NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #2762 most downloaded on PyPI
Local-first SDK for clinical extraction and de-identification workflows on hardware you control.
Latest stable
3.2.0
Last release 19 days ago
03 Nov 2026
Ships on a steady schedule
a new release about every 2 weeks
Nearly every release is documented
notes for 44 of 47 stable releases
Nothing withdrawn
no release was ever pulled
1 years old
53 releases · first in 2025
Drag across the ruler to compare, or tap a release to open it.
OpenMed 3.2.0 — global languages, synthetic clinical data and local models
OpenMed 3.2.0 expands what you can build locally with clinical text: more language-specific privacy handling, reproducible synthetic clinical datasets, and native GLiFormer, GLiNER2.5 and GLiClass model integrations. It also fixes user-reported problems with date shifting, document boundaries, HTML output and active-learning queues.
The focus is practical: preserve the original text and its offsets, make generated data traceable, and give applications explicit control over models, languages and decision policies.
This release adds localized medical-record labels, Nepali identifiers and dates, UK postcode detection, and African month-name dates. Native calendars, native digits, source-script names, local address conventions and honorifics receive more consistent handling during detection and de-identification.
New tooling covers signed offline language bundles, transactional activation, openmed i18n check, script-tokenizer audits, native-review packets, adjudication, morphology-aware label projection and translation guards. Language readiness and graduation gates keep independent recall, review and regression requirements visible.
The initial low-resource language cohort remains experimental. Added patterns, synthetic examples and user-supplied model routes do not establish clinical quality or make every language production-ready.
The synthetic APIs now cover clinical note grammars, longitudinal trajectories, medication courses, lab time series, prior-authorization packets, FHIR and OMOP records, HL7 v2 messages, C-CDA documents, code-switched dialogue and patient intake conversations.
Versioned dataset recipes bring these generators together with patient-aware split planning, source-bound truth, license and lineage checks, memorization and contamination audits, subgroup evidence, release capsules and withdrawal controls. Structured record schemas make exports easier to validate and reuse.
These are synthetic fixtures for augmentation and engineering evaluation. Constructed truth labels do not demonstrate clinical accuracy, anonymity or population equivalence. No restricted clinical records or terminology assets are bundled with this release.
Four newly public OpenMed models accompany v3.2.0, expanding local disease mention extraction in English and Spanish:
| Model | Language | Parameters |
|---|---|---|
| GLiFormer-BioNER-Disease-Base-v1 | English | 264M |
| GLiFormer-BioNER-Disease-Large-v1 | English | 576M |
| GLiNER2.5-BioNER-Disease-Base-v1 | English | 194M |
| GLiNER2.5-ClinicalNER-Disease-Spanish-v1 | Spanish | 287M |
All four are Apache-2.0 models that return disease mentions with character offsets and confidence scores, with CPU inference supported. Model weights remain a separate download; each linked model card includes usage examples, evaluation results and known limitations.
Use labels=["disease"] for the English models. For the Spanish model, use its trained label labels=["enfermedad o trastorno"], the model card's recommended threshold=0.4, and sentence_language="es" with analyze_text.
GLiFormer and current GLiNER2 / GLiNER2.5 checkpoints work through analyze_text, openmed.ner.infer and ModelLoader.create_pipeline, including cached and local checkpoint loading.
The GLiNER2.5 integration also exposes typed entity, classification, relation and JSON extraction helpers for compatible checkpoints. Those capabilities depend on the checkpoint; the disease models above are fine-tuned for disease mentions, and their other tasks have not been evaluated.
Optional GLiClass v4 torch and ONNX backends extend fixed-option decisions. Decision schema 1.1 adds named question sets, option descriptions, ordered levels and System One converters while retaining the 1.0 request contract. An opt-in local decision-server bridge and openmed decision-qualify support bounded serving and offline calibration. Model availability does not authorize autonomous clinical decisions.
dateutil (#4371).pyasn1>=0.6.5 (#4373).Read the 3.1-to-3.2 migration guide before upgrading. Two changes need particular attention:
font_fallback=True path.Existing public import paths are retained, including openmed.service.schemas.DECISION_SCHEMA_VERSION. Pin surrogate providers and review seeded-corpus drift when upgrading generated datasets.
The Python feature additions do not imply feature parity across every SDK. Experimental language status, checkpoint licenses and deployment qualification requirements remain explicit.
python -m pip install --upgrade "openmed==3.2.0"
# Optional local GLiNER2 / GLiNER2.5 support
python -m pip install "openmed[gliner2]==3.2.0"
# Optional local GLiFormer support
python -m pip install "openmed[gliformer]==3.2.0"from openmed import analyze_text
result = analyze_text(
"Synthetic note: the patient has asthma and chronic kidney disease.",
model_id="OpenMed/GLiNER2.5-BioNER-Disease-Base-v1",
revision="fccb29669b847cd6ca6384660d06c750c55366e2",
threshold=0.7,
)
for entity in result.entities:
print(entity.text, entity.label, entity.start, entity.end)The first use can download the model; subsequent use can run from the local cache. The threshold above is an example, not a clinical operating point or a change to the SDK default.
Thank you to everyone who contributed code, reported bugs and supplied concrete reproductions for this release.
Full Changelog: v3.1.0...v3.2.0
Your coding agent can read these notes before it upgrades. Set up the MCP server →
Global languages, synthetic clinical data and local decision-model backends. This release candidate was prepared on 2026-10-30 from v3.1.0 and retains the model catalog, artifact pins and registry pointer targets. Publication and the final release date remain pending. See the release notes and migration guide.
Burned-in synthetic rendering now requires covering fonts and complex-script shaping; missing default fonts require explicit font_fallback=True (#3868).
Dataset-recipe callers can opt into checked image fallback through
builtin_recipe_generators(font_fallback=True); the policy is provenance-bound.
Require Python 3.11 or newer for the v3.2 development line so dependency security fixes can be installed. Upgrade Python 3.10 environments before updating OpenMed; published v3.1.0 remains available for the older interpreter. See the dependency migration guidance.
Document existing synthetic generators, offline examples, public exports and dataset provenance limits in a shared navigation group (#3896).
Document local decision-model support, native backend limits, compatible server requirements, and the pinned MedDecider NLI qualification recipe.
Add offline decision-backend qualification with development-only threshold fitting, held-out evaluation, value-free reports and explicit calibration profile output.
Generate structured multilingual synthetic dialogue with grapheme-safe spans, Journey fact evidence and explicit adjudication truth (#2846).
Validate synthetic releases against disclosure, lineage, licenses, withdrawals, benchmark contamination and subgroup utility evidence (#3214).
Add language-family adapter distillation plans, per-language teacher loss helpers, and non-compensable edge promotion gates.
Add deterministic dialect-distance policies for explicit fallback, ensemble routing, and abstention with value-free result metadata.
Define the initial experimental low-resource language cohort and non-compensable recall, review, packaging and regression graduation gates. The live inventory includes explicit augmentation-only generators independently of model-backed language support.
Lock the language readiness Markdown contract to synthetic golden snapshots and escape version code spans safely (#3122).
Build minimal native-language reviewer packets with opaque item IDs, separate provenance manifests and closed-code decision imports (#2864).
Merge native-review labels through explicit policy-bound adjudication, invalidating stale projections and ambiguous supersession chains (#3696).
Add configurable morphology-aware clinical label projection with exact code-point, grapheme and byte offsets, and explicit ambiguity review.
Add an offline translation guard for protected terminology, numeric values, placeholders and span maps, with explicit surface-check limits.
Add explicitly marked augmentation-only synthetic notes for six national-ID-only or user-model language routes, with validator-backed identifiers and value-free failure diagnostics.
Add deterministic rare-case synthetic privacy stress records with configurable attribute tiers, value-free risk labels, and linkage-attack truth.
Add deterministic synthetic patient-intake conversations with causal turn-level PHI, assertion, revision, consent, and provenance labels.
Expose typed GLiNER2.5 entities, classification, relations and JSON extraction, six checkpoint capability records and a local fixed-option decision backend with opt-in offline checkpoint checks.
Add a pinned local GLiNER2.5 language-decision benchmark with synthetic fixture digests and value-free comparisons against deterministic routing.
Detect and convert suspected Zawgyi Burmese before routing, preserving source offsets and abstaining on mixed encoding evidence (#3889).
Extend fixed-option decisions with named question sets, per-question definitions, ordered levels and System One converters; preserve the 1.0 contract, reject unsupported question context, and keep malformed Unicode out of exception payloads.
Add an opt-in System One and Decisions server backend with loopback defaults, explicit non-PHI private-network approval, bounded responses and deadlines, and typed transport failures.
Extend synthetic golden date and identifier traps across fourteen low-resource locales, with exact spans, hard negatives, deterministic masks, and explicit coverage inventories.
Add declarative localized medical-record-number labels shared by semantic merging and safety sweeps, retaining exact identifier spans and label-free negative controls.
Detect UK postcodes through the English en_GB overlay, including bounded address context for lookalike inward codes and explicit UK detection aliases.
Add an experimental Sinhala route for user-supplied models, structural Sri Lanka NIC handling and synthetic script-aware surrogate approximations without adding bundled model support.
Recognize short Thai address lines without consuming adjacent native-script date spans.
Propagate context-gated national-ID validator rejections through semantic merging and safety sweeps while preserving unrelated and validated spans.
Keep six-digit Bengali PIN codes detectable near dotted date fragments by requiring a four-digit date year.
Add a GLiClass v4 fixed-option decision backend over optional torch and ONNX runtimes, retaining calibrated and initial scores without request text and rejecting malformed backend settings with value-free errors.
Record surrogate-provider versions and digest-only probe receipts, and expose optional strict drift checks for the seeded synthetic corpus (#3893).
Package and install signed offline language bundles with explicit resource, license, checksum and synthetic self-test gates (#2866).
Activate verified local language bundles with qualified task records, atomic registry generations and explicit rollback (#3695).
Generate seeded Devanagari and Telugu name surrogates from disjoint, well-formed aksara while preserving token shape (#3892).
Generate local-convention African address and town surrogates, with mask fallback when source-disjoint vocabulary is exhausted (#3891).
Add an opt-in live surrogate quality gate with deterministic per-language and per-label checks, hash-only reports, and explicit documented approximation exemptions.
Generate opt-in native-digit and native-calendar synthetic dates while respecting historical month bounds (#3886).
Render generated identifiers, phones and shifted dates in the source digit system and grouping (#3871).
Detect native-calendar month-name dates with exact spans and value-free calendar tags (#3866).
Shift Solar Hijri, Ethiopian and Thai Buddhist-era dates within bounded native-calendar ranges (#3867).
Shift declared Hijri and Bikram Sambat dates through bounded, digest-verified permissive calendar tables (#3888).
Add synthetic fhir records with bounded offline contracts, synthetic controls and documented behavior (#2840).
Add synthetic omop cohorts with bounded offline contracts, synthetic controls and documented behavior (#2841).
Add synthetic lab time series with bounded offline contracts, synthetic controls and documented behavior (#2843).
Add synthetic medication courses with bounded offline contracts, synthetic controls and documented behavior (#2844).
Add synthetic prior-authorization packets with bounded offline contracts, synthetic controls and documented behavior (#2842).
Add synthetic clinical note grammars with bounded offline contracts, synthetic controls and documented behavior (#2838).
Add synthetic longitudinal trajectories with bounded offline contracts, synthetic controls and documented behavior (#2839).
Compose existing synthetic generators through a versioned transactional local dataset CLI (#3693).
Add deterministic synthetic dataset partitions that preserve shared trajectory, derivation lineage and transitive ancestry, with explicit unmet-stratum export refusal (#3692).
Report independent synthetic fidelity axes, Wilson intervals and aggregate reference distribution comparisons without retaining reference records (#2852).
Bind synthetic image/text/label pairs to permissive declarations and reproducible transform chains (#2851).
Validate synthetic artifact facts, times, labels and coordinates against shared Journey truth (#3694).
Gate multilingual corpus shard uses on declared license, consent, translation lineage and restricted user-supplied custody (#2869).
Package controlled synthetic recipes and evidence into canonical offline-verifiable release capsules (#2854).
Add clinical transliteration provenance with bounded offline contracts, synthetic controls and documented behavior (#2857).
Add local abbreviation mining with bounded offline contracts, synthetic controls and documented behavior (#2858).
Add code-switch span projection with bounded offline contracts, synthetic controls and documented behavior (#2860).
Add deterministic metadata-only native-review selection with scarcity, disagreement, diversity ties, and simultaneous language, label and total budget caps (#2863).
Add independent-source cross-lingual pseudo-label consensus with explicit adjudication for conflicting or overlapping projections (#2862).
Added a bounded, value-free synthetic memorization audit that fingerprints protected reference material, blocks exact, paraphrase-style and exposure matches on configurable signals, and renders canonical digest-only JSON evidence with offline golden tests (#2853).
Added a strict, fail-closed synthetic dataset license and lineage manifest validator that binds source class, SPDX license expression, generator digest, model provenance, bounded dependencies and the redistribution decision, and renders canonical reason-code-only reports (#3099).
Add language coverage diffs with bounded offline contracts, synthetic controls and documented behavior (#3101).
Add clinical language readiness with bounded offline contracts, synthetic controls and documented behavior (#2855).
Audit synthetic script tokenization fragmentation, unknown-token rates, grapheme splits and offset recovery against pinned local limits (#2856).
Stratify language uncertainty and minimum-detectable-effect planning without ranking underpowered slices (#2865).
Gate paired multilingual checkpoint changes against every declared last-green language/label cell (#2870).
Added versioned, offline JSON Schemas for the synthetic training records the
package already emits (locale_phi_example, burned_in_annotation,
social_history_example, section_label_record,
translation_augmented_example), with record_schema_id, schema and
fingerprint export helpers, deterministic file writing and a validate_record
check that pairs schema validation with offset invariants and rejects records
claiming real PHI (#3895).
Generate seeded synthetic HL7 v2 ADT, ORU and ORM messages with a field-level
PHI truth manifest addressed by segment, field, repetition, component,
subcomponent and raw-text offsets across encoding-character variants, escape
sequences, repetitions, timestamp precisions, free text and Z-segments. A
value-free leakage harness reports residual PHI after redact_hl7v2() by
field position and label (#3884).
Generate seeded synthetic C-CDA documents (CCD, discharge summary, progress
note) with element-level PHI truth, narrative-to-structured links and
placeholder codes, plus a value-free leakage harness over redact_cda()
that reports residuals by element path and label and rejects remote schema
or entity-bearing options (#3885).
Added deterministic clinical language-pack readiness matrices with closed
capability and state vocabularies, sha256: evidence digests, alias-normalized
locale folding, byte-stable JSON/Markdown rendering and registry-derived
entries, plus offline golden tests (#3100).
Added the offline openmed i18n check locale-pack conformance command
with one independent result per component, a closed set of stable reason
codes, deterministic text/JSON output and offline golden tests (#3103).
Add deterministic Nepali (ne) patterns for +977 mobile and landline
numbers, context-gated citizenship-certificate and national identity card
numbers without checksum claims, numeric Bikram Sambat dates and postal
codes, in Devanagari and ASCII digits. Add the ne number-separator entry
and a synthetic ne.jsonl golden fixture. Nepali remains a user-supplied
model route, and Hindi and Marathi outputs are unchanged (#3864).
Detect Hausa, Yoruba, Igbo and Kinyarwanda month-name dates in day-month-year and month-day-year order from CLDR month vocabularies and common spellings. Yoruba and Igbo match NFC, NFD and unmarked forms on grapheme-safe spans; a month word without a day and year stays unflagged. Adds synthetic month-name fixtures, including a new Kinyarwanda golden fixture (#3865).
Support native GLiFormer and current GLiNER2 / GLiNER2.5 NER through
analyze_text, ModelLoader.create_pipeline, and openmed.ner.infer, with
optional gliformer / gliner2 extras, index-free checkpoint selection,
device-isolated caches, offline loading, exact spans, and disease-label
defaults for OpenMed/GLiFormer-BioNER-Disease-Base-v1 (#4329).
Shift Chinese, Japanese and Korean marker-based dates without dateutil, retaining numeric field widths, whitespace, two-digit years and source digit systems (#4328).
Keep HTML entity highlights aligned with original source offsets across escaped characters, preserving source text and metadata escaping (#4326). Crossing, nested and adjacent spans retain each label's exact source coverage.
Make security-support guidance version independent and distinguish published stable releases from candidates (#4131, #4133).
Require banking context for pattern-only SWIFT/BIC matches so clinical section headers remain intact; retain contextual codes and model detections (#4152, #4230).
Preserve every splitlines separator, including form-feed PDF page breaks, in plaintext geometry and redaction boundaries (#3567, #4228).
Keep active-learning queues retryable when event persistence fails, applying queued and labeled memory updates only after a successful append (#4203, #4220).
Preserve teacher identities on typed weak-label spans, keeping typed and mapping consensus consistent without mutating caller-owned spans (#4202, #4219).
Register OpenMed/GLiNER2.5-BioNER-Disease-Base-v1 with its disease label
default and recognize saved GLiNER2.5 boundary-extractor configurations for
offline local loading.
Retain the openmed.service.schemas.DECISION_SCHEMA_VERSION compatibility
export while adding the decision 1.1 question-set contract.
Refresh the wheel-size baseline from verified current-master artifacts after accepted library growth, retaining 10% headroom and independent CI measurement.
Refresh the exact service-image timezone-data pin to the available Debian Forky release so image builds, SBOM generation and vulnerability scans can complete.
Preserve registered non-CJK honorifics and grammatical cues in person surrogates, and retain Vietnamese family-first order (#3887).
Keep Arabic (ar) and Amharic (am) name, city, and street surrogates in
the source script. Arabic-script spans use Faker's ar_AA names and curated
synthetic Egyptian place templates unless a regional locale ships
Arabic-script data; Ethiopic spans recombine the curated am values, and
LocalePhiGenerator emits native-script names and addresses for both
(#3863).
Load GLiNER2 checkpoints with the gliner2[local] package's architecture-aware
extractor and retain native span offsets and confidence scores instead of
attempting to use older gliner classes (#4329).
Status: release candidate; publication pending. The target stable version
is 3.2.0. The package version is prepared on the release branch without an
rc suffix; this page does not claim a published tag or registry artifact.
The baseline is v3.1.0. The release scope is global languages, synthetic data and local decision backends.
openmed i18n check.The initial low-resource cohort remains experimental, with unfilled native review roles. Added deterministic patterns, synthetic fixtures or model-loader support do not graduate a language. See the cohort and graduation contract and language readiness.
The synthetic data APIs cover composable note grammars, longitudinal trajectories, medication courses, constrained lab series, prior-authorization packets, FHIR and OMOP records, HL7 v2, C-CDA, code-switched dialogues and patient-intake conversations. Locale generators also include explicitly marked augmentation-only routes and native-calendar and digit variants.
Dataset recipes bind patient and generator lineage, split planning, licenses, cross-artifact truth alignment, memorization and contamination audits, release capsules, subgroup evidence and withdrawal controls. Burned-in rendering now requires appropriate glyph coverage and shaping, with explicit checked font fallback when requested.
Synthetic outputs and their constructed truth labels are augmentation and technical test data. They do not establish clinical accuracy, anonymity or population equivalence. No restricted clinical records or terminology assets are added by this release.
OpenMed/GLiNER2.5-BioNER-Disease-Base-v1 has an English disease-only label
default and supports both its Hub id and saved boundary-extractor snapshots.openmed decision-qualify calibration tooling.See fixed-option decisions and local decision models. Decoder models use the documented local-server bridge; a new in-process MedDecider runtime is not included. Backend availability is distinct from qualification for a deployment.
The optional clinical Safe Harbor profile and global adjacent-span merging proposals remain separate follow-ups to #4152. These corrections do not change the default policy profile or claim the reporter's benchmark scores as release qualification.
Read the 3.1-to-3.2 migration guide. Python 3.11 or
newer is required. Existing public import paths are retained, including
openmed.service.schemas.DECISION_SCHEMA_VERSION.
The optional cloud extra requires pyasn1 0.6.5 or newer for the upstream
ASN.1 decoder security fixes. The core dependency set is unchanged.
After publication, the intended coordinates are:
| Surface | Coordinate |
|---|---|
| Python | pip install "openmed==3.2.0" |
| Python model extras | pip install "openmed[gliformer,gliner2]==3.2.0" |
| JavaScript | npm install openmed@3.2.0 |
| Swift | .package(url: "https://github.com/maziyarpanahi/openmed.git", from: "3.2.0") |
| Android / JitPack | com.github.maziyarpanahi:openmed:v3.2.0 |
| Service image | ghcr.io/maziyarpanahi/openmed:3.2.0 |
The new Python language, generator and decision APIs do not imply native feature parity in the other SDKs. Existing platform contracts and documented experimental boundaries remain in force.
The frozen target branch is release/v3.2.0, cut from master
4a3b5899d8f64aef8cffd3178f566750a0a80a72. The preparation PR targets that
branch. While that PR is open, reviewed fixes can target its prepare/v3.2.0
head so its checks validate the combined candidate. After the preparation PR
merges, further v3.2 fixes target release/v3.2.0. Select only needed fixes
without pulling later feature work from master into the candidate. Keep the
release branch so the release can be tagged from its reviewed final commit.
The PR checks are the live source for candidate validation. The SDK readiness
workflow uses v3.1.0, this release's migration guide, the API compatibility
report, executed golden tests and retained model evidence. The model catalog,
artifact pins, baseline metrics and registry pointers are unchanged by release
preparation. New model training and model-promotion benchmarks are therefore
not SDK-release prerequisites.
Before publication, require a READY SDK report and successful applicable
Python, dependency/security, package, docs/browser, Swift, Android, Helm and
container checks on the final candidate commit. Set the actual release date
in the changelog, repeat the version/tag-availability preflight and obtain the
maintainer's publication instruction. Publish the immutable v3.2.0 tag from
that commit through the existing tag-driven workflows; never move an existing
tag. Record registry and consumer-install verification after publication.
Federated/private learning and new multimodal/ambient work remain with their later release plans. Clinical deployment and model or language qualification remain separate decisions.
Core SDK support for Python 3.10 remains, but vulnerable legacy Pillow is no longer retained as a fallback. Existing Python 3.10 image deployments mus…
OpenMed 3.1 brings explicit authority, human review, bounded execution and recovery to clinical workflows. It expands evidence-backed clinical briefs, adds governed FHIR and SQLite OMOP execution paths, and strengthens privacy across multilingual text, documents, medical imaging and service boundaries.
The release preserves local-first defaults and keeps clinical values inside protected application boundaries. Review and audit records use controlled outcomes, counts, offsets, opaque references and digests. Applications must explicitly supply trusted authorization, custody and execution integrations.
Signed grants and purpose tickets constrain who may access which data for a specific action. Guarded dispatch checks minimum-data projections, pinned tool identities and single-use approvals before effects. Runtime revocation invalidates affected grants and approvals, while effect admission is disabled by default and supports an emergency stop.
Approval policies support multiple reviewers, distinct roles and requester exclusions. Durable local stores preserve partial approvals, review leases and replay protection across restarts. Changes to an action or its evidence require fresh review.
Bounded scheduling adds explicit read concurrency, exclusive effects, resource admission and cooperative cancellation. Recovery checkpoints and signed action ledgers retain evidence of completed work and uncertain outcomes. Ambiguous external effects require reconciliation before further execution.
Python and TypeScript workflow clients, CLI commands, authenticated REST endpoints and MCP tools expose planning, inspection, review requests and cancellation through injected application services. MCP review tools carry metadata only and do not execute clinical effects. OpenMedKit adds matching review and evidence parsers with shared conformance fixtures.
Clinical briefs can use explicitly reviewed local evidence with source, fact, profile and calibration bindings. Composition preserves citations and evidence commitments, checks the rendered packet for privacy leakage, and returns controlled refusals when evidence, budgets or local models are unavailable.
Opt-in local natural language inference (NLI) verifies summarization and grounding claims while retaining contradiction and abstention outcomes. Verification metadata excludes source and claim text. Qualification helpers separate development and held-out evaluation and bind results to the supplied local artifact.
Additional evidence tools cover Journey-to-chart abstraction, source-support scoring, SDOH extraction, clinical relations and timelines. Language-aware SDOH processing reports unsupported, skipped and failed categories explicitly; packaged cue tables and built-in extractors declare English support.
Sealed evaluation sessions combine holdout verification, overlap checks, blinded review, budgeted feedback and durable checkpoints without retaining source content. Five runnable offline synthetic workflow examples demonstrate the review and refusal contracts. Bundled JSON Schemas cover governance, clinical evidence and private-training records.
$validate preflight blocks errors and unavailable validation before application review.Existing FHIR/OpenMRS clients and legacy OMOP writers do not automatically acquire these protections. Follow the migration guide to compose the complete application-owned authorization, review and execution boundary.
This minor release includes a narrowly approved security exception for ApprovalReceipt and AggregateDPRelease. Their constructor and serialized-record changes are intentionally breaking because restoring the old behavior would restore the privacy exposure. The compatibility gate permits only the ten documented field-removal paths, including re-exports, against v3.0.0 for v3.1.0. Other compatibility requirements remain in force, including for v3.2.
ApprovalReceipt no longer exposes reviewer_role, consumed_at or expires_at. Trusted execution code obtains those attributes through ApprovalTokenVerifier.consume_authorization(...) and keeps the resulting authorization inside the protected boundary. Serialize its minimal receipt for audit; a receipt is not an execution credential.AggregateDPRelease schema v2 removes noise and seed_digest. Consumers use the released value, privacy-budget accounting and test_only flag. Existing records containing both the released value and exact noise reveal the original aggregate and must not be treated as private releases.--test-mode with --seed; production releases must omit the seed to use system randomness.Read the complete 3.0-to-3.1 migration guide before upgrading affected integrations.
Pillow-backed image processing and brand rendering now require Python 3.11 or newer with Pillow 13.0.0. The agents, beam, llamaindex, mlx, multimodal, and ocr-paddle integration extras publish this secure minimum. Core SDK support for Python 3.10 remains, but vulnerable legacy Pillow is no longer retained as a fallback. Existing Python 3.10 image deployments must upgrade Python and reinstall their image extras before processing untrusted input. The frozen dependency graph also requires Certifi 2026.10.19 or newer. Beam currently declares upper bounds below the patched dependencies, so ordinary pip cannot resolve the secure Beam extra. Use the validated Beam installation recipe, which pins the tested versions and applies the three explicit dependency overrides.
Optional integration requirements now include patched minimum versions of aiohttp 3.14.4, banks 2.5.1, Mako 1.4.2, multidict 6.9.1 and tinytag 2.3.0. MCP requires version 1.30 or newer within the existing 1.x range, and optional service/development profiles require Strawberry GraphQL 0.327.2 or newer. These optional security upgrades do not expand the base SDK dependency set.
The reviewed exception for NLTK CVE-2026-81726 expires on 3 November 2026. At candidate validation, no patched upstream version was available. NLTK is present only in selected optional integrations and is absent from the core dependency closure and service image. The exception is temporary risk acceptance, rejects a reported fix or a different package/finding/target, and stops applying after expiry. Untrusted workflows must not control affected model-artifact paths.
The scan also retains an unpatched optional accelerate MEDIUM-severity finding below the unchanged HIGH-severity gate. Passing release checks does not mean every optional dependency is free of known findings.
After publication, install or upgrade the Python package with Python 3.10 or newer:
pip install --upgrade "openmed==3.1.0"For the browser SDK:
npm install openmed@3.1.0Swift package dependencies use version 3.1.0 from https://github.com/maziyarpanahi/openmed.git; Android uses com.github.maziyarpanahi:openmed:v3.1.0. The intended service image is ghcr.io/maziyarpanahi/openmed:v3.1.0. Publication and registry availability are separate from candidate validation.
Python, browser, Swift, Android, OpenAPI and deployment version surfaces are aligned. Shared version numbers do not imply feature parity across platforms, and optional integrations retain their own runtime and artifact requirements.
PR #4237 merged into master as 2e4c470a9f62f9bb17534923ab30463c87e8fef6 after 49 hosted checks passed on repair commit 716d31ffa76952ab223b18be334445a053b35c53. All 13 applicable workflows succeeded. The three expected skips were Pages deployment and container publication on pull requests, plus the unchanged CJK/Indic paths. The merge tree matches the tested candidate.
twine check, and strict documentation staging also pass. Homepage analytics and clinical-exclusion checks remain enforced.The earlier release-preparation PR #4236 validated the additional Swift, Android and client surfaces. This repair does not change those sources. Validation and merge do not establish registry publication or deployment availability.
This release retains the existing model artifacts, pointer targets and benchmark evidence. Synthetic tests establish software contracts; they do not establish clinical efficacy, clinician agreement or production deployment qualification. OpenMed does not make autonomous clinical decisions.
Further reading: Full changelog · Migration guide · Release preparation PR #4236 · Release validation repair PR #4237
Note truncated.
Governed clinical workflows, reviewed clinical evidence, interoperability and privacy hardening. The baseline is v3.0.0. This SDK candidate retains existing model artifacts and pointer targets; synthetic validation does not establish clinical efficacy. Package publication remains a separate release action.
Keep v3.1.0 with the maintainer-approved security exception for
ApprovalReceipt and AggregateDPRelease. Preserve the existing privacy
fixes; require migration for their constructor and public-record changes.
The readiness gate admits exactly ten reviewed field removals against v3.0.0,
retains them in signed evidence, and rejects unrelated breaks or reuse of
this exception in another minor or patch release. See
3.0-to-3.1.
AggregateDPRelease now emits release schema version 2 and removes the
noise and seed_digest fields from its public dataclass and serialized
records. Update constructors and consumers to use the noisy value, budget
accounting, and new test_only flag instead; ledger schemas and digests are
unchanged. Previously written records containing both value and noise
disclose their input aggregates and must not be treated as private releases.
Seeded releases are synthetic-test-only and provide no production privacy
guarantee; the CLI now requires --test-mode with --seed. Omit the seed for
production system randomness (#3723).
Gaussian helpers now use analytic privacy-profile calibration, including
epsilon above one, and record calibration_method in spend records. Noise
scales, utility estimates, and seeded Gaussian outputs intentionally change;
unrepresentable calibration fails closed. Update deterministic expectations
without restoring the invalid classical scale (#3724).
Add five runnable, offline synthetic clinical workflow evidence examples, with deterministic digest/ID/code-only reports and fail-closed controls (#3794).
Export deterministic, self-contained private-training JSON Schemas with stable IDs, bounded metadata, privacy-safe metric conditions and offline validation guidance (#3054).
Detect GIF, WebP and BMP media types in bounded multimodal preflight (#3833).
Add an offline synthetic NLI/guarded-relation example with explicit development and calibration-double backends, evidence bindings and human-review priorities; smoke checks guard projections and API drift (#3761).
Opt-in local NLI verification for summarization and grounding, with retained contradiction and abstention outcomes, offset- and digest-bound audit records, typed backend refusals and unchanged default serialization. Source and claim text are omitted from verification metadata; no trained model is required for explicit synthetic heuristic checks (#3987).
Report exact package/version exposure for the core install and each optional extra in the vulnerability gate, with offline uv exports, value-free scanner diagnostics and a static import guard for reviewed extra-only packages. Exposure informs review without changing severity thresholds or CVE waivers (#3758).
Added a bounded, offline pure-Python Circe cohort-expression mapper with digest-bound exchange custody, explicit semantic loss reports, synthetic membership round trips, and no SQL or bundled vocabulary (#3804).
Added fail-closed Python evaluation sessions composing sealed manifest and holdout verification, injected execution/scoring, overlap forensics, blinded review and budgeted feedback, with content-free durable checkpoints and at-most-once feedback claims (#3777).
Added action-class approval quorum policies over live consumed authorizations, distinct-role and requester exclusions, value-free decisions, and durable local partial approvals with expiry, replay protection and action-change resets (#3781).
Added bounded identifier-only agent scheduling with explicit read concurrency, exclusive effects, resource admission, cooperative cancellation and conservative restart checkpoints through an injected guarded dispatcher (#3672).
Bound five governance JSON readers before decoding and return their existing typed malformed-input errors for excessive nesting, oversized input and invalid JSON. Cover every public governance JSON reader with bounded property tests and verify consent rejection through the MCP error envelope (#3779).
Added offline chart-abstraction scoring with synthetic gold fields, separate abstention errors, fact-bound source-support counts, field-type slices and deterministic value-free Wilson intervals (#3788).
Added local Journey-to-chart-abstraction evidence production with explicit source-origin declarations, digest-bound review receipts, and persistent, value-free coverage blockers for missing, unsupported or conflicting evidence (#3778).
Add an offline synthetic SDOH/temporal/timeline example with explicit reference time, authored event axes and controlled labels/offsets/states-only output; smoke tests reject API drift, external arguments and source/date output (#3760).
Registered SDOH extraction scoring with independently authored synthetic gold, a reproducible counts-only baseline, retained known misses and an eval-only native SHAC status view that reports unavailable data without zero scores.
Add explicit language-aware SDOH extraction with per-category unsupported, processed, skipped and failed outcomes. Declare English support for packaged cue tables and built-in extractors, retain legacy English dispatch, and keep audit output limited to category metadata, counts and offsets (#3797).
Stage NLP-derived OMOP loader rows as ordered approval-bound batches with exact span lineage, vocabulary snapshot checks and rollback evidence. Replacement proposals preserve prior pipeline digests and explicitly review tombstones before inserts; row values remain in protected local memory.
Add an explicit local SQLite OMOP batch committer with row-state-bound previews, mandatory injected approval/admission/custody checks, atomic idempotency receipts and read-only recovery of unknown commit outcomes.
Export aligned SDOH evidence as passive R4 Observations and offset-linked Provenance, with caller-supplied per-record terminology, pinned HL7 categories, preliminary review states, explicit losses and preserved sensitive-use labels.
Add offline clinical sdoh, clinical relations and clinical timeline
commands over validated analyzed spans. Emit value-free review records,
require explicit temporal references and links, and reserve new private
output files without loading NER models or downloading artifacts.
Import-free effect-path inventory and local/CI guards for unclassified HTTP/database effect candidates and built-in MCP, REST, CLI and Journey dependencies reaching legacy FHIR/OpenMRS write-back paths. Keep narrative extraction independent of legacy writer modules while preserving their public parser imports (#3780).
Project export-ready clinical registry cases to protected local NAACCR XML files using a caller-supplied dictionary, field map and Journey fact resolver. Verify exact export and fact custody, validate declared types and lengths, report omitted fields, and use opaque patient keys without submitting records or bundling registry dictionaries or edits (#3809).
Project reviewed, evidence-bound clinical relations into offline FHIR R4 reason links and relative-owned FamilyMemberHistory conditions, with source-free Provenance, reference checks and explicit projection losses.
Typed Python and TypeScript workflow clients with strict metadata validation, bounded inspection retries and polling, one-attempt mutations and shared offline contract vectors (#3670).
Default-disabled, authenticated REST workflow inspection, receipt submission and cancellation intent over an injected custody service, with bounded governance schemas and content-free responses (#3668).
Offline summarize and nli verify CLI commands with bounded local inputs,
value-free console results, explicit baseline selection, and exclusive private
summary/metadata destinations (#3801).
Metadata-only MCP governed workflow preflight, preview, status and human-review request tools. Server-held single-use consent protects handoff creation; no approval token or reviewer claim crosses MCP and no tool executes an effect (#3762).
OpenMedKit metadata-only agent review, native OMOP preview, approval receipt and run-evidence parsers with shared Python conformance vectors. Local custody hooks record typed receipt observations and reject replay/changed actions; parsed evidence never grants clinical execution authority (#3669).
Include governance, FHIR/OMOP write-back, lineage and MCP namespaces in the per-PR public API compatibility gate, with the same deprecation window and dependency-light capture of declared exports.
Verify patient-keyed calendar-day shifts and emitted subject surrogates across five Journey source surfaces using bounded local witness processors. Keep protected values in memory and emit keyed proofs and counts; report missing format coverage explicitly. Release schema v1.1 adds consistency blockers while preserving v1.0 report parsing and historical signature verification. Bind the candidate's synthetic fixture to shared patient/date witnesses and regenerate its golden output; preserve the historical v3.0 inputs at their immutable tag (#3806).
Add typed Python service-client grounding, profiling and cohort resolution, with explicit OpenAPI path coverage and documented operational exclusions (#3754).
Project governed FHIR create/update attempts and refusals into deterministic, value-free R4 AuditEvents with injected clocks, opaque references, controlled reviewer categories and receipt digests. Preserve unknown commit state for recovery (#3787).
Added an offline exact-source v3.1 agent candidate manifest runner,
openmed gates agent-release and a release script with deterministic
HMAC-signed, content-free gate decisions and fail-closed source checks (#3776).
Added content-free v3.1 release-gate adapters for executed adversarial and recovery reports and versioned replay, reference-server and clinician-agreement inputs, with canonical source digests and fail-closed coverage checks (#3775).
Added an offline DQD 2.9.0 results normalizer and subprocess-compatible command for the OMOP quality bridge, with value-free diagnostics, strict version and table allowlists, and synthetic leakage and status regression fixtures (#3799).
Record purpose-ticket allow and deny decisions as value-free access events, with injectable sinks, a local append-only digest chain, and aggregate counts that exclude record selectors and values (#3793).
Added an opt-in, exporter-free agent OpenTelemetry bridge with validated value-free phase spans, aggregate event counters and latency, omitted or hashed run IDs, and synthetic in-memory privacy/determinism tests (#3792).
Added a locally signed native-measure evidence adapter with aggregate counts, pinned definition, value-set and engine custody, time-boundary conversion, and explicit drift-comparator parity and differences (#3789).
Add offline qualification of caller-supplied local planners against synthetic grant, schema, minimum-data, review-stop and hostile-result scenarios, with counts-only reports and non-compensable critical gates (#3790).
Added local revisioned agent-handoff storage with bounded review leases, atomic competing-decision refusals, exact action/evidence commitments and append-only opaque decision receipts preserved across restart (#3666).
Added fail-closed purpose-ticket authorization for complete typed tool-result batches, with independent patient/encounter/namespace/snapshot scope checks, strict nested field projections, original evidence identity and content-free quarantine diagnostics. Privacy scanning remains a separate boundary (#3671).
Added local, sealed blinded-review result import for summary evaluation, with exact artifact/rubric binding, citation and reviewer-agreement metrics, explicit incomplete/disputed states and synthetic-only release refusal (#3655).
Added explicit, confirmation-bound terminal recovery journal retirement, sealed digest/count records, non-resumable retired dispositions, and restart-safe checkpoint cleanup with summary anchor verification (#3791).
Added Doccano relation-extraction JSONL import and export for the versioned
annotation envelope, with de-identification digest binding, declared loss for
unsupported fields, review states and embeddings, controlled label and
relation normalization, and --format doccano CLI support (#3807).
Add governed workflow CLI plan/preview, inspection, human-review request, cancellation and explicit resume over caller-injected local services. Bind resume to exact action/state/receipt custody, keep inputs protected and outputs value-free, and distinguish denial, review, cancellation, conflict and missing adapters with stable exit codes (#3667).
Add versioned synthetic multilingual full-pipeline clinical-brief regressions for Latin, RTL, Indic and code-switched evidence, public surface/native packet parity, protected-span offset round trips and governed conflict refusals. Reports separate contract parity from unsupported language capabilities and unevaluated model quality. Accept unset optional replacement aliases without weakening populated-alias consistency checks (#3659).
Added passive preliminary FHIR R4 clinical-brief documents in Python and OpenMedKit, preserving evidence commitments, citation order, digests and human-review limitations with closed-subset round trips, explicit conversion losses and final narrative/metadata privacy gates (#3658).
Fixed-order clinical brief composition over explicitly reviewed synthetic evidence, with local generation, calibrated NLI integration, exact citations, rendered-packet privacy checks and value-free review/provenance output.
Add opt-in cooperative clinical-brief cancellation and existing-budget deadlines, empty interruption outcomes, CLI/service parity, and native late-result guards.
Add opt-in v1 explicit evidence bindings for paraphrased clinical brief claims, bounded local generation, traceable offsets and unchanged verification/privacy gates, with aligned OpenMedKit validation and synthetic offline tests (#3654).
Add offline qualification of caller-supplied clinical NLI artifacts, separate development/held-out evaluation, digest-bound receipts and a drift-checked BriefContext callback; synthetic data never establishes clinical qualification (#3656).
Added a local Journey-fact ClinicalBrief context provider with explicit review verification, source/fact/profile/calibration bindings, typed mapping refusals and offline Python, CLI, REST and MCP injection tests (#3652).
Add the separately versioned reviewed-local clinical evidence contract and opt-in Python/OpenMedKit brief admission, with current source/review custody, digest-bound expiring receipts, typed refusals and synthetic offline fixtures (#3651). Existing evidence packets remain synthetic-only.
Added immutable run-scoped agent tool catalog bindings, exact-version pinned dispatch, deterministic restore eligibility and content-free re-review outcomes for implementation, schema and side-effect drift (#3665).
Added injected runtime clinical-agent authority revocation checks before sensitive reads, effects and recovery, with monotonic grant/ticket generations, descendant and approval invalidation, and content-free offline tests (#3664).
Added guarded dispatch on the existing workflow executor, enforcing signed grants, purpose tickets, minimum-data projections, pinned tool identity and exact single-use approval, with content-free cancellation and uncertain-effect recovery outcomes through injected local providers (#3661).
Add default-off agent effect admission, signed local transition receipts, independent rollback detection, emergency stop and fresh-enable CLI commands with synthetic approval/recovery safety controls (#3765).
Add opt-in FHIR R4 server $validate preflight through an injected target-bound
transport and opaque credential handle. Return controlled issue summaries and
request digests; block errors and unavailable validation before application review.
Add injected, bounded FHIR R4 write execution for existing conditional plans and assembled transactions. Bind exact wire details to consumed approval receipts, repeat local authorization and lineage gates, preserve predicates, version preconditions and idempotency keys, and require durable attempt reservations. Ambiguous outcomes require reconciliation without retry or automatic compensation (#3662).
Added deterministic, bounded FHIR R4 transaction assembly for approved creates and updates, preserving conditional/version requests and recording value-free Provenance with exact canonical Bundle digests (#3771).
Added offline FHIR write-origin labeling and provisional-status normalization for Observation, Condition and AllergyIntolerance, with exact-label rejection, role-gated attestation and value-free findings (#3766).
Added offline FHIR write-reference scope verification against purpose-bound ticket selectors, injected non-compartment resolution, guarded preview dispatch, internal transaction/contained checks and value-free denials (#3764).
Added opt-in private SQLite approval nonce claims shared across processes and restarts, fail-closed durable replay protection, and a compatible consumption store adapter for MCP consent receipts (#3763).
Added an offline, synthetic governed clinical reference workflow covering approval, FHIR and OMOP effects, action evidence, replay, and recovery (#3196).
Added metadata-only signed action ledgers and deterministic replay evidence, permanent circuit-breaker stops, cooperative resource budgets, and completed run invariant reports for governed agent runtimes (#2766, #2767, #2769, #2770, #3001). These contracts do not dispatch effects or replace adapter validation.
Added local FHIR Subscription intake checkpoints with bounded deduplication, ordered claims, payload-free quarantine, and explicit gap replay (#2776).
Added an offline Draft 2020-12 catalog for approval tokens and receipts, reviewer handoff packets, capability grant manifests and workflow recovery checkpoints and decisions, with local references, bounded fields, version, enum and digest drift tests (#3783).
Add the offline agent governance self-check and openmed agent self-check
command. Independently verify the actual schema catalog, bundled golden
summaries, closed fields, outcomes, correlations, timing and commitments,
returning deterministic value-free reports with negative controls.
Include the existing MIT-licensed JSON Schema validator in the base install so the governance self-check needs no development extra. Keep the license inventory, SBOM closure and standalone manifest aligned with that dependency.
Bundle seven metadata-only agent run-summary golden vectors with strict, offline validation of canonical JSON, Markdown digests and commitments (#3044).
Export deterministic, import-light Draft 2020-12 JSON Schemas for public agent outcome, correlation, timing and run-summary records.
Added the offline openmed models slm-verify command, which verifies a local
clinical SLM package, probes explicitly requested capabilities against the
declared metadata, and runs the memory preflight for an explicit device
budget, with a value-free JSON report, stable CLI error codes, and no runtime
import or network access (#3805).
Added bundled Draft 2020-12 JSON Schemas for clinical brief audit and response
records, evidence packets, NLI verification results and SDOH evidence reports,
with a committed fingerprint snapshot, schema drift comparison, optional
jsonschema validation and value-free contract tests (#3803).
Map reviewed brief profile fields to existing evidence-class budgets, bind value-free estimates into audit records and distinguish length overflow from unavailable local models. Export controlled summarizer errors and preserve their typed reason codes without private exception contexts. Use script-aware first-three-sentence boundaries for explicit comparison extraction; retain reviewed fact-coverage selection and native unloaded-model refusal (#3746, #3747, #3749, #3750).
Select explicit extractive brief sentences by unique reviewed fact coverage, independent omission rules and existing length allocations. Preserve ordered citations and expose insufficient-budget/resource-limit diagnostics without partial output; retain the first-three-sentence comparison baseline (#3653). Recomputed synthetic summary reports retain missing-adjudication failures.
Issue v2 human-approval tokens with signed key identifiers and issuance times, injected local key rotation, bounded lifetime/skew checks before nonce claims, and explicit v1 verification compatibility. Approval receipts now use a v2 codes-and-digests-only schema; failure examples cover the migration (#3768).
Require patched pyasn1 0.6.5 in the optional cloud profile and frozen
lockfile to address four ASN.1 decoder CPU-exhaustion advisories (#4372).
Require patched OpenTelemetry API 1.45.1 in the optional dev, otel and
service profiles and GitPython 3.2.1 in the docs profile, and lock
patched virtualenv 21.14.13, Brotli 1.2.1 and graphql-core 3.2.12. A
lock-only packaging override lifts Prefect 3.7's
packaging<26.3 cap so the lock keeps Airflow on patched 3.3.2; the base
SDK's dependency set is unchanged.
Refresh the service image's exact Debian forky pins for ca-certificates, liblzma5 and tzdata after the archive retired the previous versions.
Renew the maintainer-reviewed exception for unpatched NLTK
CVE-2026-81726 in the universal lockfile through 2026-11-03. NLTK remains
optional and absent from the core closure and service image. The exception
rejects other packages or targets, any reported fix, and expiry; this is
temporary risk acceptance, not remediation.
Raise optional dependency floors to patched aiohttp 3.14.4, banks 2.5.1, Mako 1.4.2, multidict 6.9.1 and tinytag 2.3.0. Apply the same floors to published extras and the frozen lockfile, preserving AWQ's Linux-only boundary and the base SDK's dependency set.
Require MCP 1.30 or newer for patched session handling in the optional MCP and development dependency profiles (#4221).
Refresh explicitly configured SMART credentials through atomic injected custody, validate token type/lifetime/granted permission unions, rotate refresh secrets, and revoke uncertain or invalid refreshes with value-free findings. Bulk export now rejects invalid or insufficient token responses and expired dispatches (#3773).
Align Python and OpenMedKit summary leakage guards for Unicode variants, unspaced scripts, short identifiers and recognized Hangul suffixes. Reuse detector normalization, retain value-free counts and source digests, and add shared synthetic brief/refusal parity controls (#3729).
Registered MLX summarizers now require a provisioned, symlink-free package with a trusted manifest pin before model construction. Admission verifies every file, refuses undeclared members, and derives task support, context limits and quantization from digest-bound metadata. Package failures expose controlled reason codes; unprovisioned aliases fail closed without fallback.
Confine OpenMedKit MLX manifest paths during download, cache inspection and local loading. Reject traversal and escaping symlinks with value-free errors, remove stale ready markers and confined partial writes on rejection, and retain root and nested tokenizer layouts (#3728).
Restore service-container builds with matching exact Debian OpenSSL 3.6.5-1 pins and retain lockfile SARIF diagnostics after image-build failures.
Require Strawberry GraphQL 0.327.2 or newer in the optional service and development profiles to fix completed-subscription retention and awaitable permission-result handling (CVE-2026-107727 and CVE-2026-107728).
Reject emails whose redacted headers cannot be serialized as clean MIME
with UnsupportedDocumentError instead of an unhandled stdlib error. The
rejection omits the stdlib error, which embeds raw header bytes.
Handle decoding failures through both batch readers' per-file error paths, continue later files in permissive mode, and retain strict propagation without including decoder details in error reports (#4201).
Read JSONL/NDJSON by physical file lines, preserving Unicode separators inside strings, blank-line handling, error-line numbering, and object-pair hooks (#4205).
Require MCP Python SDK 1.30 or newer within the existing 1.x compatibility range to fix unbounded Streamable HTTP session retention (CVE-2026-59951). MCP remains optional; the base SDK dependency set is unchanged.
Accept signed consent receipts through the real MCP call transport while retaining canonical argument schemas, bounded untrusted metadata, remote authorization and single-use verification (#4147).
Report explicit CLI refusal and check outcomes, including cancelled briefs; inspect bounded local cache metadata and report summarizer package pin readiness without downloading artifacts or constructing models (#3755, #3759).
Require review of empty cohort populations with value-free source coverage, explicit unknown counts, and typed warnings across saved runs, local DuckDB/Parquet resolution, and existing Journey reads. Read legacy saved executions without rewriting their immutable storage bytes (#3798).
Screen FHIR narrative markup/entities and bounded UTF-8 attachments before agent dispatch, quarantine unsupported attachments, and replace FHIR key-name exemptions with coded-element paths and content-free offset findings (#3769).
Normalize SMART v1 read/write/all permissions, wildcard resources, granular queries and non-clinical identity/launch/session scopes in offline scope audits. Compare constraints conservatively and report malformed input without echoing values; granular audit evidence contains only constraint digests (#3767).
Remove DICOM overlay, curve and icon carriers recursively; refuse unprocessed encapsulated documents, expose truthful pixel-cleaning outcomes, and stage combined header/pixel processing before writing its final artifact (#3725).
Apply the pinned DICOM 2026d Basic Profile catalog recursively, declare explicit retention/cleaning options with method codes, clean retained text through a caller detector, and replace identifying file metadata (#3727).
Parse clinical and Journey ISO calendar timestamps through one explicit profile on every supported Python version. Reject basic, week-date, ordinal and compact-offset input with value-free errors while retaining supported extended values, caller timezone rules and golden digests.
Honor per-call ONNX revision overrides before configured and default revisions (#3504).
Retain service concurrency permits and drain tokens until timed-out model work actually finishes, including GraphQL and batched requests. Add a value-free orphaned-work gauge and bound parallel GraphQL resolver work without changing REST timeout envelopes. Drain and orphaned-work accounting also cover deployments without concurrency or rate limits (#3902).
Omit unchecked PDF annotation appearances in both raster redaction paths; add explicit local detection of mapped annotation contents and appearances, fail-closed unsupported geometry, and value-free annotation reports (#3726).
Route Thai and Hebrew to their native packs and fail closed on unsupported letter-bearing scripts unless an explicit fallback is configured (#3738).
Share offset-safe decimal-digit and bidi-control normalization across both privacy APIs, fold Persian/Urdu keyboard variants for matching, and add seeded in-value bidi and native-digit robustness operators (#3739, #3741, #3742).
Detect validated dotted/prefixed Gregorian dates and derive date shifting and surrogate order from the shared locale table (#3740, #3743).
Redact additional HL7 v2 relative/prior identifiers with typed XAD surrogates, and C-CDA guardian, next-of-kin, informant and birthplace values; expose value-free coverage inventories for unmapped fields/elements (#3744, #3745).
Remove quadratic suffix copying from list boundary construction (#3518).
Treat eval-cache entries removed between lookup and read as cache misses, allowing recomputation (#3505).
Confine service bulk paths to operator roots, restrict request-selected models, and enforce complete REST route and GraphQL field scopes before execution.
Reject invalid JWT NumericDates, bound production token lifetimes, omit credentials from config serialization, and use private atomic config writes.
Launch reference services with content-free structured access logs only.
Synchronize Python, browser npm, Android, Swift demo, OpenAPI, standalone, Helm/operator and documentation version surfaces at 3.1.0. Update SDK candidate defaults to v3.1.0 and retain v3.0.0 as the API baseline.
Reconcile the complete v3.0.0-to-master range, including changes accidentally recorded under the historical v3.0.0 section; preserve that released section as recorded by its immutable tag.
Allow 60 minutes for the full CI pytest step after the Windows suite completed its tests at the former 45-minute boundary. Preserve full discovery, coverage, exit status and the bounded timeout contract.
Coordinate the REST shedding fixture through the event loop instead of an additional blocking executor waiter. Retain the same response, privacy, admission and post-release latency assertions, and release its controlled worker on fixture failure.
Classify exported typing.Any aliases consistently on Python 3.10 through
3.13, without treating interpreter representation changes as removed APIs.
Real changes from an alias to a class, callable or constant still fail the
public API guard (#3770).
Distinguish ordinary SDK PRs from version-changing release candidates using immutable base/head package metadata. Candidate PRs derive their stable baseline and migration guide from that boundary; explicit dispatches always evaluate all gates. Ordinary PRs produce a NOT_A_CANDIDATE scope receipt, never a READY report (#3770).
Map direct FHIR/OpenMRS writes, OMOP persistence and transaction exports to their explicit authority, review and execution boundaries in a v3.0-to-v3.1 migration guide with offline synthetic examples and marked fragments (#3785).
Add a governed-agent lifecycle overview and organize existing authority, approval, write-back, evidence, recovery, evaluation and assurance guides into one navigation section while preserving their public URLs (#3784).
Execute governed-agent Python documentation blocks offline with per-page synthetic context. Require explicit, content-bound exceptions for application adapter sketches, and fail on API drift or attempted external effects (#3795).
Explain SDOH extraction, registered determinants, custom extractors and cue table loading with runnable offline synthetic examples and source-value limits.
Document clinical brief trust boundaries across Python, CLI, REST, MCP and OpenMedKit, with existing mitigation tests and explicit residual risks; validate mitigation test links and source citations offline (#3752).
Explain local summarizer/NLI configuration, pre-load artifact admission and reviewed brief context wiring with executable synthetic doubles and drift checks for every public backend error and brief refusal (#3751).
Map v3.0 guarded clinical intelligence to source modules, guides and synthetic examples. Group clinical navigation by area and expose four previously unlisted guides, preserving page URLs and the existing Pages byte budgets.
OpenMed 3.0.0 — From Clinical Text to Patient Journeys
A clinical record is more than a collection of notes. It is a history of encounters, observations, medications, corrections and decisions—and every useful answer should lead back to the evidence behind it.
OpenMed 3.0 connects those pieces. This major release introduces Patient Journey: a longitudinal layer for turning fragmented clinical inputs into versioned facts, reviewable timelines, reproducible cohorts and interoperable datasets. Alongside it, new clinical-brief workflows bring local summarization, citation checks and human-review metadata into a shared pipeline.
The result is a broader toolkit for building clinical-data applications without making a hosted service mandatory. Start with local files and SQLite, add PostgreSQL or optional integrations when needed, and keep source evidence, access policies and correction history attached to the data.
Journey provides a common foundation for longitudinal clinical data:
A corrected fact does not erase its history. An ambiguous match does not silently become a patient link. Applications can inspect both the current view and the evidence that produced it.
New ingestion contracts make retries, interrupted jobs and source updates explicit:
Evidence-coordinate support is separate from complete media de-identification; applications must still verify the relevant document, metadata and pixel carriers.
Build cohort and registry workflows on top of the same longitudinal evidence:
Restricted patient datasets and vocabularies remain caller supplied. OpenMed does not bundle restricted patient corpora, UMLS, SNOMED CT or CPT assets.
OpenMed 3.0 adds:
Missing data remains insufficient evidence—not an inferred clinical conclusion.
Trial matches are candidates for review, not enrollment decisions. Adverse-event report signals do not establish patient-level causality.
The new composed clinical-brief pipeline connects reviewed evidence, local generation, clinical NLI, citations, rendered-output privacy checks and review/provenance metadata.
model="extractive".The composed clinical-brief evidence-packet contract is synthetic-only in this release. The walkthrough uses clearly labelled fixture providers; it is not a qualified real-patient summarization system. Missing models, runtime capabilities, calibration or review evidence refuse execution rather than silently selecting a cloud fallback. Heuristic NLI requires explicit backend="heuristic".
New training recipes and benchmark/promotion tooling are included, but this SDK release does not ship newly qualified clinical checkpoints. Training is optional; model qualification remains a separate evaluation and review process.
Read the clinical-brief guide · Run the synthetic walkthrough.
Python provides the broadest Journey implementation. Shared wire contracts and version numbers do not imply identical native execution coverage across Swift, Android and web clients.
HLA-B-27.These changes do not constitute blanket security clearance for every optional runtime. The optional Accelerate checkpoint-index advisory remains unresolved; deployments using that profile must review the dependency policy. The existing NLTK exception is narrowly scoped and time-limited, not a general downstream waiver.
This is a major release. Review the 2.5 → 3.0 migration guide before upgrading.
ground aliases now use text_or_entities instead of spans and return GroundingResult, not a bare list of grounded spans.OpenMedGraphQLContext requires explicit Journey resources and access policy; mount_graphql requires resource_getter.OPENMED_GLINER_SPAN_ARTIFACT explicitly for the optional sample-document GLiNER smoke test.Version-pinned installation coordinates:
# Python
pip install --upgrade "openmed==3.0.0"
# Web SDK
npm install openmed@3.0.0.package(url: "https://github.com/maziyarpanahi/openmed.git", from: "3.0.0")implementation("com.github.maziyarpanahi:openmed:v3.0.0")ghcr.io/maziyarpanahi/openmed:v3.0.0OpenMed is an assistive developer toolkit, not an autonomous clinical decision-maker. Clinical output requires qualified human review; synthetic tests and package compatibility do not establish clinical efficacy, regulatory compliance or medical-device suitability.
3.0 is about making clinical data connected, traceable and usable—while keeping control with the people building and reviewing the system.
Feature map · Full changelog · Compare 2.5.0 → 3.0.0
Note truncated.
Offline synthetic clinical-brief walkthrough, golden pipeline test, explicit fixture-provider disclaimers and a recording script.
OpenMedKit guarded clinical-brief packets, local Maple brief generation, native leakage/envelope/citation validation and shared Python wire fixtures.
Fail-closed summary model-promotion gates and seeded, counts-only extractive/MLX benchmark evidence. Failed reports remain failed and do not qualify models; training and model publication are optional, separate from the library/SDK release.
Aligned the published and scanned service-container runtime repairs, omitted unrelated desktop source trees, checked installed dependency consistency and removed the installer from the final runtime; no vulnerability waiver added.
Guarded clinical-brief CLI, REST, MCP, Python and TypeScript interfaces with local review lookup, separate protected/audit CLI files and content-free logs.
Fixed-order clinical brief composition over explicitly reviewed synthetic evidence, with local generation, calibrated NLI integration, exact citations, rendered-packet privacy checks and value-free review/provenance output.
Added cache-only, registry-pinned MLX summarization with capability, memory,
prompt-digest and source-token leakage guards. The deterministic CPU baseline
now requires explicit model="extractive"; missing runtime or artifacts fail
closed instead of silently changing backends (#3234).
Added offline benchmark suite listing and description, explicit task metadata, and aggregate report comparison with missing-evidence and regression verdicts (#286).
Added a Polish (pl) PII language pack with native locale surrogates,
validated PESEL handling, date, phone, address, and postcode patterns, and
synthetic offset and zero-leakage regression fixtures (#294).
Added deterministic citation minimality checks for guarded atomic claims, with opaque offset-only reports, token counts, explicit context budgets, and human-review flags (#2742).
Added deterministic claim-level citation support metrics with optional clinician adjudication, privacy-safe orphan and unused-evidence counts, and synthetic offline regression coverage (#2754).
Added deterministic de-identification-aware citation boundary validation with post-redaction digest binding, replacement-boundary rejection, and value-free review reports (#2741).
Added a deterministic, local unsupported-claim rate for clinical summaries, with four-state scoring, per-claim-class bootstrap intervals, aggregate-only reports, and synthetic offline regression coverage (#2749).
Added deterministic, value-free guarded clinical-output provenance manifests with input/evidence drift detection, model and policy fingerprints, and contiguous human-review transitions (#2578).
Added an offline clinical NLI sequence-classifier adapter for pinned PyTorch and ONNX artifacts with explicit class mapping, calibrated abstention, and structured contradiction prechecks (#3235).
Added deterministic, domain-separated SHA-256 commitments for validated agent run summaries, with constant-time categorical verification, stable golden vectors, and value-free malformed-input handling (#3039).
Added domain-separated, metadata-only fingerprints for validated federated update schemas, with constant-time schema comparison (#3055).
Added deterministic, metadata-only multimodal processing summary differences with signed aggregate deltas and sorted input/output digest changes (#3052).
Added deterministic, aggregate-only review-yield metrics for guarded clinical relation candidates, with accepted, corrected, rejected, duplicate, and deferred outcomes by relation class (#2753).
Added assertion-aware FHIR R4 AllergyIntolerance, Immunization, and
Encounter exporters with RxNorm, SNOMED CT, and CVX coding support.
Added immutable, versioned longitudinal Journey contracts for clinical artifacts, evidence locators, facts, conflicts, resolution events, and dataset snapshots, including deterministic JSON, bundled schemas, and derivation-graph validation (#3320).
Added durable local Journey storage with content-addressed artifact bytes, append-only SQLite metadata and fact history, atomic graph transactions, typed outcomes, deterministic migrations, integrity checks, and point-in-time reads (#3321).
Added backend-neutral Journey storage composition, PostgreSQL metadata parity, transactional migration recovery, and allowlisted content-addressed object namespaces with local-only defaults (#3322).
Added replay-safe ingestion jobs with versioned manifests, exclusive leases, gap-safe checkpoints, value-free retries, explicit cancellation, quarantine promotion, SQLite and PostgreSQL ledgers, and restart recovery tests (#3323).
Added physically separated identified and de-identified projection stores, capability-scoped transform vault access, purpose and consent policy gates, cache revocation, and durable value-free access auditing (#3324).
Added conservative patient and encounter identity resolution with opaque source keys, deterministic exact links, review-isolated candidate plugins, transactional merge and split decisions, and durable provenance (#3325).
Added offline structured-evidence adapters for text, existing document results, FHIR R4, HL7 v2, CDA R2, CSV, and XLSX with exact coordinates, parser provenance, bounded quarantine, and value-free results (#3326).
Added a versioned clinical model-pack contract and local-only task router with pinned artifact integrity, license and quantization gates, explicit runtime capability outcomes, and deterministic fallbacks (#3328).
Added model-independent clinical-fact normalization across extraction, assertion, temporality, relation, medication, laboratory, observation, and social-determinant outputs with evidence and derivation gates (#3330).
Added a local-first ingestion-to-fact orchestrator with versioned stage manifests, value-free derivation edges, exact replay, descendant-only reprocessing, typed quarantine, and atomic Journey writes across five synthetic structured source formats (#3327).
Added deterministic terminology resolution cascades with versioned mapping provenance, explicit ambiguous and unmapped outcomes, an append-only local review queue, opt-in local semantic candidates, and value-free coverage summaries (#3331).
Added evidence-preserving fact deduplication, dimension-specific conflict sets, versioned policy and human resolutions, append-only correction history, guarded review packets, and counts-only queue SLA summaries (#3332).
Added paginated point-in-time longitudinal journey views with typed patient, encounter, time, event, status, source, and review filters; complete immutable provenance drill-down; correction-aware state; and a value-free timeline graph (#3333).
Added versioned PDF, OCR, image, and DICOM evidence coordinates with invertible rotation, crop, and scale chains; exact source-space round trips; explicit DICOM tag provenance; and typed quarantine behavior (#3334).
Added a reproducible four-task Journey specialist pack with pinned low-cost encoder/span recipes, deterministic offline split manifests, a hard shared USD 1,000 GPU ledger, frozen-holdout calibration and failure slices, quantized-delta gates, model cards, and fail-closed promotion fallback (#3329).
Added deterministic projection of resolved clinical facts into versioned OMOP 5.4 person, visit, note, condition, drug, procedure, measurement, observation, and source-to-concept tables with explicit concept-0 outcomes, evidence custody, correction-aware replacement, retained ETL lineage, vocabulary-license gates, and split-leakage protection (#3335).
Added an optional OMOP 5.4 quality bridge with digest-bound local and caller-injected remote adapters, signed aggregate conformance, completeness, and plausibility reports, typed non-success states, and reproducible reconciliation against a frozen synthetic reference projection (#3336).
Added provenance-preserving FHIR R4 round trips for Journey facts and resolution events, byte-stable cohort-definition exchange with snapshot and criterion custody, explicit conversion-loss reports, and an optional digest-bound out-of-process cohort service bridge (#3337).
Added immutable saved-cohort definition versions and execution manifests, deterministic point-in-time reruns, four-state membership with criterion and time-window evidence, opaque-key workflow explanations, local append-only persistence, and vocabulary-license boundaries (#3338).
Added governed dataset snapshots from saved cohorts and ingestion jobs with deterministic JSONL, annotation, and Parquet artifacts; fail-closed split and distribution gates; license custody; and separately authorized, audited identified exports (#3339).
Added content-addressed clinical registry definitions, evidence-bound cases, owner-scoped assignments, non-skippable review and adjudication, correction history, counts-only queue summaries, and policy-bound value-free exports (#3341).
Added deterministic native clinical measures and pinned optional CQL/ELM JSON evaluator bridges with explicit population, unknown, error, evidence, value-set, time-window, trace, aggregate-summary, and drift contracts (#3342).
Added evidence-bound met, open, not-applicable, and insufficient-data care-gap states with conservative missing-data behavior, mandatory unresolved-input review, immutable corrections, and versioned history (#3343).
Added opt-in public clinical-trial metadata synchronization with immutable study versions, source and cache integrity checks, typed drift failures, deterministic incremental updates, and offline metadata queries (#3344).
Added deterministic public trial-criteria parsing, local candidate reranking, five-state evidence-bound criterion evaluation, value-free review packets, fail-closed aggregate matching, and a provenance-bound benchmark (#3345).
Added opt-in public adverse-event ingestion, report-level PRR and ROR calculations, explicit count suppression, immutable dataset/filter/policy provenance, and separately typed chart-level suspected relations (#3346).
Added an evidence-grounded patient and cohort query planner with closed read-only tool calls, bounded SQL guards, snapshot citations, access receipts, explicit abstention states, and inert prompt-injection boundaries (#3347).
Added a single versioned Journey resource contract across Python, REST, GraphQL, and bounded read-only SQL, including cursor pagination, namespace/purpose/field policy, typed non-success states, generated schema snapshots, and client parity (#3348).
Added deterministic half-open window planning for offline streaming audio, with integer-millisecond offsets, real per-window overlap, keep/merge/drop tail policies, arithmetic window-count bounds, and value-free rejection of boolean, negative and overflowing input (#3005).
Added decoded-memory planning for multimodal batches: estimates image, DICOM, and audio memory from validated manifests and an explicit caller-supplied factor policy, then accepts, splits into ordered batches, or rejects against an inclusive byte budget with per-batch overhead. Missing factors or geometry, including PDF page counts, stay unevaluable, and saturating 64-bit arithmetic never yields an accept (#3091).
Added a bounded, dependency-free PDF page geometry preflight that reads the version, page count, and inherited media box, crop box, and rotation for each page, including FlateDecode object streams, and reports numbers and stable reason codes only while rejecting encrypted or over-limit files (#3088).
Added a metadata-only image orientation preflight that compares all eight EXIF orientations with declared mirror and rotation transforms and optional dimensions, returning aligned, transform-required, ambiguous, or invalid verdicts with stable reason codes and deterministic JSON (#3090).
Added deterministic, privacy-safe diffs of agent run summaries with sorted added and removed workflow identifiers and artifact digests, signed outcome, tool-call, and duration deltas, and byte-stable JSON and Markdown output (#3040).
Added offline lifetime and audience validation for agent capability grants, with caller-supplied evaluation instants, symmetric clock-skew tolerance, an optional lifetime ceiling, ordered reason codes, and byte-stable reports that carry no grant payload (#3080).
Added PHI-safe categorical error envelopes for agent execution, mapping the documented public failure classes onto stable codes, validated class/stage combinations and derived retryability, with no free-text field and no exception message, argument or traceback retained (#2995).
Added a closed allowlist contract for agent event attributes covering correlation and governance identifiers, execution stages, outcome codes, digests, bounded counts, durations and flags, with a distinct sensitive-key refusal that never echoes the submitted key (#2994).
Added deterministic page batching for document inference, planning ordered half-open page ranges under page-count and per-batch pixel budgets with reject or isolate oversize policies, exact gap-free partitioning, and value-free bound and overflow checks (#3048).
Added bounded classic TIFF header preflight for both byte orders, with an allowlisted first-IFD tag set, overflow-checked counts, offset bounds and cycle detection inside a loaded prefix, value-free errors, and synthetic plus Pillow-written file-level regression tests (#3046).
Added immutable local ASR input profiles and a metadata-only compatibility check over privacy-safe WAV headers, returning compatible, resample, downmix, review or incompatible with ordered reason codes and closed, bounded, sorted-unique profile validation (#3050).
Added bounded PNG IHDR and JPEG frame-header geometry preflight with CRC and length validation, a marker scan bounded by fill, marker-count and byte budgets, value-free errors, and synthetic plus Pillow-written file-level regression tests (#3008).
Added deterministic dependency validation and topological ordering for agent action graphs, with identifier-only nodes, categorical duplicate, dangling, self-edge and cycle findings, and documented lexicographic tie-breaking that makes the order independent of input order (#2996).
Added deterministic, metadata-only validation of append-only agent run event sequences with duplicate, gap, ordering, cross-run and post-terminal findings, stable finding order, and byte-stable JSON reports that never accept or echo event payloads (#2997).
Added content-free audio format distribution summaries with validated channel, sample-rate, bit-depth, duration-bucket, and format categories, small-cell suppression, and deterministic sorting, with synthetic golden regression tests (#3093).
Added content-free frame-sampling manifests for clinical video with strict monotonicity, bounds, declared-strategy validation, and a recomputed coverage digest, with synthetic regression tests (#3092).
Added deterministic SPDX identifier normalization for synthetic lineage with a committed permissive allowlist, deprecated-alias mapping, LicenseRef support, and stable unknown and malformed outcomes, with synthetic regression tests (#3120).
Added deterministic, allocation-free audio resampling plans with reduced rational rates, floor, nearest-even, and ceiling rounding policies, exact or rounded status, overflow-checked arithmetic, and an optional duration-error tolerance, with synthetic regression tests (#3049).
Added allocation-safe image geometry derivation with checked pixel counts, reduced aspect ratios, orientation classes, and overflow-checked optional memory estimates from validated dimensions, with synthetic regression tests (#3047).
Added Unicode whitespace regression coverage for vital-sign parsing and fixed the
cases it found: non-breaking, narrow, thin, and other Unicode spaces inside labels
such as oxygen saturation no longer turn a vital sign into unknown, and units
such as mm Hg match their ASCII spelling (#3106).
Added serialize_measurement_trends, compact JSON for measurement trends that is
byte-identical for equivalent input orderings and rejects non-finite derived values,
with golden fixtures for ordered, mixed, unknown, and incomparable trends. Points
tied on one timepoint are now ordered by value instead of input position (#3107).
Added a synthetic local-extension example and documentation page for the status vocabulary, covering an explicit local path, provenance validation, and a duplicate-cue guard (#3108).
Training and new-checkpoint publication are optional research, without a library-release milestone or feature-blocking relationship. SDK readiness verifies retained artifacts and pointers separately from model promotion.
Align active Python, npm, Android, Swift demo and Helm version surfaces at 3.0.0. Document grounding and GraphQL migrations for the major release, and include Journey and clinical-brief golden coverage in SDK readiness evidence.
Clinical NLI now fails closed when no released local checkpoint is registered;
lexical verification requires backend="heuristic", and results expose
value-free four-state metadata instead of source or claim text (#3235).
Reject non-positive bootstrap sample counts and invalid alpha probabilities before producing a confidence interval.
Release an interrupted circuit-breaker recovery probe only when its ownership token still matches, preserving health evidence.
Raise typed client errors for unfollowed non-2xx responses, including redirects in JSON and streaming requests.
Keep load-test clients connected through final response bodies and count incomplete responses as failures.
Align grounding-index recall references, queries and report metadata with the selected vocabulary systems.
Grow the dense alias search budget when duplicate aliases would hide available distinct concepts.
Allow model loading when optional architecture metadata is absent or empty; retain advisory warnings and normal factory failures.
Accept source-aligned literal hash tokens in space-delimited CoNLL rows without displacing ordinary hash comments.
Reject non-finite numeric keep-alive durations and conversion overflow while preserving explicit expiry opt-out aliases.
Return the result-cache instance selected under the global lock so concurrent capacity changes cannot replace a caller's return value.
Preserve already-qualified OpenMed logger names so module-level logging configuration applies without a duplicated openmed namespace.
Close newly created Pulsar clients when subscription or adapter setup fails, preserving the setup error and successful caller ownership.
Clean up temporary evaluation reports after failed writes, closes, or publication without masking the original error or replacing an existing report.
Measure awaited coroutine execution in the profiling decorator, preserving coroutine identity and recording elapsed time on failure or cancellation.
Scope shared tokenizer cache entries to loader identity, including bound methods, so different loader implementations cannot share the wrong tokenizer.
Canonical span hashes, pipeline audit-record hashes, and trace pseudonyms now use private random HMAC keys by default. Reuse a pipeline or redactor instance, or supply the same non-empty private key, when stable hashes across calls are required. Web and React Native calls also accept explicit keys for this purpose. Empty explicit keys are rejected; redaction labels and offsets are unchanged.
GitHub Actions are pinned to immutable commits, enforced by CI, and container publishing permissions are limited to the publish job.
Require Torch 2.13.0 or newer in optional model extras and current runtime recipes to address checkpoint/native-operation vulnerabilities. Torch remains absent from the base SDK; deployments must requalify platform and converter compatibility. Separate optional dependency findings are not waived.
Make the Swift clinical-document GLiNER characterization explicitly opt-in
through OPENMED_GLINER_SPAN_ARTIFACT, matching other local model smoke tests.
SDK tests no longer silently select an unrelated home-cache model; clinical
coverage assertions and thresholds remain unchanged.
Require patched Banks, Datasets, HTTP/2 and OAuthlib dependencies only in the affected optional integrations, with matching lock constraints and regression checks; the base dependency set is unchanged. OAuthlib 4 removes deprecated JSONP revocation support and changes server-side grant validation ordering.
Reuse Windows process-memory API types and bindings during streaming-table RSS sampling so the sampler does not retain a new ctypes pointer type per batch. Actual current-memory readings and unchanged overflow limits remain enforced, including the 16 MiB large-file regression.
Require patched PyJWT versions in the locked dependency graph and bound CI tests with per-test timeout diagnostics rather than six-hour silent hangs.
Continue optional dependency diagnostics when a module raises during import, recording the failure in the report (#3500).
Reject unsupported RequestBudget mapping keys instead of silently ignoring misspelled limits (#3509).
Preserve hash characters inside quoted configuration values while stripping trailing comments (#3502).
Strip BIO prefixes only at label beginnings, preserving interior labels such as HLA-B-27 (#3503).
Preserve tokenizer-provided subword offsets from any standard Mapping
output when projecting token labels, keeping the absent-offset fallback (#3561).
Key GLiNER model cache entries by requested device so a cached instance is not moved under a later caller (#3501).
Include breaking commits with unmapped types in release notes.
Hash dataset files with bounded memory (#3510).
Report non-object catalog JSONL rows through the normal coherence diagnostic path with physical line numbers, without exposing row values or paths (#3562).
Respect explicitly empty gold annotations (#3511).
Reject missing requested splits in named CMeEE directories before loading another split, while preserving explicit-file and generic-directory use (#3563).
Load prefetched Hugging Face models from the standard cache during offline inference, including Transformers 5.x pipeline and component loading (#1983).
Validate calibration targets as booleans or numeric zero/one, including aliases, instead of treating non-empty strings as positive labels (#3568).
Reject non-integer sharding counts before reading documents.
Include ranked grounding alternatives in MedMentions top-k accuracy without crediting withheld selections or turning abstentions into predictions (#3569).
Reject overlapping sibling items at nested list levels.
Require strict decoder validation before auto-detecting ISCII, preserving malformed Latin-1 strings through privacy preprocessing instead of raising or partially rewriting the input (#3242).
Keep zero-length spans empty when converting legacy encoding offsets (#3507).
The static Python API comparison against v2.3.0 reports zero removed symbols, zero narrowed signatures, and zero new deprecations . The REST API retai…
OpenMed 2.5 brings clinical privacy and extraction previews, richer FHIR and OMOP exports, safer document intake, and new privacy-policy and audit tools to the Python SDK. It preserves the v2.3 family-based model-registry API while adding an explicit API for managing separate model tiers and formats.
This release covers changes since v2.3.0. Python 3.10+ remains supported, and the core dependency set is unchanged.
ClinicalPrivacyProcessor combines explicit language, category, and role controls with document-level review status. The clinical_preserve profile targets identifiers while retaining clinical context such as diagnoses, medications, doses, and measurements. Full-document ONNX batching preserves source offsets and input order, with bounded cancellation and per-document outcomes.fa pack adds Persian patterns, Eastern-digit normalization, Jalali date handling, and Iranian national-ID surrogate support. Persian is also available in the REST language enum. Its default model route remains a compatibility placeholder; this release does not supply newly qualified dedicated Persian weights.Clinical-preserving processing is opt-in and remains a preview. No language/model route is clinically qualified by default; applications must establish qualification for their intended data, language, policy, and runtime.
to_fhir() facade dispatches supported entities to Condition, Observation, MedicationStatement, and Procedure resources, assembles deterministic Bundles, and reports unmapped labels. A conservative DiagnosticReport exporter adds explicit status handling and reference normalization.measurement, procedure_occurrence, visit_occurrence, observation_period, and note_nlp rows, preserving supported numeric values, units, dates, and source offsets. Cohort validation checks keys, relationships, vocabulary references, and note provenance.Synthetic FHIR/OMOP conformance coverage includes the official HL7 R4 validator, a malformed-resource negative control, and OMOP column, key, and reference checks. The Java validator and restricted vocabulary content are not bundled.
preflight_asset brings manifest, media-type, profile, resource-limit, and digest checks into one ordered accept-or-abstain report. An unevaluable check is reported as an abstention rather than a successful validation.
MOBILE_V1 and DESKTOP_V1 profiles bound bytes, dimensions, pages, frames, and duration before decoding. Batch summaries detect duplicate identifiers and digests and validate aggregate limits.New local tools make privacy decisions and their supporting evidence easier to inspect:
These reporting APIs use bounded metadata, counts, offsets, hashes, and provenance instead of retaining protected source values. Deletion, upload, trace cleanup, and other external effects remain explicit operations.
The optional Snowpark adapter adds caller-managed warehouse de-identification. The separate privacy-proxy application requires an injected transport, keeps mappings scoped to each request, and supports placeholder restoration across streaming responses.
Agent integrations gain strict provider-result and run-summary schemas, opaque correlation identifiers, typed governance identifiers, and artifact references with digest and size metadata.
Federated-training helpers add scheduling windows, round-status summaries, validated update metadata, and aggregate metrics with clipping declarations and minimum-group suppression. Retraining tools rank aggregate evidence, score trigger signals, and prepare recipe proposals. Compute, cost, energy, and carbon accounting adds explicit budget reports.
These additions provide orchestration and evidence contracts. They do not automatically train, deploy, or qualify a model.
Existing v2.3 registry callers keep their family-based API. RegistryService, family CLI selectors, unambiguous model aliases, and caller-owned schema-v1 files remain supported. An SDK upgrade does not require migrating those files.
Applications that need separate channels for model tiers or formats can opt into SlotRegistryService and schema-v2 slots keyed by family::tier::format. Stored slot versions are independent of version-like tokens in model repository names. The compatibility adapter retains existing schema-v2 files and refuses ambiguous multi-slot family operations without writing.
The static Python API comparison against v2.3.0 reports zero removed symbols, zero narrowed signatures, and zero new deprecations. The REST API retains its existing paths and component schemas, with Persian added to the language enum.
Other fixes include compatible Swift dependency resolution, stricter ONNX metadata validation, request-scoped privacy-proxy restoration, Windows artifact-deletion identity checks, and refreshed documentation and container dependencies.
The clinical, governance, and orchestration additions described above are Python capabilities. Matching npm, Swift, and Android version numbers do not imply that every new Python API is implemented on those platforms.
Use these coordinates once the v2.5.0 packages and tag are published:
# Python SDK
pip install --upgrade "openmed==2.5.0"
# Optional Hugging Face and FHIR integrations
pip install --upgrade "openmed[hf,fhir]==2.5.0"
# JavaScript package
npm install openmed@2.5.0Swift Package Manager:
.package(url: "https://github.com/maziyarpanahi/openmed.git", from: "2.5.0")Android through JitPack:
implementation("com.github.maziyarpanahi:openmed:v2.5.0")Container:
docker pull ghcr.io/maziyarpanahi/openmed:v2.5.0For Helm deployments, select the release image with image.tag=v2.5.0. Review the migration guide before adopting slot registries, new serialized schemas, or the clinical-preserving preview.
OpenMed is assistive software. Clinical outputs require qualified review and must not automatically determine diagnosis, treatment, or billing. Structural validation and synthetic regression tests do not establish clinical efficacy, deployment-specific privacy performance, or regulatory compliance.
Before deployment, evaluate identifier recall, critical leakage, source-span integrity, clinical-text preservation, date handling, and surrogate consistency on the intended languages and runtimes. Model artifacts and their existing release targets are unchanged by this SDK release.
Versioned links become available with the v2.5.0 tag:
Full Changelog: v2.3.0...v2.5.0
OpenMed 2.5 adds clinical privacy and extraction previews, local privacy and audit controls, FHIR and OMOP validation, bounded multimodal intake, and registry and training orchestration. This release compares against v2.3.0; there is no intervening v2.4.0 tag. See the release notes and migration guide.
Added deterministic, privacy-safe directionality validation for guarded causal, treatment, procedure-indication, and medication-indication relations, with typed fail-closed endpoint and direction errors (#2737).
Added deterministic, local unit-dimension compatibility checks for dose, rate, concentration, and laboratory relation candidates, with explicit review findings for incompatible or unknown units and value-free reports (#2738).
Added deterministic, offline SDOH negated-need resolution with determinant-specific assertion scope, value-free provenance, and human review for double-negated, contradictory, and uncertain findings (#2731).
Added deterministic, value-free SDOH experiencer classification and patient-level filtering for patient, household, family, and unknown evidence with reviewable offset provenance (#2724).
Added duplicate-cue validation for status vocabularies: cues that collide after the existing Unicode, case, and whitespace normalization are rejected on load, both within one status and across statuses, with value-free errors (#3104).
Added duplicate and conflicting sense validation for abbreviation inventories: equivalent candidates, short forms that normalize together, and repeated JSON keys now fail on load instead of being silently dropped, and one long form under distinct semantic types is kept as separate alternatives when a local inventory is merged (#3105).
Added a functional-status zero-shot NER domain with ADL, assistance, mobility, functional-scale, assistive-device, and cognitive-status labels, synthetic span fixtures, and offline per-label coverage reporting (#911).
Added complete detection of bounded German postal-address fields and fragment protection inside known clinical phrases, with person-name counterexamples and independent mask/remove/replace regression checks.
Added preview clinical-preserving privacy processing with explicit language, category and role controls, full-document ONNX tensor batching, bounded cancellation and per-document review status. Clinical protection, source offsets and output policy remain consistent across the safety sweep.
Added German clinical context, temporal and quantity extraction regressions, memory-streamed Tesseract OCR and PDF reading-order/redaction checks. These preview capabilities require independent task and language qualification.
Added deterministic schema compatibility evaluation for versioned agent artifacts with closed outcomes, stable reason codes, caller-declared ranges, fail-closed unknown kinds, and strict SemVer 2.0 precedence (#3000).
Added bounded BMP CORE/INFO header geometry preflight with explicit limits, value-free errors, and synthetic file-level regression tests (#3114).
Added bounded GIF logical-screen and bounded global-color-table preflight with explicit limits, value-free errors, and synthetic file-level regression tests (#3115).
Added tiny RF64 refusal fixtures that pin WAV envelope rejection, short-read boundaries, and stream restoration without changing the parser (#3117).
Added bounded VP8, VP8L, and VP8X WebP geometry preflight with explicit limits, value-free errors, and synthetic file-level regression tests (#3116).
Added structural locale normalization and explicit alias duplicate/unsupported-format checks with synthetic file-level regression fixtures (#3119).
Added a privacy-safe multimodal preflight report (preflight_asset) that
runs manifest validation, bounded media-type detection, modality profile
checks, limit-profile evaluation, and a bounded digest pass in a fixed order
and returns one accept-or-abstain PreflightReport with ordered, allowlisted
findings, a preflight AbstentionRecord, and byte-stable JSON; unevaluable
checks, including a PDF's pixel rules, abstain rather than accept (#2980).
Added immutable pre-decode limit profiles for multimodal assets (MOBILE_V1,
DESKTOP_V1) with inclusive ceilings for bytes, pages, pixels per unit, total
pixels, frames, and audio duration, evaluated over the privacy-safe asset
manifest into deterministic LimitFinding records; unevaluable rules,
including a PDF's pixel rules, are reported as insufficient_metadata rather
than assumed safe (#2956).
Added an exact schema version and strict, bounded dictionary and JSON parsers for privacy-safe agent run summaries, including duplicate-field, non-finite number, unknown-field, and unsupported-version rejection (#3038).
Added strict, content-free multimodal provider result envelopes with bounded counts and timing, deterministic serialization, and value-free failures (#3006).
Added an optional Snowpark adapter and generated Python UDF SQL for in-warehouse text de-identification with lazy dependency loading, compatible Snowpark registration, and escaped SQL literals (#2369).
Added an immutable, provenance-aware local terminology cache keyed by exact vocabulary releases, with deterministic response fingerprints, stale-release refusal, response-free reports, and bounded value-free validation (#2400).
Added a bounded, deterministic, offline policy-migration checker with fail-closed schema and protection-type changes, privacy-safe reports, and a report-bound human acknowledgement gate for weakening changes (#2407).
Added an in-memory no-PHI telemetry exporter with closed counter families, bounded dimensions and totals, atomic event validation, exception-type-only categorization, deterministic JSON and Prometheus rendering, and no mandatory network transport (#2414).
Added a bounded, deterministic tabular re-identification risk report with aggregate-only JSON and Markdown renderers, immutable report state, fail-closed consistency checks, and locally derived threshold outcomes (#2411).
Added a deterministic local pre-push privacy scanner that checks every new commit blob for direct identifiers, secrets, and sensitive structured fields; emits value-free reports; supports narrowly versioned synthetic-fixture allowlists; and installs atomically while preserving existing hooks (#2298).
Added a versioned, bounded privacy policy-as-data schema for jurisdiction, recall floors, de-identification actions, surrogate strategy, and privacy-safe audit retention, with deterministic local-only loading, strict duplicate and alias handling, and value-free validation failures (#2406).
Added a bounded, thread-safe privacy budget ledger for named aggregate release contexts with atomic epsilon/delta charging, counts-only evidence, immutable configuration views, and value-free failures (#2410).
Added caller-owned HMAC-SHA256 audit-report key rotation with bounded key material, key-ID based current and retained-key verification, canonical mapping checks, fail-closed provider handling, and value-free failures (#2408).
Added a bounded, deterministic minimum-necessary structured field selector with caller-declared purpose mappings, policy allowlists and denylists, fail-closed unknown declarations, value-free decision explanations, and projection restricted to selector-approved fields (#2412).
Added a bounded, counts-only audit-artifact retention planner with explicit disposition rules, deletion evidence, remaining-set verification, strict input fields, and fail-closed future timestamps (#2409).
Added a deterministic, offline CycloneDX 1.6 evidence generator for the base runtime dependency closure, with source and manifest hashes, bounded local inputs, atomic output, and no embedded URLs or build paths (#2416).
Added a deterministic offline dependency risk report that correlates local locked versions with caller-supplied advisory snapshots, emits bounded value-free risk summaries, and performs no package-manager or network calls (#2417).
Added deterministic counts-only trace privacy audit artifacts with canonical policy and file hashes, immutable category counts, value-free JSON and Markdown renderings, stable file fingerprinting, and private atomic writes (#2302).
Added deterministic, PHI-free local release compute, cost, energy, and carbon tracking with orchestrator-linked stage timings, per-run and rolling budget verdicts, family/tier/workload breakdowns, optional advisory queue throttling, and hash-verified ledger replay (#1244).
Added a bounded, deterministic nested-resource redaction contract with explicit scalar paths and actions, stable arrays and identifiers, closed policy validation, and raw-value-free reports and failures (#2413).
Added declarative field-level FHIR and OMOP de-identification policies with fail-closed identifier handling, patient-consistent date shifting, schema linting, CSV/Parquet support, and resumable FHIR NDJSON integration (#2187).
Added the canonical grounded-span to_fhir() facade with label-driven
Condition, Observation, MedicationStatement, and Procedure dispatch,
deterministic Bundle assembly, PHI-free exported/unmapped label counts, and
graceful skipping for labels without an exporter. The facade remains the
same callable as the established grounded exporter and never synthesizes a
Patient resource.
Added a deterministic, local key-custody metadata validator for synthetic signing and surrogate workflows, with lifecycle transition checks, purpose/algorithm compatibility, digest-only reports, and fail-closed rejection of bytes, secret-like, or unknown fields (#2648).
Added bounded local deletion verification for fingerprinted sensitive artifacts, with symlink, alias, and hard-link refusal, independent recovery copies, commit-stage rollback, and counts-only evidence (#2418).
Added a bounded, manifest-driven deletion impact planner with deterministic counts-only reports, reverse-dependency analysis, ownership checks, and explicit plan-bound confirmation before injected local execution (#2529).
Added a deterministic, offline OMOP cohort export validator for key, relationship, vocabulary, and NOTE/NOTE_NLP provenance invariants, with aggregate counts and content-derived row fingerprints instead of source values (#2402).
Added a bounded, deterministic FHIR R5 Bundle round-trip fidelity diff with stable entry matching, explicit serializer-difference declarations, and value-free reports containing structural paths, types, and SHA-256 digests (#2401).
Added bounded, deterministic structured access reviews that compare workflow read and export declarations with resource schemas and deny policies while keeping schema values out of JSON, Markdown, and validation failures (#2419).
Added bounded, policy-aware diffs for aggregate redaction summaries, with closed value-free inputs, deterministic policy fingerprints, and structured action, category, and count changes (#2426).
Added bounded privacy-policy composition with explicit scope-overrides, deterministic scope precedence and inheritance, and validated value-free decision traces (#2522).
Added a bounded, metadata-only synthetic privacy regression corpus manifest with deterministic fixture hashes, policy and severity coverage validation, immutable invariants, and atomic local persistence (#2420).
Added a bounded, local evidence-bundle integrity verifier with file and manifest hashes, policy and provenance checks, and value-free reports (#2427).
Added a bounded tabular schema-drift privacy gate with counts-only evidence, conservative stable-ID matching, and release blocking for unsafe role or structural drift (#2524).
Added a bounded, deterministic referential-integrity auditor for surrogate maps with cardinality, collision, orphan, and cross-table consistency checks, closed input schemas, and counts-only value-free reports (#2538).
Added a bounded, deterministic nested structured-redaction idempotence checker for comparing shape, action, surrogate, policy, and count evidence across synthetic FHIR- and OMOP-shaped passes without retaining protected values (#2523).
Added a bounded, offline privacy evidence replay verifier with counts-only synthetic manifests, stable policy/environment/result fingerprints, and privacy-safe schema, environment, policy, and result drift reports (#2527).
Added an offline manifest-coherence regenerator and CI drift gate for the runtime model registry, PII language defaults, governed README counts, registry model cards, and generated model and benchmark documentation (#77).
Added deterministic agent action phases with an explicit reviewed resume path, terminal boundaries, and payload-free transition validation (#2998).
Added exact OMOP CDM v5.4 visit_occurrence, observation_period, and
note_nlp exporters with deterministic local keys, bounded clinical dates,
source offsets, and assertion-derived NLP term fields (#2360).
Added exact OMOP CDM v5.4 measurement and procedure_occurrence row
exporters with shared Athena concept resolution, deterministic unmapped
fallback, and preservation of numeric lab values, units, and ranges (#275).
Added dependency-free US Core 9.0.0 conformance checks for exported Condition, laboratory Observation, MedicationRequest, and AllergyIntolerance resources, including base-R4-first validation, must-support warnings, required-binding errors, canonical profile resolution, and compact CC0 constraint metadata (#2366).
Completed the synthetic grounding/export conformance suite with fail-closed
out-of-process HL7 FHIR R4 validation, an official-validator malformed
resource negative control, expanded ACHILLES-style OMOP column/key/reference
checks, and paired JSON/Markdown BenchmarkReport artifacts (#2359).
Added a versioned federated update metadata envelope with coordinator-owned parameter expectations, bounded exact shape arithmetic, deterministic JSON, clipping declarations, and value-free rejection of unknown or identifying fields (#3010).
Added typed, canonical governance identifiers for capabilities, purposes, policies, workflows, and tools, with shared validation and value-free diagnostics (#3042).
Added a versioned no-PHI exception taxonomy for telemetry and audit records, with owner-free approval metadata, bounded digest-only evidence, explicit UTC expiry checks, deterministic serialization, and value-free validation failures (#2528).
Added a bounded, deterministic audit-envelope parser with redacted payload metadata, canonical fingerprints, strict schema and signature validation, and value-free diagnostics (#2594).
Added a deterministic, local-only privacy exception budget gate that counts bounded synthetic waiver metadata by severity, scope, expiry, and policy fingerprint, failing closed on exceeded or unbounded exceptions (#2591).
Added local, deterministic FHIR ValueSet expansion over caller-loaded free
vocabulary snapshots plus explicit FHIR $expand and ECL delegation to a
caller-supplied terminology endpoint. Results include versioned provenance;
caching is user-controlled, and restricted member codes are never persisted
without a second explicit policy opt-in (#926).
Added closed, versioned federated aggregate metric envelopes with finite clipping bounds, minimum-group suppression, coarse participant bands, controlled privacy mechanisms, confidence intervals, deterministic JSON, and value-free rejection of client-level or unknown fields (#3011).
Added dependency-free base FHIR R4 structural validation for eight exported clinical resource types, including deterministic structured findings, cardinality and primitive datatype checks, fixed required bindings, Bundle aggregation, and a compact CC0-derived constraint table (#2364).
Added typed, 128-bit opaque correlation identifiers for agent runs and actions, with strict kind-aware parsing, deterministic metadata-only JSON, parent-action validation, and value-free failures (#2973).
Added strict, content-free agent artifact references with opaque identifiers, a closed artifact-kind vocabulary, versioned schema IDs, digest and size metadata, deterministic JSON, and value-free validation failures, including oversized integers and deeply nested JSON (#2999).
Added a conservative, deterministic FHIR DiagnosticReport exporter with
R4/R5 union allowlisting (32-field), explicit unknown status, type-gated
scalars and Reference normalization, effective[x] mutual exclusivity,
deep-copy evidence preservation, field-name-only value-free errors, and
no network or clock dependency (#2566).
Added privacy-safe multimodal asset batches with opaque batch identifiers, canonical asset ordering, duplicate identifier and digest detection, a bounded asset count, derived byte, page, frame, and duration totals, and sorted value-free findings for invalid, oversized, overflowing, or inconsistent batches (#3002).
Rekeyed the committed model-registry state to schema v2: sparse
family::tier::format release-channel slots (the baseline_key
convention shared by gates/baseline.json, gates/rollout_state.json,
and the release ledger), created only by coordinate-matched RELEASABLE
promotions, with assigned per-slot SemVer that is validated as stored
state and never recomputed from repo-id version tokens. Ships a
fail-closed one-time v1 migration (registry_ctl.py migrate) that maps
pointers through committed baseline coordinate evidence and leaves the
file unchanged on any ambiguity (#1804).
Preserve the v2.3 family registry API, CLI selectors, serialized views, and
unambiguous aliases; expose slot operations through SlotRegistryService
with an explicit v2 state contract and fail-closed compatibility adapter.
Pin Swift tokenization to the validated 0.1.24 release so clean package resolution cannot select an incompatible MLX dependency graph.
Add SDK-only readiness evidence for unchanged model artifacts and pointer targets while retaining signed model gates for model releases.
Validate ONNX label metadata before importing optional runtimes; malformed labels now fail at the metadata boundary.
Keep local privacy-proxy request mappings scoped to one request and reject unknown, duplicate, or malformed placeholders on inbound restoration.
Refresh Debian certificate and OpenSSL package pins used by the container build and validate release artifact size budgets against measured growth.
Fixed verified artifact deletion and rollback on Windows Python 3.12 by comparing explicit creation timestamps across pathname and descriptor stat results, while retaining identity and in-read mutation checks.
OpenMed 2.5 expands the Python SDK with clinical privacy and extraction previews, local privacy and audit controls, bounded multimodal intake, FHIR and OMOP validation, and metadata-only training orchestration. The release baseline is v2.3.0; there is no intervening v2.4.0 release.
This page describes the candidate. Registry publication, immutable package tags, and deployed documentation are verified separately after publication.
The release adds local policy schemas, composition and migration checks, minimum-necessary field selection, field-level FHIR/OMOP policies, tabular risk and schema-drift reports, surrogate integrity checks, and structured redaction and idempotence evidence.
Audit tooling covers access scope and expiry, key custody and rotation, evidence integrity, replay, lineage and freshness, report size and cardinality, waiver lifecycle, exception budgets, policy coverage and simulation, and release-evidence aggregation. Reports use declared identifiers, counts, offsets, hashes, and bounded metadata rather than protected source values.
Encrypted surrogate mappings, local deletion planning and verification, retention cleanup, pre-push and CI privacy scanning, dataset-upload guards, and session-end trace scrubbing are explicit local operations. The optional Snowpark adapter supports caller-managed warehouse processing; it does not turn the core package into a network service.
Read the 2.3-to-2.5 migration guide, including
registry compatibility and the versioned model-state migration. Python 3.10+
remains supported. The core dependency set stays unchanged; Snowpark is
optional and the grounding-validate extra does not bundle the Java validator.
After publication, the intended coordinates are:
pip install --upgrade "openmed==2.5.0"
pip install --upgrade "openmed[hf,fhir]==2.5.0"
npm install openmed@2.5.0
.package(url: "https://github.com/maziyarpanahi/openmed.git", from: "2.5.0")
implementation("com.github.maziyarpanahi:openmed:v2.5.0")
Container consumers select ghcr.io/maziyarpanahi/openmed:v2.5.0.
The Helm application version and default tag are 2.5.0; explicit image
selection uses image.tag=v2.5.0.
The new clinical, privacy-governance, and orchestration APIs described above are Python capabilities. Updating the npm, Apple, and Android package versions does not imply these new APIs run natively on every platform.
Synthetic regression tests and package checks establish the behavior they exercise. They do not establish clinical efficacy, deployment-specific identifier recall, regulatory compliance, or clinical-device suitability. Clinical outputs require qualified review and must not automatically determine diagnosis, treatment, or billing.
This SDK candidate preserves the existing model pointer targets and model artifacts. The registry state representation changes independently of those targets. A model promotion requires its own staged golden and public SHIELD benchmark evidence, signed gates, and authorization under the release-stream policy.
The complete change record is in the changelog.
One column per month.