FlagOS workload capability matrix (F4)

F4 turns the checked-in F1, F2, and F3 hardware records into one machine-readable answer about workloads that have been demonstrated through FlagQuantum's public backend="flagos" and device="flagos" boundary. It does not inspect or infer the provider's inner communication route.

Current result

The authoritative F4 artifact is artifacts/flagos_workload_capability_f4_20260826.json. It classifies the fixed evidence set as follows:

Capability	Status	Exact evidence boundary
flagos.statevector.forward	development_verified	complex64 and complex128, one node, 2/4/8 cards, sharded statevector
flagos.statevector.training	development_verified	complex64 and complex128 bounded backward, SGD, and Adam trajectories, one node, 2/4/8 cards
flagos.collectives.full	unsupported	complex reduce_scatter_tensor failed on the two-card F1 matrix; the other required collectives passed

development_verified means executable development hardware evidence exists for the stated workload and nothing broader. It is not production support, general scalability evidence, a performance claim, or a multi-node result. unsupported describes the tested full collective matrix; it does not prevent statevector workloads whose required collective subset passed.

Audit contract

Run the deterministic aggregator from the repository root:

python tools/audit_flagos_workload_capability.py

The tool hashes each source artifact, validates raw cases and rank semantics, and then writes the F4 artifact atomically. It fails closed if schemas, 2/4/8-card ladders, sharding, gradient or optimizer semantics, source revisions, provider claim flags, or explicit unsupported results disagree.

The output deliberately records:

• outer_backend = "flagos";
• inner_communication_route = "unattributed";
• flagcx_route_verified = false;
• all communication, scalability, production, and release claim flags as false.

The historical source artifacts may contain older diagnostic wording. F4 does not propagate that wording into a provider-identity claim.

Evidence inputs

• F1.1 two-card distributed conformance, including the complex collective matrix;
• F2 one-node 2/4/8-card sharded forward scale ladder;
• F3 one-node 2/4/8-card sharded backward and optimizer ladder.

All three inputs use the same locked Torch-FL source revision. Their individual FlagQuantum source revisions remain visible in the artifact descriptors.

F4 is a historical aggregation of those three inputs, so its single_device_capacity_failure_not_measured blocker remains unchanged. The later matched 32-qubit capacity result is recorded separately in F5; it does not retroactively rewrite the F4 artifact.