Detect low-confidence spans from token logprobs, live mode, release binaries

.github/workflows/ci.yml

@@ -16,5 +16,11 @@ jobs:
           components: clippy, rustfmt
       - uses: Swatinem/rust-cache@v2
       - run: cargo fmt --all -- --check
-      - run: cargo clippy -- -D warnings
+      - run: cargo clippy --all-targets -- -D warnings
+      - run: cargo clippy --all-targets --no-default-features -- -D warnings
       - run: cargo test
+      - name: Smoke test the offline detector on a bundled sample
+        run: |
+          cargo run -q -- --logprobs-file examples/logprobs/cuyp.json --threshold 0.6 --format json | tee report.json
+          grep -q '"text": " Dordrecht"' report.json
+          cargo run -q --example detect

.github/workflows/release.yml

@@ -0,0 +1,57 @@
+name: Release
+on:
+  push:
+    tags: [ "v*" ]
+permissions:
+  contents: write
+env:
+  CARGO_TERM_COLOR: always
+  BIN: llm-token-visualizer
+jobs:
+  build:
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - { os: ubuntu-latest, target: x86_64-unknown-linux-gnu }
+          - { os: macos-latest, target: aarch64-apple-darwin }
+          - { os: macos-latest, target: x86_64-apple-darwin }
+          - { os: windows-latest, target: x86_64-pc-windows-msvc }
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: dtolnay/rust-toolchain@stable
+        with:
+          targets: ${{ matrix.target }}
+      - uses: Swatinem/rust-cache@v2
+        with:
+          key: ${{ matrix.target }}
+      - run: cargo build --release --locked --target ${{ matrix.target }}
+      - name: Package
+        shell: bash
+        run: |
+          name="$BIN-${GITHUB_REF_NAME}-${{ matrix.target }}"
+          mkdir -p "dist/$name"
+          if [ "${{ runner.os }}" = "Windows" ]; then exe="$BIN.exe"; else exe="$BIN"; fi
+          cp "target/${{ matrix.target }}/release/$exe" README.md LICENSE "dist/$name/"
+          cp -r examples/logprobs "dist/$name/samples"
+          cd dist
+          if [ "${{ runner.os }}" = "Windows" ]; then 7z a "$name.zip" "$name" >/dev/null; else tar czf "$name.tar.gz" "$name"; fi
+      - uses: actions/upload-artifact@v4
+        with:
+          name: ${{ matrix.target }}
+          path: dist/*.*z*
+  release:
+    needs: build
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          path: dist
+          merge-multiple: true
+      - name: Checksums
+        run: cd dist && sha256sum * > SHA256SUMS.txt
+      - uses: softprops/action-gh-release@v2
+        with:
+          files: dist/*
+          generate_release_notes: true

CHANGELOG.md

@@ -0,0 +1,15 @@
+# Changelog
+
+## 0.2.0 (2026-09-25)
+
+- New: hallucination-risk detection from token logprobs. `--logprobs-file` reads a Chat Completions response (or its `logprobs` block, or a bare token array) and flags words below `--threshold` (default 0.5), with the alternatives the model considered.
+- New: `--live "<prompt>"` calls any OpenAI-compatible API (`OPENAI_API_KEY`, optional `OPENAI_BASE_URL`, `--model`, `--save`). Behind the default `live` feature.
+- New: `--format json` report and `--fail-on-flag` (exit status 2) for scripts and CI.
+- New: four real Llama 3.1 8B sample responses in `examples/logprobs/` and `cargo run --example detect`.
+- Fix: `TerminalRenderer::render` no longer prints to stdout itself, so library callers of `visualize_tokens(.., "terminal", ..)` control output. The CLI prints the returned string.
+- Release workflow builds Linux, macOS and Windows binaries on tag push.
+- Repo: build output (`target/`) is no longer tracked; sketch directories renamed.
+
+## 0.1.0
+
+- Token confidence visualizer: terminal, HTML and Markdown renderers.

CONTRIBUTING.md

@@ -1,22 +1,34 @@
-# Contributing to LLM-Hallucination-Detection-Script
+# Contributing
 
-Thanks for your interest. Contributions that improve detection accuracy, add new methods, extend provider integrations, or add test coverage are all welcome.
+Thanks for helping. Small, focused PRs are easiest to review.
 
-## What we want
+## Setup
 
-- **New detection methods** -- additional heuristics or ML-based approaches
-- **Provider integrations** -- examples using Mistral, Gemini, local models
-- **Rust MVP improvements** -- better streaming detection, new dashboard features
-- **Bug fixes** -- correctness issues, edge cases, false positive reduction
-- **Tests** -- more labeled examples, accuracy benchmarks
+You need a stable Rust toolchain.
 
-## How to contribute
+```bash
+git clone https://github.com/Mattbusel/LLM-Hallucination-Detection-Script
+cd LLM-Hallucination-Detection-Script
+cargo test
+cargo run -- --logprobs-file examples/logprobs/cuyp.json --threshold 0.6
+```
 
-1. Fork and clone
-2. Python: `python -m pytest tests/` must pass
-3. Rust: `cargo test --all` must pass
-4. Open a PR with a description of the change and why
+## Before you open a PR
 
-## Questions
+CI runs these; run them locally first:
 
-Open a [Discussion](https://github.com/Mattbusel/LLM-Hallucination-Detection-Script/discussions) for design questions or ideas before building.
+```bash
+cargo fmt --all -- --check
+cargo clippy --all-targets -- -D warnings
+cargo clippy --all-targets --no-default-features -- -D warnings
+cargo test
+cargo run --example detect
+```
+
+## Good places to start
+
+- Detection logic lives in `src/detect.rs`, with unit tests at the bottom of the file. New heuristics should come with a test and, ideally, a real sample response in `examples/logprobs/`.
+- Sample responses must be real API output (say which model and provider in the PR). Do not hand-edit logprobs.
+- `rust_mvps/` is design sketches that are not built. Turning one into a working crate is welcome, but open an issue first.
+
+Issues labelled `good first issue` are sized for a first PR.