evalseal 0.1.0__tar.gz

evalseal-0.1.0/.env.example

@@ -0,0 +1,4 @@
+# Copy to .env and fill in. .env is gitignored — never commit a real key.
+# Free Gemini key: https://aistudio.google.com/apikey
+# Only used when recording (EVALSEAL_RECORD=1); replay and CI need no key.
+EVALSEAL_API_KEY=

evalseal-0.1.0/.github/workflows/ci.yml

@@ -0,0 +1,33 @@
+name: CI
+on: [push, pull_request]
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.11", "3.12"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with: { python-version: "${{ matrix.python-version }}", cache: pip }
+      - run: pip install -e ".[dev]"
+      - name: Tests (no API key — cassettes replay)
+        run: pytest -q --cov=evalseal
+      - name: Demo run reproduces from cassette
+        run: |
+          if [ ! -f tests/cassettes/run.json ]; then
+            echo "::warning::tests/cassettes/run.json not recorded yet; skipping demo replay."
+            exit 0
+          fi
+          set +e
+          evalseal run \
+            --dataset examples/borderline_judge/dataset.jsonl \
+            --target-config examples/borderline_judge/target.json \
+            --scorer-config examples/borderline_judge/scorer.json \
+            --n 5 --cassette tests/cassettes/run.json
+          code=$?
+          # 0 = all stable, 3 = unstable cases found (expected for this demo).
+          # Anything else means the replay itself broke.
+          if [ "$code" -ne 0 ] && [ "$code" -ne 3 ]; then exit "$code"; fi
+      - name: Ledger integrity
+        run: evalseal verify

evalseal-0.1.0/.github/workflows/release.yml

@@ -0,0 +1,52 @@
+name: Release
+on:
+  push:
+    tags: ["v*"]
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with: { python-version: "3.12", cache: pip }
+      - name: Tag matches package version
+        run: |
+          version=$(python -c 'import tomllib; print(tomllib.load(open("pyproject.toml", "rb"))["project"]["version"])')
+          if [ "${GITHUB_REF_NAME#v}" != "$version" ]; then
+            echo "::error::tag $GITHUB_REF_NAME does not match pyproject version $version"; exit 1
+          fi
+      - run: pip install -e ".[dev]" build twine
+      - name: Tests (no API key — cassettes replay)
+        run: pytest -q
+      - run: python -m build
+      - run: twine check dist/*
+      - uses: actions/upload-artifact@v4
+        with: { name: dist, path: dist/ }
+
+  publish-pypi:
+    needs: build
+    runs-on: ubuntu-latest
+    # Trusted publishing: PyPI verifies this workflow via OIDC. No token is stored.
+    environment:
+      name: pypi
+      url: https://pypi.org/project/evalseal/
+    permissions:
+      id-token: write
+    steps:
+      - uses: actions/download-artifact@v4
+        with: { name: dist, path: dist/ }
+      - uses: pypa/gh-action-pypi-publish@release/v1
+
+  github-release:
+    needs: publish-pypi
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    steps:
+      - uses: actions/download-artifact@v4
+        with: { name: dist, path: dist/ }
+      - name: Create GitHub release
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: gh release create "$GITHUB_REF_NAME" dist/* --repo "$GITHUB_REPOSITORY" --title "$GITHUB_REF_NAME" --generate-notes

evalseal-0.1.0/.gitignore

@@ -0,0 +1,223 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[codz]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#   Usually these files are written by a python script from a template
+#   before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py.cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+# Pipfile.lock
+
+# UV
+#   Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+# uv.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+# poetry.lock
+# poetry.toml
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#   pdm recommends including project-wide configuration in pdm.toml, but excluding .pdm-python.
+#   https://pdm-project.org/en/latest/usage/project/#working-with-version-control
+# pdm.lock
+# pdm.toml
+.pdm-python
+.pdm-build/
+
+# pixi
+#   Similar to Pipfile.lock, it is generally recommended to include pixi.lock in version control.
+# pixi.lock
+#   Pixi creates a virtual environment in the .pixi directory, just like venv module creates one
+#   in the .venv directory. It is recommended not to include this directory in version control.
+.pixi
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# Redis
+*.rdb
+*.aof
+*.pid
+
+# RabbitMQ
+mnesia/
+rabbitmq/
+rabbitmq-data/
+
+# ActiveMQ
+activemq-data/
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.envrc
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#   JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#   be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#   and can be added to the global gitignore or merged into this file.  For a more nuclear
+#   option (not recommended) you can uncomment the following to ignore the entire idea folder.
+# .idea/
+
+# Abstra
+#   Abstra is an AI-powered process automation framework.
+#   Ignore directories containing user credentials, local state, and settings.
+#   Learn more at https://abstra.io/docs
+.abstra/
+
+# Visual Studio Code
+#   Visual Studio Code specific template is maintained in a separate VisualStudioCode.gitignore 
+#   that can be found at https://github.com/github/gitignore/blob/main/Global/VisualStudioCode.gitignore
+#   and can be added to the global gitignore or merged into this file. However, if you prefer, 
+#   you could uncomment the following to ignore the entire vscode folder
+# .vscode/
+# Temporary file for partial code execution
+tempCodeRunnerFile.py
+
+# Ruff stuff:
+.ruff_cache/
+
+# PyPI configuration file
+.pypirc
+
+# Marimo
+marimo/_static/
+marimo/_lsp/
+__marimo__/
+
+# Streamlit
+.streamlit/secrets.toml
+
+# EvalSeal local state and generated reports
+.evalseal/
+report.json
+report.md

evalseal-0.1.0/DESIGN.md

@@ -0,0 +1,76 @@
+# EvalSeal design
+
+## What it claims, and what it doesn't
+
+**EvalSeal measures reproducibility. It does not create determinism.** Setting a
+temperature or seed does not make a hosted model deterministic, and EvalSeal never says it
+does. It runs every case N times and reports how often the verdict disagrees with itself.
+
+**A green CI replay proves the harness is reproducible, not that the model is.** Replaying
+a cassette gives back the recorded responses, so the analysis, scores and report can be
+regenerated from them exactly. It says nothing about whether the live model would answer
+the same way today. To learn that, record again and `diff` the runs.
+
+**The ledger is tamper-evident, not tamper-proof.** Each record's SHA-256 covers its whole
+content plus the previous record's hash. Editing a past score breaks that record's hash,
+and re-hashing it breaks the next record's link. Anyone who can rewrite the *whole* file
+can build a new valid chain, though. Anchoring the head hash somewhere external
+(signing, notarization) is out of scope for v0.
+
+## Components
+
+| module | role |
+|---|---|
+| `analyze.py` | Pure functions: seeded bootstrap CI, flip rate, stability class. No I/O. |
+| `adapters/recording.py` | Cassette keyed by the SHA-256 of the effective request body. |
+| `adapters/target.py` | `Target` protocol; `LocalCallableTarget` (tests), `OpenAICompatibleTarget`. |
+| `adapters/scorer.py` | Exact, regex, and LLM-judge scorers. The judge is a `Target`. |
+| `executor.py` | N-run loop, provenance capture, provenance-gap warnings. |
+| `ledger.py` | Hash-linked append-only JSONL; `verify_chain`. |
+| `report.py`, `cli.py` | `report.md` / `report.json`; `run`, `verify`, `diff`. |
+
+## Decisions
+
+**The cassette stores responses as ordered lists.** Repeats of a case send identical
+requests. If the cassette kept one response per request hash, repeat #1 would be replayed
+N times and every recorded flip would disappear. Instead, the k-th identical call gets the
+k-th recorded response. This depends on calls happening in the same order in record and
+replay mode, which holds because the executor is sequential.
+
+**Only the effective request is hashed.** The cassette key is the URL plus the JSON body
+actually sent. Unset parameters are left out of the body entirely, so "temperature not
+set" and "temperature=1.0" are different requests. API keys never enter the key or the
+cassette.
+
+**Provenance is checked on every response.** A served-model mismatch on call 37 matters
+even when call 100 looks clean. The manifest records the first response's metadata, and
+drift warnings flag any change in served model or system fingerprint during the run.
+
+**Dated snapshots are not mismatches.** Requesting `gpt-4o-mini` and being served
+`gpt-4o-mini-2024-07-18` is the provider resolving an alias. The snapshot is still stored
+in `served_model`. Any other difference produces a warning.
+
+**Judge verdict parsing is strict.** The first standalone `PASS`/`FAIL` token decides, and
+anything else counts as FAIL. A plain substring check would read "FAIL — does not PASS" as
+a pass.
+
+**Exit codes separate outcomes.** `run` exits 3 when any case is UNSTABLE, which is
+different from 1 (error) and 2 (usage). CI can then accept an expected unstable demo while
+still failing on a broken replay.
+
+## Known limits
+
+- **Small-N bootstrap CIs are approximate.** With N=5 binary verdicts the bootstrap
+  distribution is coarse (steps of 0.2) and tends to under-cover. The CI's seed is fixed,
+  so the interval itself is reproducible, but it is still only an approximation.
+- **Flip rate needs N ≥ 5 to mean much.** At N=3 a single dissent already reads as 33%.
+  Use a larger N on the cases you care about.
+- **Stability thresholds are conventions.** `BORDERLINE_MAX_FLIP = 0.20` is a named
+  constant chosen for auditability, not a statistically derived cutoff.
+- **The `diff` noise floor is conservative.** It uses the widest per-case CI half-width
+  across both runs. That rarely claims a false "REAL CHANGE", but it will miss small real
+  shifts in the aggregate. A paired test across cases would be more powerful.
+- **Float scores use a median-crossing pseudo-flip.** This gives a variance signal
+  without a threshold, but it is a heuristic. All built-in scorers are binary.
+- **Canonical hosts are an allowlist.** Any self-hosted or gateway endpoint gets a
+  NON-CANONICAL warning by design. The warning means "provenance unverified", not "wrong".

evalseal-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Venkata Manideep Patibandla
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

evalseal-0.1.0/PKG-INFO

@@ -0,0 +1,116 @@
+Metadata-Version: 2.5
+Name: evalseal
+Version: 0.1.0
+Summary: Reproducibility and provenance receipts for LLM evaluations.
+Project-URL: Homepage, https://github.com/patibandlavenkatamanideep/evalseal
+Project-URL: Issues, https://github.com/patibandlavenkatamanideep/evalseal/issues
+Author: Venkata Manideep Patibandla
+License: MIT
+License-File: LICENSE
+Keywords: evaluation,llm,llm-as-judge,provenance,reproducibility
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Software Development :: Testing
+Requires-Python: >=3.11
+Requires-Dist: httpx>=0.27
+Requires-Dist: pydantic>=2.6
+Requires-Dist: rich>=13.7
+Requires-Dist: typer>=0.12
+Provides-Extra: dev
+Requires-Dist: pytest-cov>=5.0; extra == 'dev'
+Requires-Dist: pytest>=8.0; extra == 'dev'
+Description-Content-Type: text/markdown
+
+# EvalSeal
+
+**Reproducibility receipts for LLM evals: run it N times, report the score with its noise, and seal what actually ran.**
+
+## The problem
+
+An eval score from a single run is one sample of a random process. Run the same eval again
+and borderline items quietly flip from PASS to FAIL, especially when an LLM judge grades
+them. Reports also name the model you *asked* for, not the one that *answered*. EvalSeal
+measures the flips, records the real provenance, and seals both into a tamper-evident ledger.
+
+## Quickstart
+
+```bash
+pip install evalseal
+
+# The recorded demo (examples + cassette) lives in the repo.
+git clone https://github.com/patibandlavenkatamanideep/evalseal && cd evalseal
+
+# Replays the committed cassette: no API key, no network.
+evalseal run \
+  --dataset examples/borderline_judge/dataset.jsonl \
+  --target-config examples/borderline_judge/target.json \
+  --scorer-config examples/borderline_judge/scorer.json \
+  --n 5
+```
+
+## Real flip rates
+
+Recorded 2026-09-15 and committed in `tests/cassettes/run.json`: `gemini-2.5-flash` as both
+the target and the judge, temperature left at the provider default, 20 arguable prompts,
+5 runs each. The quickstart above replays exactly this run.
+
+**Mean score: 0.92, but 5 of 20 cases did not get the same verdict every time.**
+
+| case | prompt | verdicts | mean | 95% CI | flip rate | stability |
+|---|---|---|---|---|---|---|
+| b01 | Is a hot dog a sandwich? | `FPPFP` | 0.60 | [0.20, 1.00] | 40% | UNSTABLE |
+| b10 | Blockchain for a child in exactly 20 words | `FPFPP` | 0.60 | [0.20, 1.00] | 40% | UNSTABLE |
+| b16 | "Do we only use 10% of our brains?" in a jokey tone | `PPFFP` | 0.60 | [0.20, 1.00] | 40% | UNSTABLE |
+| b05 | A borderline-polite refusal to a coworker | `PFPPP` | 0.80 | [0.40, 1.00] | 20% | BORDERLINE |
+| b19 | A technically accurate haiku about recursion | `PPFPP` | 0.80 | [0.40, 1.00] | 20% | BORDERLINE |
+| 15 others | | `PPPPP` | 1.00 | [1.00, 1.00] | 0% | STABLE |
+
+Treat the k-th repeat of every case as one ordinary single-run eval, and the five
+"single runs" of this identical eval scored **0.90, 0.95, 0.85, 0.90 and 1.00**. A single
+run can't tell you which of those numbers you got.
+
+The report also flagged `TEMPERATURE NOT SET` for both the target and the judge, which is
+the reason these borderline verdicts can come out differently from run to run.
+
+To re-record with your own key, copy `.env.example` to `.env`, add a free
+[Google AI Studio](https://aistudio.google.com/apikey) key, and run the quickstart with
+`EVALSEAL_RECORD=1`. If the free tier rate-limits you, run the same command again later;
+responses already recorded are kept.
+
+## Commands
+
+| command | what it does | exit code |
+|---|---|---|
+| `evalseal run` | Runs each case N times, analyzes variance, seals a record, writes `report.json` + `report.md`. | `0` all stable/borderline · `3` any case UNSTABLE · `1` error |
+| `evalseal verify` | Recomputes every hash in `.evalseal/ledger.jsonl` and checks the chain links. | `0` intact · `1` tampered or broken |
+| `evalseal diff A B` | Compares two ledger runs and says whether the mean moved beyond the noise floor. Use `--` for negative indices: `evalseal diff -- 0 -1`. | `0` |
+
+**Stability classes** are based on the flip rate, the share of a case's N verdicts that
+disagree with its majority: `STABLE` (0), `BORDERLINE` (≤ 20%), `UNSTABLE` (> 20%).
+
+**Provenance warnings** show up in the report when:
+- the served model differs from the requested one (for the target or the judge),
+- the endpoint isn't a canonical provider host,
+- temperature was left at the provider default,
+- the served model or system fingerprint changed partway through the run.
+
+## How it works
+
+The executor sends each prompt to the target N times and scores every response. An LLM
+judge is itself a target, so its own randomness is measured instead of assumed away.
+`analyze.py` computes the mean, a seeded bootstrap 95% CI, and the flip rate for each case.
+Every request goes through a cassette. In record mode, real responses are saved in call
+order; in replay mode, which is the default and what CI uses, they are served back, and a
+missing entry fails loudly. Each run is saved as a `RunRecord`: its manifest (requested vs.
+served model, fingerprint, parameters and whether they were set explicitly, rubric hash,
+dataset hash) plus its results. The record is hashed and linked to the previous record's
+hash in an append-only JSONL ledger, so editing any past score breaks `verify`.
+
+See [DESIGN.md](https://github.com/patibandlavenkatamanideep/evalseal/blob/main/DESIGN.md) for what this does and does not prove.

evalseal-0.1.0/README.md

@@ -0,0 +1,86 @@
+# EvalSeal
+
+**Reproducibility receipts for LLM evals: run it N times, report the score with its noise, and seal what actually ran.**
+
+## The problem
+
+An eval score from a single run is one sample of a random process. Run the same eval again
+and borderline items quietly flip from PASS to FAIL, especially when an LLM judge grades
+them. Reports also name the model you *asked* for, not the one that *answered*. EvalSeal
+measures the flips, records the real provenance, and seals both into a tamper-evident ledger.
+
+## Quickstart
+
+```bash
+pip install evalseal
+
+# The recorded demo (examples + cassette) lives in the repo.
+git clone https://github.com/patibandlavenkatamanideep/evalseal && cd evalseal
+
+# Replays the committed cassette: no API key, no network.
+evalseal run \
+  --dataset examples/borderline_judge/dataset.jsonl \
+  --target-config examples/borderline_judge/target.json \
+  --scorer-config examples/borderline_judge/scorer.json \
+  --n 5
+```
+
+## Real flip rates
+
+Recorded 2026-09-15 and committed in `tests/cassettes/run.json`: `gemini-2.5-flash` as both
+the target and the judge, temperature left at the provider default, 20 arguable prompts,
+5 runs each. The quickstart above replays exactly this run.
+
+**Mean score: 0.92, but 5 of 20 cases did not get the same verdict every time.**
+
+| case | prompt | verdicts | mean | 95% CI | flip rate | stability |
+|---|---|---|---|---|---|---|
+| b01 | Is a hot dog a sandwich? | `FPPFP` | 0.60 | [0.20, 1.00] | 40% | UNSTABLE |
+| b10 | Blockchain for a child in exactly 20 words | `FPFPP` | 0.60 | [0.20, 1.00] | 40% | UNSTABLE |
+| b16 | "Do we only use 10% of our brains?" in a jokey tone | `PPFFP` | 0.60 | [0.20, 1.00] | 40% | UNSTABLE |
+| b05 | A borderline-polite refusal to a coworker | `PFPPP` | 0.80 | [0.40, 1.00] | 20% | BORDERLINE |
+| b19 | A technically accurate haiku about recursion | `PPFPP` | 0.80 | [0.40, 1.00] | 20% | BORDERLINE |
+| 15 others | | `PPPPP` | 1.00 | [1.00, 1.00] | 0% | STABLE |
+
+Treat the k-th repeat of every case as one ordinary single-run eval, and the five
+"single runs" of this identical eval scored **0.90, 0.95, 0.85, 0.90 and 1.00**. A single
+run can't tell you which of those numbers you got.
+
+The report also flagged `TEMPERATURE NOT SET` for both the target and the judge, which is
+the reason these borderline verdicts can come out differently from run to run.
+
+To re-record with your own key, copy `.env.example` to `.env`, add a free
+[Google AI Studio](https://aistudio.google.com/apikey) key, and run the quickstart with
+`EVALSEAL_RECORD=1`. If the free tier rate-limits you, run the same command again later;
+responses already recorded are kept.
+
+## Commands
+
+| command | what it does | exit code |
+|---|---|---|
+| `evalseal run` | Runs each case N times, analyzes variance, seals a record, writes `report.json` + `report.md`. | `0` all stable/borderline · `3` any case UNSTABLE · `1` error |
+| `evalseal verify` | Recomputes every hash in `.evalseal/ledger.jsonl` and checks the chain links. | `0` intact · `1` tampered or broken |
+| `evalseal diff A B` | Compares two ledger runs and says whether the mean moved beyond the noise floor. Use `--` for negative indices: `evalseal diff -- 0 -1`. | `0` |
+
+**Stability classes** are based on the flip rate, the share of a case's N verdicts that
+disagree with its majority: `STABLE` (0), `BORDERLINE` (≤ 20%), `UNSTABLE` (> 20%).
+
+**Provenance warnings** show up in the report when:
+- the served model differs from the requested one (for the target or the judge),
+- the endpoint isn't a canonical provider host,
+- temperature was left at the provider default,
+- the served model or system fingerprint changed partway through the run.
+
+## How it works
+
+The executor sends each prompt to the target N times and scores every response. An LLM
+judge is itself a target, so its own randomness is measured instead of assumed away.
+`analyze.py` computes the mean, a seeded bootstrap 95% CI, and the flip rate for each case.
+Every request goes through a cassette. In record mode, real responses are saved in call
+order; in replay mode, which is the default and what CI uses, they are served back, and a
+missing entry fails loudly. Each run is saved as a `RunRecord`: its manifest (requested vs.
+served model, fingerprint, parameters and whether they were set explicitly, rubric hash,
+dataset hash) plus its results. The record is hashed and linked to the previous record's
+hash in an append-only JSONL ledger, so editing any past score breaks `verify`.
+
+See [DESIGN.md](https://github.com/patibandlavenkatamanideep/evalseal/blob/main/DESIGN.md) for what this does and does not prove.

evalseal-0.1.0/examples/borderline_judge/dataset.jsonl

@@ -0,0 +1,20 @@
+{"case_id": "b01", "prompt": "Is a hot dog a sandwich? Answer yes or no with one sentence.", "expected": null}
+{"case_id": "b02", "prompt": "Summarize the causes of WWI in exactly one sentence.", "expected": null}
+{"case_id": "b03", "prompt": "Give a 'mostly correct' explanation of how vaccines work in 2 sentences.", "expected": null}
+{"case_id": "b04", "prompt": "Answer ambiguously whether Pluto is a planet.", "expected": null}
+{"case_id": "b05", "prompt": "Write a borderline-polite refusal to a coworker's request.", "expected": null}
+{"case_id": "b06", "prompt": "Explain in one sentence why the sky is blue, without using the word 'scattering'.", "expected": null}
+{"case_id": "b07", "prompt": "Is a tomato a vegetable? Answer in one sentence.", "expected": null}
+{"case_id": "b08", "prompt": "Name the largest desert in the world in one word.", "expected": null}
+{"case_id": "b09", "prompt": "In one sentence, was Napoleon short?", "expected": null}
+{"case_id": "b10", "prompt": "Describe how a blockchain works to a child in exactly 20 words.", "expected": null}
+{"case_id": "b11", "prompt": "What is the capital of Australia? Include one fun fact, keep it under 15 words.", "expected": null}
+{"case_id": "b12", "prompt": "Summarize the plot of Hamlet in exactly two sentences.", "expected": null}
+{"case_id": "b13", "prompt": "Is zero an even number? Explain briefly, but hedge a little.", "expected": null}
+{"case_id": "b14", "prompt": "How many continents are there? One sentence.", "expected": null}
+{"case_id": "b15", "prompt": "Give a one-sentence definition of 'machine learning' that a statistician would accept.", "expected": null}
+{"case_id": "b16", "prompt": "Do humans only use 10% of their brains? Answer in a casual, jokey tone in one sentence.", "expected": null}
+{"case_id": "b17", "prompt": "Translate 'break a leg' into French, keeping the idiomatic meaning, in one line.", "expected": null}
+{"case_id": "b18", "prompt": "Is glass a liquid? Give a short answer with one caveat.", "expected": null}
+{"case_id": "b19", "prompt": "Write a haiku about recursion that is technically accurate.", "expected": null}
+{"case_id": "b20", "prompt": "In one sentence, who invented the light bulb?", "expected": null}

evalseal-0.1.0/examples/borderline_judge/scorer.json

@@ -0,0 +1,6 @@
+{
+  "type": "llm_judge",
+  "judge_model": "gemini-2.5-flash",
+  "base_url": "https://generativelanguage.googleapis.com/v1beta/openai",
+  "rubric": "You are a strict grader. PASS only if the response is factually correct, directly answers the prompt, and is well-formed. Otherwise FAIL."
+}

evalseal-0.1.0/examples/borderline_judge/target.json

@@ -0,0 +1 @@
+{ "model": "gemini-2.5-flash", "base_url": "https://generativelanguage.googleapis.com/v1beta/openai" }