Module record_serve 

2026-09-26: The serve settings a gate record discloses about the server it measured (GateRecord::serve_resolved).

served_by names the in-tree recipe (gate::recipe_closure), and serve_overrides only the keys changed for the run; neither states what the server resolved those settings to (2026-10-02). The keys are defined here, in the crate that owns the record, and the server CLI fills them from its rendered serve flags.

Owner: bench gate (records). Invariants:

• A disclosure always carries SPECULATIVE, WEIGHT_QUANTIZATION and ACTIVATION_QUANTIZATION; every other key is present only when resolved or on.

Structs

LiveForward

2026-09-28: What a server reports about its forward (GET /forward): the server fills it from its model, the harness reads it into the record.

Constants

ACTIVATION_QUANTIZATION

2026-09-30: Key for --activation-quantization, always present on a record written since the flag exists, with the value’s canonical form (adaptive, declared, a ladder). A record without it predates the flag, and its server ran what adaptive names.

EXPERT_QUANTIZATION

2026-09-27: Key for --expert-quantization, present only for a tier other than the default fp8, with the tier’s name as the value (nvfp4-gate-up, nvfp4). The flag has no environment fallback, so an absent key means fp8.

FORWARD

2026-09-28: Key for --forward, present only for a forward other than legacy.

KERNEL_TREE

2026-10-03: Key for the SHA-256 of the kernel tree the serve’s memory plan read, present whenever the serve planned (every serve whose loader reads the MoE expert-table decision).

MAX_BATCH_SIZE

2026-10-01: Key for the slot count --max-batch-size auto resolved to, written auto:<n>; present only for auto.

MOE_EXPERT_TABLES

2026-10-02: Key for the MoE expert-table decision, present (skip) only when the serve’s memory plan dropped the transposed MoE prefill tables. Absent means built, as every serve before the decision existed did.

MTP_GATE

2026-09-26: Key for the --mtp-gate regime in force: force or auto.

PLAN_DIGEST

2026-09-28: Key for the live decode plan’s digest (metrale_circuit::digest::plan_digest), present only when the server runs a circuit forward.

PREFILL_CODISPATCH

2026-09-26: Key for --prefill-codispatch, present (true) only when the rendered serve gave the flag.

SPECULATIVE

2026-09-26: Key for whether --speculative was on: true or false.

W4A4_DOWNCAST

2026-09-26: Key for --w4a4-downcast, present (true) only when on. The flag defaults to false and has no environment fallback, so an absent key means off.

WEIGHT_QUANTIZATION

2026-09-28: Key for --weight-quantization, always present on a record written since the flag exists, with the tier’s name (declared, nvfp4). A record without it predates the flag, and its server ran what nvfp4 names.

Functions

disclosure

2026-09-26: The disclosure for a server whose rendered flags resolved to these values.

merge_live_forward

2026-09-28: Add the live forward to resolved: FORWARD when it is not legacy, and PLAN_DIGEST. requested is the forward the rendered serve asked for; a server running another one, or a circuit without a digest, or legacy with one, is refused: the record would state a configuration the measurement did not run.