Skip to content

Artifacts

This guide summarizes the deterministic artifacts emitted by Ascribe runs and how they support reproducible, audit‑ready research. Tables are Arrow‑friendly and stable across runs that share configuration and seeds.

What is recorded

A run writes a minimal, typed record of the CRV/CIRVA loop:

  • Model snapshot
  • model.parquet with metadata.json describing configuration, seeds, and environment.
  • Event log
  • events.parquet capturing observations and actions, including channel, scope, and delay; see crv.world.events.
  • Agent signals
  • agents_tokens.parquet with step‑stamped per‑agent, per‑object signals and readouts such as value scores and holdings.
  • Relations
  • Tables describing identity and affect structure, for example crv.core.tables.identity_edges, crv.core.tables.holdings, crv.core.tables.holdings_valuation, and other derived relations.

Each file is append‑only within a run and can be joined by time and identifiers for analysis.

Structure and typing

Artifacts mirror the typed contracts in crv.core. A few examples (see module pages for full schemas):

This correspondence allows you to write invariant checks and audits against the same types that implementations use.

IO and manifests

Paths, bucket math, and on‑disk layout are managed in crv.io and crv.io.config. A run’s directory is self‑contained:

  • metadata.json serves as the minimal manifest of configuration used.
  • Data files are placed in predictable subpaths to enable discovery and tooling.

See tests under tests/io/ for contract expectations and path composition.

Determinism and comparison

Because both state and timing are explicit, two runs with the same configuration and seed produce byte‑for‑byte comparable artifacts. This supports:

  • Replay: reproduce a trajectory exactly for inspection and debugging.
  • Ablation: change one assumption at a time and attribute deltas to that change.
  • Audits: assert invariant properties over relations and outcomes.

Practical tips

  • Use uv run crv-app --run <path> to explore a run interactively (see the App guide).
  • Read tables with Polars or Arrow; schemas are stable and typed.
  • For quick sweeps, build an offline valuation policy in crv.lab and reuse it across scenarios; outputs remain comparable.