Optional backend conformance

The optional peptide pharmacokinetics and uptake adapters (PeptiVerse, PeptiVerse CPP and PlifePred2) share one small conformance contract. It records what the wrapper was developed against separately from the files it actually resolved on this machine, and it never downloads a model during prediction.

Three records carry that information:

• BackendSpec declares the endpoint, upstream compatibility target, license, serialization format, prediction-only entry point, and reviewed platform and interpreter support.
• ArtifactIdentity records the role, size, and SHA-256 of each exact code, weight, preprocessing, or auxiliary file.
• BackendInventory combines those records with output-affecting settings into a path-independent identity suitable for prediction metadata and cache keys.

Artifact state and runtime capability are reported separately:

Artifact status	Meaning
missing	At least one required file was not found.
mismatch	A file differs from the reviewed checksum.
unverified	Files were hashed, but no reviewed checksum was declared.
verified	Every file matches its reviewed checksum.

Capability	Meaning
blocked	Required files are missing or mismatched.
artifacts_located	Files exist, but their provenance is unverified.
artifacts_verified	Files match the reviewed inventory; inference has not run.
inference_reproduced	The verified backend completed an offline prediction.

These labels prevent file discovery from being presented as reproduced inference. A dataset, training script, interactive design program, web-only service, or checkpoint for another endpoint cannot satisfy this contract.

Exact input and cache identity

Input, output, and backend identity stay on separate records, each with its own role:

Record	Fields	Role
Prediction	peptide, flanks, source name/offset, allele/TCR, predictor name/version	output and source-occurrence projection
PeptideInput	sequence, terminal chemistry, optional attachments/context and source provenance	shared exact peptide input and conditional-fragment helper; cleavage rejects unsupported attachments
AssayMeasurement	free-text chemistry, species, matrix, cell type, conditions	benchmark source record, not predictor input
BackendInventory	code/model/feature files and output-affecting settings	backend content identity

PeptideInput is the shared, versioned input record for reviewed PK, uptake and individual cleavage adapters. It keeps the canonical L-peptide sequence, terminal chemistry and attachments separate from occurrence/source provenance and from a nested PeptideContext. The context can describe route, formulation, cargo, administered and released material, measured analyte, study, assay species, matrix, cell type/subtype, maturation state, readout, timepoint, and additional conditions. Missing fields are serialized as null; they are not filled with physiological defaults or inferred from a patient.

Plain strings remain a backward-compatible shorthand for an unmodified canonical L-peptide with free termini. When an exact input names terminal modification or an attachment that a sequence-only adapter cannot consume, the adapter rejects it instead of stripping the chemistry. With on_unsupported="record", reviewed adapters preserve its batch position as an explicit unsupported result and score the supported remainder.

prediction_cache_key() hashes the input's chemical form and descriptive model context together with BackendInventory.identity_sha256, which already contains exact model/code/feature assets and output-affecting settings. Source occurrence IDs, genes, and offsets remain on the returned record but do not poison inference reuse: repeated occurrences may share a cache key while still appearing once per occurrence in ordered output.

run_python_sidecar() supplies the shared runtime boundary. It runs a reviewed, noninteractive Python prediction entry point under the selected isolated interpreter, disables user-site imports, activates standard offline modes, blocks Python socket connections, captures failures, and enforces a timeout. This is defense in depth for reviewed sidecars, not a sandbox for hostile code: pickle, joblib, and unrestricted torch checkpoints can execute arbitrary code. They must match reviewed checksums by default, and opting into an unverified artifact is an explicit trust decision.

PeptiVerse inventory

PeptiVerse verifies the pinned upstream inference.py, the exact transformer_wt_log checkpoint/configuration/calibration files, and the ESM2 weights, configuration, and tokenizer files. Its manifest names Transformer_WT_Log directly; it never accepts upstream's fallback to transformer_wt. The ESM2 snapshot must be local before construction, and the sidecar replaces the two unused SMILES embedders so prediction cannot trigger PeptideCLM or ChemBERTa downloads.

PeptiVerse CPP shares those source and embedding resources but requires only its exact svm_gpu_wt CPU SVC and original best_models.txt. The bridge checks scikit-learn 1.7.2 before loading the checkpoint, then verifies the estimator, class orientation and native threshold 0.5493. CPP-only installs need no half-life artifacts.

predictor_version contains both the revisions mhctools was developed against and a path-independent SHA-256 identity of the files actually supplied. After a successful call, artifact_inventory.capability changes from artifacts_verified to inference_reproduced; merely finding the files does not make that claim.

PlifePred2 inventory

PlifePred2 verifies the natural-peptide forest from the official 1.0 wheel, Pfeature's QSO implementation at revision 93636eb95bed9df2893b7a0c56b1215e648ecdbf, both QSO distance matrices, and the three additional data files that Pfeature reads at process startup. Only those reviewed resources are copied into the isolated feature workspace.

The official wheel pins scikit-learn 1.4.2. The sidecar checks that version before loading the joblib forest, so an interpreter that merely happens to unpickle it with warnings is not reported as exact reproduction. Both the outer predictor and nested Pfeature process run with socket access disabled and timeouts. The output-setting choice (assume_log10_seconds) is part of the identity because it changes whether a duration is emitted.

Capability matrix

This release makes no blanket platform claim for real-model inference. The conformance suite exercises artifact checks, input/output contracts, offline execution, and isolated stub sidecars on Linux (Python 3.11-3.14 in CI) and macOS (Python 3.11 and 3.14 locally); actual third-party inference remains an opt-in smoke test. A verified predictor instance advances to inference_reproduced only after that local smoke succeeds.

Backend/candidate	Endpoint-specific release	Static artifact gate	Real inference
PeptiVerse	Human-serum half-life, exact transformer_wt_log	Exact source/model/calibration/ESM2 inventory; automated conformance	Opt-in smoke; no published platform combination yet
PeptiVerse CPP	Canonical CPP classification, exact CPU svm_gpu_wt	Exact source/SVC/threshold/ESM2 inventory; runtime version guard	Direct native API and fixed-source conformance on Linux/macOS, Python 3.11, CPU
PlifePred2	Undocumented blood-half-life native regression	Exact forest/QSO/resources inventory; automated conformance	Reproduced on macOS arm64 / Python 3.12 / CPU; otherwise opt-in
POSEIDON	None established	Blocked: research/training repository is not an inference-complete endpoint release	Not run
PERSEU	None established	Blocked: interactive design path and serialized models do not provide a reviewed prediction-only entry point	Not run

Finding a repository, dataset, or checkpoint is therefore never shown as an inference capability. A platform claim can be added only alongside a recorded real-model smoke result for that exact inventory.

The PlifePred2 entry records a local smoke on 2026-09-14 with scikit-learn 1.4.2 and asset identity d1fb877035f93b1302a975c47c08a084d8757964692c64f89f108792da0915c0. The verified backend returned native scores 3.157266404183623 and 3.7934723194625297 for the two opt-in reference peptides and advanced to inference_reproduced. This is evidence only for the listed runtime; it does not establish the endpoint's undocumented biological semantics.