (CLI) — complete reference for AI tools
 
 Canonical hosted copy: https://www.evalshift.dev/cli-llms-full.txt
-Package: evalshift (PyPI) | CLI entry point: evalshift | version: 1.4.1
+Package: evalshift (PyPI) | CLI entry point: evalshift | version: 1.4.2
 Python: >=3.11 | license: Apache-2.0 | status: stable
 Install: pip install evalshift   (or: uv pip install evalshift)
 

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "evalshift"
-version = "1.4.1"
+version = "1.4.2"
 description = "Run your prompts on two LLMs and find out, with statistical confidence, what regressed."
 readme = "README.md"
 license = "Apache-2.0"