Capability-to-evidence manifest

This manifest maps every advertised FastPLMs 1.0 capability to its user documentation, runnable example, and required validation tier. It is a coverage contract, not a statement that an unreported run passed. The exact checkpoint list and family declarations come from src/fastplms/models.toml.

The Example column links a curated CLI when that interface exposes the whole capability. Programmatic-only forms instead link their runnable CPU contract so the manifest does not imply broader CLI coverage than the example provides.

Frozen ESMC release evidence

Status: pending. Default documentation generation never discovers or trusts reports implicitly. Release rendering requires an explicitly selected, complete schema-v3 set of exactly 30 records on one exact GH200/aarch64 target: 18 measured eager, SDPA, and Flex records plus 12 structured FlashAttention 2/3 unavailable records across three checkpoints and two immutable sequence panels. The set must also carry the final candidate/reference image identities, dependency lock, installed inventory, and official source attestations. A partial, stale, malformed, self-digest-invalid, or cross-device set fails closed and cannot replace this status.

Locked oracle package compatibility exception

The frozen oracle lock permits exactly one nonzero pip check diagnostic: nvidia-cusparselt-cu13 0.8.1 is not supported on this platform. It applies only to nvidia-cusparselt-cu13==0.8.1 on NVIDIA GH200 480GB / linux / aarch64. The vendor filename tag is py3-none-manylinux2014_aarch64, while the wheel metadata declares py3-none-manylinux2014_sbsa. The exact wheel is nvidia_cusparselt_cu13-0.8.1-py3-none-manylinux2014_aarch64.whl with SHA-256 4dca476c50bf4780d46cd0bfbd82e2bc10a08e4fef7950917ce8d7578d22a23f. FastPLMs accepts this vendor metadata mismatch only after the lock, installed inventory, wheel bytes, metadata tag, and target identity all match. The wheel is not rewritten (validated-vendor-metadata-exception-no-wheel-rewrite). Any additional diagnostic or identity drift fails closed.

Executable evidence selectors

Only the selectors below are claimed. Their scopes are intentionally narrower than a whole family or validation tier. A tier appearing on another row does not automatically apply to the capability in this row.

Selector	Tier/job	Executable target	Scope
cpu:autoclass-runtime	cpu_contract	tests/cpu/test_autoclass_evidence_matrix.py::test_autoclass_runtime_evidence_matrix_exactly_matches_all_45_entries tests/cpu/test_autoclass_evidence_matrix.py::test_autoclass_runtime_evidence_targets_are_collected_cpu_tests	Every family-level AutoClass entry and its explicit tiny runtime contracts.
artifact:checkpoint-autoclasses	artifact	tests/release/test_published_automodel.py::test_local_artifact_offline_autoclass_parity	Every advertised AutoClass for every built checkpoint, grouped by checkpoint.
compliance:sequence-primary-head	compliance	tests/parity/test_native_results.py::test_native_exact_checkpoint_contract tests/parity/test_native_results.py::test_native_every_checkpoint_bf16_inference	The official-parity head only: AutoModel for ANKH or a family without MaskedLM; otherwise AutoModelForMaskedLM.
compliance:ankh-seq2seq	compliance	tests/parity/test_native_results.py::test_native_ankh_explicit_decoder_prompt_generation	ANKH AutoModelForSeq2SeqLM explicit-prompt generation only.
compliance:structure-automodel	compliance	tests/structure/test_esmfold_folding_compliance.py tests/structure/test_esmfold2_folding_compliance.py	ESMFold and ESMFold2 AutoModel folding paths only.
benchmark:claim-eligible-primary-head	benchmark	benchmarks/suite.py::benchmark_cases[claim_eligible=True]	The benchmark-selected head for representative sequence checkpoints and ESMFold2 projection cases; startup and embedding cases are excluded.
cpu:attention-contracts	cpu_contract	tests/cpu/test_attention_contracts.py	Portable dispatch, masks, fallback, fake FA2/FA3, ESMC Flex/FA3, and eager/SDPA gradient contracts.
nightly:sequence-backends	nightly	tests/integration/test_backend_consistency.py	Current GH200 eager, SDPA, and Flex forward/backward paths. Flash kernels are not downloaded, built, or executed in the current locked environment.
historical:fa2-focused	historical	tools/remote/run.py::_kernel_capability_preflight	Policy records prior real FlashAttention 2 focused execution, but the immutable execution report is not bundled in this repository and no current GH200 numerical claim is inferred from it.
compliance:flash-unavailable-gh200	compliance	tests/parity/test_native_results.py::test_esmc_bf16_calibration_and_biological_holdout	Complete report-bound FA2/FA3 unavailability records and fail-closed dispatch on the frozen release environment.
compliance:deep-backends	compliance	tests/parity/test_native_results.py::test_native_representatives_all_backends	Every advertised backend on the pinned deep sequence representative per family.
benchmark:claim-eligible-backends	benchmark	benchmarks/suite.py::benchmark_cases[claim_eligible=True]	Backends emitted by claim-eligible sequence and ESMFold2 benchmark cases.
cpu:embedding-contracts	cpu_contract	tests/cpu/test_embedding_contracts.py	Ordered inputs, biological masking, pooling, streaming, and persistence.
cpu:e1-embeddings	cpu_contract	tests/cpu/test_e1_contracts.py	E1 raw-sequence and MSA embedding persistence.
feature:e1-rag	feature	tests/integration/test_e1_rag.py	E1 retrieval, MSA preparation, cache, scoring, and embedding flows.
cpu:ankh-contracts	cpu_contract	tests/cpu/test_ankh_contracts.py	ANKH encoder and explicit-decoder embeddings, layers, masks, and T5 views.
cpu:generation-contracts	cpu_contract	tests/cpu/test_generation_contracts.py	Tiny deterministic DPLM, DPLM2, and ESM3 generation contracts.
feature:generation	feature	tests/integration/test_dplm_generation.py tests/integration/test_esm3.py	DPLM, DPLM2, and ESM3 generation behavior in the feature suite.
cpu:peft	cpu_contract	tests/cpu/test_peft_contracts.py	Real initializer, collators, one optimizer step, and adapter/classifier reload.
nightly:peft	nightly	tests/unit/test_fine_tuning_example.py	Fine-tuning example contracts in the nightly feature job.
cpu:ttt	cpu_contract	tests/cpu/test_ttt_contracts.py	Seeded TTT initialization, update, reset, save, reload, and family isolation.
feature:ttt	feature	tests/integration/test_ttt.py	TTT integration behavior in the feature suite.
cpu:structure-contracts	cpu_contract	tests/cpu/test_structure_contracts.py	Tiny injected structure cores, public outputs, save/reload, and binder batching.
structure:public-contracts	structure	tests/structure/test_structure_public_helpers.py	Seeded Boltz helper, linker masking, real features, losses, and binder gradients.
structure:full-suite	structure	tests/structure	The declared GPU structure suite for folding and preparation behavior.
feature:binder	feature	tests/integration/test_binder_design.py	Seeded binder workflow, atom padding, critic ranking, and traceability.
cpu:artifact-example	cpu_contract	tests/cpu/test_documentation_contracts.py::test_artifact_loading_example_executes_local_only_autoconfig	The offline local-artifact example with AutoConfig.
cpu:task-head-example	cpu_contract	tests/cpu/test_documentation_contracts.py::test_task_head_example_executes_all_advertised_heads_offline	Offline ESM2 masked-LM scoring, contacts, sequence classification, and token classification through the documented example.

Curated offline example execution

Every curated example is routed to the exact collected CPU test nodes below. These tests run under the required offline cpu_contract gate.

Example	Tier	Exact executable CPU node
embedding_and_retrieval.py	cpu_contract	tests/cpu/test_documentation_contracts.py::test_embedding_and_retrieval_example_executes_with_ordered_sqlite
attention_switching.py	cpu_contract	tests/cpu/test_documentation_contracts.py::test_attention_switching_main_executes_optimized_and_masked_fallback
ankh_embeddings.py	cpu_contract	tests/cpu/test_documentation_contracts.py::test_ankh_embedding_example_executes_encoder_and_decoder_layers
generation.py	cpu_contract	tests/cpu/test_documentation_contracts.py::test_generation_example_executes_seeded_dplm_branch_offline
generation.py	cpu_contract	tests/cpu/test_documentation_contracts.py::test_generation_example_executes_seeded_dplm2_branch_offline
generation.py	cpu_contract	tests/cpu/test_documentation_contracts.py::test_generation_example_executes_seeded_esm3_trace
e1_rag.py	cpu_contract	tests/cpu/test_documentation_contracts.py::test_e1_rag_example_executes_local_msa_and_shared_persistence
ttt.py	cpu_contract	tests/cpu/test_documentation_contracts.py::test_ttt_example_executes_seeded_adapt_save_and_reset
structure_preparation.py	cpu_contract	tests/cpu/test_documentation_contracts.py::test_structure_preparation_example_executes_each_public_branch
artifact_loading.py	cpu_contract	tests/cpu/test_documentation_contracts.py::test_artifact_loading_example_executes_local_only_autoconfig
task_heads.py	cpu_contract	tests/cpu/test_documentation_contracts.py::test_task_head_example_executes_all_advertised_heads_offline
fine_tuning.py	cpu_contract	tests/cpu/test_peft_contracts.py::test_fine_tuning_main_wires_both_tasks_without_external_io
fine_tuning.py	cpu_contract	tests/cpu/test_peft_contracts.py::test_shipped_collators_create_tokenizer_aware_sequence_and_pair_batches
fine_tuning.py	cpu_contract	tests/cpu/test_peft_contracts.py::test_shipped_initializer_drives_one_peft_step_and_atomic_final_reload
binder_design_fastplms.py	cpu_contract	tests/cpu/test_structure_contracts.py::test_public_binder_workflow_pads_heterogeneous_prepared_atoms_without_truncation
binder_design_fastplms.py	cpu_contract	tests/cpu/test_structure_contracts.py::test_binder_example_main_wires_explicit_offline_cli_arguments
binder_design_fastplms.py	cpu_contract	tests/cpu/test_structure_contracts.py::test_binder_structure_loss_is_finite_and_differentiable

Families and AutoClasses

Family	Tokenizer mode	AutoClass	Weight status	Guide	Family workflow and runnable entry-point contract	Required evidence
esm2	tokenizer	AutoConfig	FastPLMs extension	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
esm2	tokenizer	AutoModel	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
esm2	tokenizer	AutoModelForMaskedLM	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses, cpu:task-head-example, compliance:sequence-primary-head, benchmark:claim-eligible-primary-head
esm2	tokenizer	AutoModelForSequenceClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses, cpu:task-head-example
esm2	tokenizer	AutoModelForTokenClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses, cpu:task-head-example
esm_plusplus	tokenizer	AutoConfig	FastPLMs extension	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
esm_plusplus	tokenizer	AutoModel	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
esm_plusplus	tokenizer	AutoModelForMaskedLM	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses, compliance:sequence-primary-head, benchmark:claim-eligible-primary-head
esm_plusplus	tokenizer	AutoModelForSequenceClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
esm_plusplus	tokenizer	AutoModelForTokenClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
esm3	tokenizer	AutoConfig	FastPLMs extension	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
esm3	tokenizer	AutoModel	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses, compliance:sequence-primary-head, benchmark:claim-eligible-primary-head
esm3	tokenizer	AutoModelForSequenceClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
esm3	tokenizer	AutoModelForTokenClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
e1	sequence	AutoConfig	FastPLMs extension	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
e1	sequence	AutoModel	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
e1	sequence	AutoModelForMaskedLM	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses, compliance:sequence-primary-head, benchmark:claim-eligible-primary-head
e1	sequence	AutoModelForSequenceClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
e1	sequence	AutoModelForTokenClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
dplm	tokenizer	AutoConfig	FastPLMs extension	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
dplm	tokenizer	AutoModel	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
dplm	tokenizer	AutoModelForMaskedLM	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses, compliance:sequence-primary-head, benchmark:claim-eligible-primary-head
dplm	tokenizer	AutoModelForSequenceClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
dplm	tokenizer	AutoModelForTokenClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
dplm2	tokenizer	AutoConfig	FastPLMs extension	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
dplm2	tokenizer	AutoModel	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
dplm2	tokenizer	AutoModelForMaskedLM	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses, compliance:sequence-primary-head, benchmark:claim-eligible-primary-head
dplm2	tokenizer	AutoModelForSequenceClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
dplm2	tokenizer	AutoModelForTokenClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
ankh	tokenizer	AutoConfig	FastPLMs extension	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
ankh	tokenizer	AutoModel	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses, compliance:sequence-primary-head, benchmark:claim-eligible-primary-head
ankh	tokenizer	AutoModelForMaskedLM	FastPLMs extension	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
ankh	tokenizer	AutoModelForSeq2SeqLM	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses, compliance:ankh-seq2seq
ankh	tokenizer	AutoModelForSequenceClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
ankh	tokenizer	AutoModelForTokenClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
boltz2	structure	AutoConfig	FastPLMs extension	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
boltz2	structure	AutoModel	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
esmfold	structure	AutoConfig	FastPLMs extension	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
esmfold	structure	AutoModel	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses, compliance:structure-automodel
esmfold	structure	AutoModelForSequenceClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
esmfold	structure	AutoModelForTokenClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
esmfold2	structure	AutoConfig	FastPLMs extension	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
esmfold2	structure	AutoModel	pretrained	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses, compliance:structure-automodel, benchmark:claim-eligible-primary-head
esmfold2	structure	AutoModelForSequenceClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses
esmfold2	structure	AutoModelForTokenClassification	base weights + untrained task head	guide	family workflow; runnable AutoClass contract	cpu:autoclass-runtime, artifact:checkpoint-autoclasses

Attention backends

Backend	Advertising families	Guide	Example	Required evidence
eager	ankh, boltz2, dplm, esm2, esm3, esm_plusplus, esmfold, esmfold2	guide	example	cpu:attention-contracts, nightly:sequence-backends, compliance:deep-backends, benchmark:claim-eligible-backends
flash_attention_2	esm2, esm_plusplus	guide	example	cpu:attention-contracts, historical:fa2-focused, compliance:flash-unavailable-gh200
flash_attention_3	dplm, esm2, esm_plusplus	guide	example	cpu:attention-contracts, compliance:flash-unavailable-gh200
flex_attention	dplm, e1, esm2, esm3, esm_plusplus, esmfold, esmfold2	guide	example	cpu:attention-contracts, nightly:sequence-backends, compliance:deep-backends, benchmark:claim-eligible-backends
sdpa	ankh, dplm, dplm2, e1, esm2, esm3, esm_plusplus, esmfold, esmfold2	guide	example	cpu:attention-contracts, nightly:sequence-backends, compliance:deep-backends, benchmark:claim-eligible-backends

Input, embedding, and storage contracts

Capability	Guide	Example	Required evidence
Sequence list or streaming FASTA	embedding API	embedding and retrieval	cpu:embedding-contracts
Ordered mapping or one-shot generator	embedding API	runnable API contracts	cpu:embedding-contracts
Biological-residue max_length, bounded token windows, and stable order	embedding API	runnable API contracts	cpu:embedding-contracts
Mean and standard-deviation pooling	embedding API	embedding and retrieval	cpu:embedding-contracts
Max/norm/median/variance/CLS/PARTI pooling	embedding API	runnable pooler contract	cpu:embedding-contracts
Full-residue and all-selected-layer output	embedding API	ANKH layers	cpu:embedding-contracts, cpu:ankh-contracts
Transactional sharded safetensors and exact resume	embedding API	embedding and retrieval	cpu:embedding-contracts
Read-only SQLite and ordered duplicate-preserving filters	embedding API	embedding and retrieval	cpu:embedding-contracts
Legacy SQLite conversion without pickle deserialization	embedding API	runnable converter contract	cpu:embedding-contracts
E1 raw-sequence and MSA-aware ordered embeddings	E1 guide	E1 RAG	cpu:e1-embeddings, feature:e1-rag
ANKH encoder/explicit-decoder hidden-state selection	ANKH guide	ANKH layers	cpu:ankh-contracts

Generation and adaptation contracts

Capability	Guide	Example	Required evidence
ESM2 pretrained masked-LM scoring and contact prediction	ESM2	task heads	cpu:task-head-example, cpu:autoclass-runtime
ESM2 sequence/token classification with explicitly untrained task heads	ESM2	task heads	cpu:task-head-example, cpu:autoclass-runtime
DPLM amino-acid diffusion generation	DPLM	generation	cpu:generation-contracts, feature:generation
DPLM2 modality-aware sequence/structure co-generation	DPLM2	generation	cpu:generation-contracts, feature:generation
ESM3 multimodal-conditioned generation	ESM3	generation	cpu:generation-contracts, feature:generation
ANKH task-prompted sequence-to-sequence generation	ANKH	ANKH embeddings and generation	cpu:ankh-contracts, compliance:ankh-seq2seq
Trainer/PEFT LoRA with immutable inputs and verified save/reload	fine-tuning	fine-tuning	cpu:peft, nightly:peft
Seeded TTT adapter initialize/update/reset/save/reload	TTT	TTT	cpu:ttt, feature:ttt

Structure contracts

Capability	Guide	Example	Required evidence
ESMFold single-chain folding and multimer-linker confidence masking	models	structure preparation	cpu:structure-contracts, structure:public-contracts, structure:full-suite, compliance:structure-automodel
Seed-scoped Boltz2 protein helper and BF16 execution policy	Boltz2	structure preparation	cpu:structure-contracts, structure:public-contracts, structure:full-suite
Atom-dense binder optimization and critic reporting	binder design	binder design	cpu:structure-contracts, structure:public-contracts, feature:binder
Offline local artifact AutoClass loading	artifacts	artifact loading	cpu:artifact-example, artifact:checkpoint-autoclasses
esmfold2 48-block full ESMFold2: single-sequence or optional MSA-conditioned protein inputs, typed complexes, ligands, nucleic acids, modifications, and bonds; pocket and distogram requests fail closed	ESMFold2	structure preparation	cpu:structure-contracts, structure:full-suite, compliance:structure-automodel
esmfold2_fast 24-block Fast ESMFold2: inference-optimized single-sequence conditioning with typed multichain and multimolecule inputs; every protein must have msa=None and MSA inputs fail closed	ESMFold2	structure preparation	cpu:structure-contracts, structure:full-suite, compliance:structure-automodel
esmfold2_experimental_cutoff2025 48-block full ESMFold2: single-sequence or optional MSA-conditioned protein inputs, typed complexes, ligands, nucleic acids, modifications, and bonds; pocket and distogram requests fail closed	ESMFold2	structure preparation	cpu:structure-contracts, structure:full-suite, compliance:structure-automodel
esmfold2_experimental_fast_cutoff2025 24-block Fast ESMFold2: inference-optimized single-sequence conditioning with typed multichain and multimolecule inputs; every protein must have msa=None and MSA inputs fail closed	ESMFold2	structure preparation	cpu:structure-contracts, structure:full-suite, compliance:structure-automodel

Release evidence must name the exact head, checkpoint and runtime revisions, tokenizer identity, backend, dtype, hardware, sequence or structure panel, seed, environment, and input hash. Missing evidence remains visibly pending; it must not be replaced by a synthetic benchmark number or an inferred claim.