Anim
Anim is local research software that checks whether an event-based model (EBM) result changes when you run predeclared, scientifically reasonable alternatives. You connect your EBM as a local worker. Anim does not import, select, download, certify, or silently replace your model.
It is not a diagnostic, prognostic, treatment, causal, regulatory, or medical-device tool. An emitted event order is not evidence that a recoverable disease-order signal exists.
Version: 0.2.0. See the changelog for this release. The 0.1.1 release artifacts and their recorded hashes remain immutable.
Installation
Anim requires CPython 3.12:
python3.12 -m pip install 'anim==0.2.0'
ebm-audit doctor
The compatibility matrix covers CPython 3.12 on macOS and Linux.
macOS worker execution requires Seatbelt; Linux requires a working Bubblewrap
installation at /usr/bin/bwrap with namespaces enabled. Without a provider,
worker execution fails closed with PRIVACY.CONTAINMENT_UNAVAILABLE.
Windows execution and other Python minors are unsupported.
To work from this source checkout instead:
python3.12 --version
uv sync --frozen
uv run ebm-audit doctor
See the packaging validation runbook for fresh wheel installation, enforced offline synthetic smoke, and the CI matrix.
Smallest Runnable Example
Run the project-owned synthetic demo before opening or connecting any participant data:
ebm-audit demo --conformance-ebm
It runs offline and uses only synthetic project-owned data. Its expected bounded
outcome is PARTIAL with an INCOMPLETE report at
ebm-audit-demo/report/. That result exercises the worker and report paths.
It is not scientific validation and does not accept a backend.
For a transferred offline kit, follow the offline kit guide. Its installed command can be run from the supplied virtual environment:
cd "$PROOF_ROOT" || exit 1
"$VENV_ROOT/bin/ebm-audit" demo --conformance-ebm
- Open
ebm-audit-demo/report/report.htmllocally and inspect the matchingebm-audit-demo/report/report.json,ebm-audit-demo/report/universes.csv, andebm-audit-demo/warnings.jsonlfiles. Inspectwarnings.jsonlfor visible cautions and diagnostics. - Read How the synthetic route works before changing a worker. It follows a deliberately tiny two-event synthetic example from a request shape to the visible audit artifacts.
- To start a researcher-owned worker project, run
ebm-audit adapter init /approved/local-config/my-ebm-worker, then follow the generatedREADME.md. The generated backend is aSYNTHETIC-ONLYtransport example, not an EBM. - If the model lives in a private Jupyter notebook, follow the frozen notebook handoff. The notebook and data remain private and local.
Worker integration and audit workflow
Version 0.2.0 adds local worker pinning and capability checks, saved-run summaries and comparisons, and fresh-attempt replay with progress and memory admission. Start with the synthetic demo above; inspect its saved evidence with:
ebm-audit summary --run-dir ebm-audit-demo
Use the adapter runbook for adapter pin,
adapter check, and the separately provisioned synthetic-only open-source EBM
example. Software dependencies and public source code are prepared explicitly;
audit runtime remains offline, with no telemetry or LLM interpretation.
For ordinary configured runs, reproduction and recovery
explains rerun and its refusal of identity drift. Replay recipes live beside
sealed results in <run-name>.operations/; the ephemeral demo has no recipe.
Execution controls describes cancellation, JSON progress on
stderr, and memory reservations that limit concurrency without dropping planned
candidates. Reservations are not measured RSS or an OS memory cap.
Report comparison explains summary and diff.
Missing, invalid, failed and incomparable evidence remains explicit. A successful
software operation does not make its scientific result complete or valid.
What The Audit Checks
For evidence that a worker can actually supply, the auditor keeps these questions separate:
- How much order or stage uncertainty exists within one fit.
- How much results change across independent chains or seeds.
- How sampling, declared analyst decisions, and participant removal change it.
- Whether synthetic no-signal controls show an apparent result when they should not.
- Whether a worker omitted a capability, failed, or returned invalid evidence.
An emitted order is not proof that a disease-order signal is recoverable. The auditor makes sensitivity and missing evidence visible; it does not diagnose, predict, recommend treatment, or establish a causal result.
Read Next
| Need | Read |
|---|---|
| See the end-to-end synthetic route and public Fit SDK | How the synthetic route works |
| Build a local worker around an EBM | Custom worker guide |
| Preserve a private Jupyter baseline | Frozen notebook handoff |
| Understand accepted input and privacy rules | Input-data dictionary |
| Understand the optional real-data handoff and current report limit | Optional downstream real-data integration |
| Verify a transferred offline kit instead of using this checkout | Offline kit verification |
| Read the exact worker wire contract | Worker protocol schema and canonical records schema |
| Read the readiness claim boundary | EBM integration readiness contract |
| Read the execution-boundary specification | Adapter protocol |
| Read report and claim wording rules | Reporting and claim language |
Do not copy participant rows, private column names, raw values, reversible mappings, or local paths into this repository, reports, tickets, chat, or a corpus note. A real-data integration is optional downstream work that needs its own local permission, privacy review, scientific review, and worker evidence.
The supported public integration surface is the ebm-audit CLI and the Python
package ebm_audit.worker_sdk. cli_workflows and reporting modules are
auditor internals, not alternate worker-integration APIs.
Read the result states
- Warnings are visible cautions or diagnostics, not automatic failure or permission to ignore a scientific gate.
UNSUPPORTED_CAPABILITYmeans the worker cannot perform a requested output. It is an explicit non-success universe, not missing evidence and not a pass.UNAVAILABLEin the training-stage status fields means required evidence cannot be supplied. It remains visible. It is neither pass nor fail.NOT_APPLICABLEin the training-stage status fields means evidence is outside the declared capability or analysis scope. It remains visible. It is neither pass nor fail.- Inspect
capability_evidence.training_stage.posterior.status,capability_evidence.training_stage.hard_stage.status, andcapability_evidence.training_stage.expected_stage.statusfor the declared training-stage capability state. - Failed universes remain visible with their terminal
final_statusincandidate_records[].final_statusandebm-audit-demo/report/universes.csv. Do not silently drop them or interpret them as successful scientific evidence.
What this PARTIAL/INCOMPLETE result can establish
A PARTIAL/INCOMPLETE audit supports only the visible protocol and
capability-limited synthetic evidence actually present in its report. It does
not establish a recoverable disease-order signal, scientific validity,
diagnosis, prognosis, treatment, or causal claims. Unavailable evidence remains
missing and is neither pass nor fail. Product readiness and worker integration do
not certify pysaebm, PySuStaIn, or any named or future EBM backend.
Status And Authority
The only product-readiness state is:
READY FOR RESEARCHERS TO INTEGRATE AN EBM AND RUN THE AUDITOR LOCALLY
The backend-neutral integration and local audit path has completed the project's synthetic readiness review. This is software readiness, not scientific approval of an EBM or dataset.
It means a researcher can connect a local EBM worker and run the auditor without the original developer's help. It does not accept a named backend, validate an untested integration, establish a disease-order signal, or authorize participant-data use. The exact claim boundary is the normative EBM integration readiness contract.
See the changelog for the public release scope.
Anim is licensed under the Apache License 2.0.
Metadata
Release files for anim 0.2.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| anim-0.2.0.tar.gz | 1.8 MB | With provenance | Details for anim-0.2.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| anim-0.2.0-py3-none-any.whl | Python 3 | none | any | Details for anim-0.2.0-py3-none-any.whl |
Total release size: 3.7 MB
Release files / anim-0.2.0.tar.gz
| Download URL | anim-0.2.0.tar.gz |
|---|---|
| Size | 1.8 MB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
d63bb98fb267089d068b94d5095f9711e6634258de007f146417cc8e35c85b3b
|
|
BLAKE2b-256 checksum How to use checksums |
9b33d3028f29ea8f0768358b43c8342a1e999a4e2bf72e9e423b584d33c0a3be
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 5, 2026.
Transparency logRelease files / anim-0.2.0-py3-none-any.whl
| Download URL | anim-0.2.0-py3-none-any.whl |
|---|---|
| Size | 2.0 MB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
4d727ad94ff7c98c61d8cb9a979037a2a29abcdb116998044a38f2413efb1772
|
|
BLAKE2b-256 checksum How to use checksums |
f45561e1f21a77cff7a03df5c4b515fbc42170f7a4bfdfd35433140f458f8112
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 5, 2026.
Transparency logSecurity for anim 0.2.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for anim-0.2.0.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 5, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.