eleanity 1.2.0

Same model. Same input. Find the first divergence.

Eleanity

Same model. Same input. Find the first causal divergence.

CLI-first tool for LLM runtime parity. It compares inference stacks on the same scenario and reports:

• status (PASS, PASS_WITH_TOLERANCE, PASS_WITH_LIMITED_COVERAGE, INCONCLUSIVE, DIVERGENT, ERROR)
• first divergent layer (when present)
• coverage of required layers and diagnosis confidence
• functional impact (separate from internal parity)
• a reproduction command

It does not score model quality (no MMLU). It checks whether two runtimes still share the same causal path: artifact → template → special tokens → token IDs → generation / API.

Product surfaces: CLI (text / json / quiet / sarif) and a Python API for embedding without subprocess — see docs/api.md.

Try the core idea offline

uvx eleanity demo

status:            DIVERGENT
first_divergence:  template
baseline_template: 'user: Hello\nassistant:'
candidate_template:'user: Hello'
probable_cause:    CHAT_TEMPLATE_DIFFERENT (confidence=0.92)

No model download, GPU, server, account, or network endpoint is required.

Hosting note: the repository currently lives at TheusHen/eleanity. Transfer to eleanity/eleanity is planned when the org is available (ROADMAP.md).

---

Install (clean machine)

From Git (available now)

# one-shot CLI without a permanent install
uvx --from git+https://github.com/TheusHen/eleanity.git eleanity --help

# or install into an environment
pip install "git+https://github.com/TheusHen/eleanity.git"
# optional HF backend (PEP 508 direct reference + extras)
pip install "eleanity[transformers] @ git+https://github.com/TheusHen/eleanity.git"

uv add git+https://github.com/TheusHen/eleanity.git
# or with transformers extra:
uv add "eleanity[transformers] @ git+https://github.com/TheusHen/eleanity.git"

From a local checkout

git clone https://github.com/TheusHen/eleanity.git
cd eleanity
uv sync --group dev
uv run eleanity --help

PyPI

pip install eleanity
# pin the stable API line:
pip install "eleanity==1.0.0"

When a reviewed version change reaches main, release.yml runs the complete quality suite, publishes that exact version to PyPI, then creates an immutable vX.Y.Z tag and GitHub Release with the wheel, sdist, and SHA256 checksums. The workflow is idempotent and uses the built-in GitHub token; it never writes version commits directly to the protected branch.

# Required repository secret (once):
gh secret set PYPI_API_TOKEN -R TheusHen/eleanity

The job verifies wheel/sdist metadata and writes SHA256SUMS.

Requires Python 3.11+.

---

Offline parity check

uv run eleanity compare --model demo --backends fake,fake \
  --format quiet --no-parallel --no-gates

Exact output from a live run:

status=PASS impact=NONE coverage=100.0 confidence=0.85 first_divergence=none gates=True run_id=16c3f05b-70f9-4c6a-9170-3e5942f91bd6

---

Real model self-parity (PASS)

Model: HuggingFaceTB/SmolLM2-135M-Instruct (CPU).

Commands

uv sync --group dev --extra transformers

uv run eleanity pull HuggingFaceTB/SmolLM2-135M-Instruct --tokenizer-only

uv run eleanity compare \
  --model HuggingFaceTB/SmolLM2-135M-Instruct \
  --backends transformers,transformers \
  --format quiet --no-parallel --no-gates \
  --observe artifact,template,special_tokens,tokens,generation

Exact quiet output (live run)

status=PASS impact=NONE coverage=100.0 confidence=0.85 first_divergence=none gates=True run_id=0aba046a-5846-45b2-94d9-4584bb4fe98a

Metric	Value
Engine total	~10276 ms
First observation	~9259 ms
Second (warm)	~1017 ms
Token match	31 / 31
Required coverage	100% (min 75%)

uv run eleanity report 0aba046a-5846-45b2-94d9-4584bb4fe98a --format text

Layer table excerpt:

Layer            Baseline obs   Candidate obs   Compare
artifact         OBSERVED       OBSERVED        PASS
special_tokens   OBSERVED       OBSERVED        PASS
template         OBSERVED       OBSERVED        PASS
tokens           OBSERVED       OBSERVED        PASS
generation       OBSERVED       OBSERVED        PASS

Template diff: PASS
Token diff:    PASS (count: 31)

This validates the engine + Transformers path.

---

Cross-runtime DIVERGENT (Transformers × real HF OpenAI-compat HTTP)

Both sides run real Hugging Face weights (SmolLM2-135M-Instruct):

Side	Runtime path
Baseline	transformers adapter (in-process)
Candidate	vllm adapter → local OpenAI-compat server that loads the same HF model via Transformers

The HTTP server does not expose template/tokenize endpoints (same honesty profile as many LM Studio / gateway setups). By default it applies the chat template without the assistant generation prompt — a common production bug — so generation diverges for a causal reason, not a hard-coded string.

Commands (exact)

uv sync --group dev --extra transformers

# starts real HF server, runs doctor + compare, stops server
uv run python scripts/examples/run_cross_runtime_demo.py

Under the hood:

# terminal A
uv run python scripts/examples/hf_openai_server.py \
  --port 8765 --preload --omit-generation-prompt \
  --model HuggingFaceTB/SmolLM2-135M-Instruct

# terminal B
export ELEANITY_VLLM_URL=http://127.0.0.1:8765
uv run eleanity doctor --check-backends --backends vllm --format json
uv run eleanity compare \
  --model HuggingFaceTB/SmolLM2-135M-Instruct \
  --backends transformers,vllm \
  --backend-url vllm=http://127.0.0.1:8765 \
  --policy quantized \
  --format quiet --no-parallel --no-gates \
  --observe artifact,template,tokens,generation,api

Parity attempt (server uses full AGP): ELEANITY_DEMO_MATCH=1 uv run python scripts/examples/run_cross_runtime_demo.py

Point the same commands at LM Studio / vLLM serve by changing --backend-url. Offline fixed-string stub (CI without GPU weights): scripts/examples/mock_openai_diverge.py.

Exact quiet output (live run)

status=DIVERGENT impact=HIGH coverage=50.0 confidence=0.762 first_divergence=generation gates=False run_id=90028893-8848-463f-9331-daf5268f60b5

Field	Value
Baseline	transformers · HuggingFaceTB/SmolLM2-135M-Instruct (HF weights, in-process)
Candidate	vllm adapter → hf_openai_server.py (same HF weights over HTTP, --omit-generation-prompt)
Policy	quantized
Status	DIVERGENT
First divergence	generation
Coverage	50% required layers (template/tokens not mutually observed on HTTP)
Verified	artifact, generation
Not verified	template (NOT_SUPPORTED on HTTP), tokens (NOT_EXPOSED on HTTP)
Generation texts	transformers: Hello! How can I help you today? · HTTP: assistant\nHello! How can I help you today?
Engine total	~8287 ms

Reproduction command stored on the run:

eleanity compare --model HuggingFaceTB/SmolLM2-135M-Instruct --backends transformers,vllm \
  --baseline transformers --policy quantized \
  --observe artifact,template,tokens,generation,api --no-gates \
  --backend-url vllm=http://127.0.0.1:8765 --format text

This is the cross-stack claim: two real runtimes, same model weights, localized first divergence, honest missing layers.

---

Template first-divergence (character-level)

Classic chat-template bug (missing assistant generation prompt), fully localized:

uv run python scripts/examples/demo_template_divergence.py

Exact output (live run):

status:            DIVERGENT
first_divergence:  template
character:         11
byte:              11
baseline template: 'user: Hello\nassistant:'
candidate template:'user: Hello'
probable_cause:    [CHAT_TEMPLATE_DIFFERENT] conf=0.92

More write-ups: docs/examples/first-divergence.md.

---

Compared to common alternatives

Approach	Result type	Causal first layer	CI exit contract	Missing-data honesty
Manual print / eyeball	vibes	no	no	no
String golden tests	brittle text diff	rare	sometimes	often silent
lm-eval / quality benches	capability scores	no	yes	n/a
API latency dashboards	ops metrics	no	maybe	n/a
Eleanity	parity + location	yes	0/1/2	yes (coverage)

---

Adapter capability matrix (summary)

Full table: docs/adapter-capabilities.md.

vllm and llamacpp are HTTP-client adapters. Run the corresponding server runtime in its own restricted environment; Eleanity does not install either server runtime as an extra.

Adapter	template	tokens	logits	generation	streaming	API
fake	yes	yes	no	yes	yes	partial
transformers	yes	yes	partial	yes	no	no
vllm HTTP	partial	partial	no	yes	partial	yes
llamacpp HTTP	partial	partial	no	yes	partial	yes
ollama	no	no	no	yes	partial	yes
sglang / tgi / openai	partial/no	partial/no	no	yes	partial	yes

---

Everyday CLI

eleanity doctor --check-backends
eleanity compare --backends transformers,vllm --format text
eleanity test fixtures/public/tokenizer-edge.yaml --backends fake,fake --format quiet
eleanity report <run-id> --format text
eleanity replay <run-id>
eleanity stabilize --backend fake --repetitions 3 --format quiet
eleanity policy-spec --policy quantized

Exit codes: 0 pass family · 1 divergent / gate fail · 2 config / dependency error.

Full reference: docs/cli.md

---

Python API (no subprocess)

from eleanity import Eleanity

client = Eleanity()  # or Eleanity.from_yaml("eleanity.yaml")
result = client.compare(model="demo", backends=["fake", "fake"], no_gates=True)
print(result.status, result.first_divergence, result.coverage)
raise SystemExit(result.exit_code)

Low-level observe/compare: from eleanity.api import observe_backend, compare_traces, make_scenario.
Full reference: docs/api.md.

---

CI

Workflow	Role
ci.yml	Required quality: actionlint + ruff + mypy + unit/contract/integration + CLI smoke
ci.yml typecheck job	Required mypy gate
parity-local-ai.yml	Downloads SmolLM2-135M and runs real Transformers self-parity
parity-public-fixtures.yml	Public fixture suites
release.yml	Idempotent PyPI publish, immutable tag, and GitHub Release for reviewed versions
eleanity.yml	Reusable monorepo gate

Local:

uv run ruff check src tests
uv run ruff format --check src tests
uv run pytest -q

---

Project docs

Doc	Purpose
docs/cli.md	CLI reference
docs/api.md	Python client + low-level API
docs/parity-specification.md	Status + comparator tables
docs/adapter-capabilities.md	Honesty matrix
docs/trace-specification.md	Trace Spec v1
ROADMAP.md	Stable roadmap and known boundaries
SUPPORT.md	Where to get help
SECURITY.md	Vulnerability reporting

---

License

Apache License 2.0 — full text in LICENSE.