Documentation

This page describes the core objects and outputs you can expect from Sigilant Labs. Product capabilities evolve; the console reflects the exact fields exported for your runs.

Core objects

Run

A single execution request: model + target profile + candidate configuration space.

Inputs: model identifier, target profile, quant list, ctx/batch constraints, prompts/workload.

Variant

A specific configuration under evaluation (e.g., quant, context size, batch size, and runtime flags).

Each variant produces measured metrics and gate outcomes.

Artifact

Exported JSON/CSV describing the run summary, per-variant results, and the recommended configuration.

Designed for reproducibility and internal review.

Metrics (typical)

TTFT: time-to-first-token (latency proxy)
Throughput: tokens/sec once generation is underway
Peak memory: observed peak RSS/VRAM depending on profile
Quality gates: optional checks depending on workload and configured policies

Exact definitions depend on runtime and target profile; exports include the measurement context.

Recommendation logic

Recommendations balance throughput, latency, memory constraints, and optional quality gates. If no variant passes configured gates, the export will state the reason and provide the “best available” option with explicit caveats.

Support

Email support@sigilantlabs.com for access provisioning, invoicing, or operational questions.