Config

The current typed config surface is split between:

• graphids.exp.config for run and experiment configs plus typed stage payloads (FitRunPayload, CacheRunPayload, ExtractRunPayload, AnalyzeRunPayload)
• graphids.primitives for data, model, loss, scaler, representation, and discovery primitives

The older plan-chassis documentation is kept as historical reference in docs/reference/orchestration.md.

temporal_hybrid

The temporal_hybrid model primitive exposes the newer architecture as a modular config option in the same stack as the temporal classifier, GAT, RNN, and VGAE baselines.

Supported top-level knobs:

• objective: supervised, anomaly, or joint
• memory: type: tgn, enabled, and reset_on_stream_end
• backbone: type: none | gru | ssm_lite | mamba, layers, and dropout
• heads: classification, next_id, iat, and payload_delta
• anomaly: mode: regression | nll plus optional log-scale clamps
• rhythm: optional causal per-ID IAT summary branch
• motif: optional recent destination-ID/IAT motif branch
• loss_weights and anomaly_score_weights

Head defaults follow the objective. Supervised runs enable the classification head, anomaly runs enable the self-supervised heads, and joint runs enable both. Invalid combinations fail during config/model construction: anomaly runs cannot receive a classifier loss, supervised runs require classification, and anomaly or joint runs require at least one anomaly head.

memory.time_encoding_dim enables CAN-TGN-style elapsed-time encodings for the source and destination ID memories. memory.use_source and memory.use_destination support memory ablations. anomaly.mode: nll changes the IAT and payload-delta anomaly terms from SmoothL1 errors to Gaussian negative log-likelihoods; next-ID remains categorical NLL in both modes.

Canonical smoke examples live in configs/experiments/temporal_hybrid_*_smoke.yml. Final benchmark-matrix configs are intentionally separate from the core smoke integration so the architecture can be validated before broader scheduling.