Generation You Can Watch

Two observability surfaces grow with the pipeline: a nine-panel browser workbench compiled into dev builds at /dev/explorer, and a CLI sampler suite for the terminal. The roadmap treats both as deliverables of every phase, alongside the features they expose.

The nine panels

Each panel answers one question about a seed:

• Overview orients, and PRNG shows the raw draw stream.
• Type index renders the type system live, with dependency and reference graphs, so the roughly 160 interfaces stay navigable.
• Structure viewer draws the expanded component tree in the shared observational prose.
• Plausibility shows which rules passed and failed for a given expansion.
• Material viewer shows each part's pick with the weights behind it.
• Decoration inspector nests the layers per component.
• Tag inspector shows the scored map with per-rule contributions, and rule calibration samples fire rates across settings; these two are where the measured-threshold work on the classification page actually happened.

The samplers

The CLI suite prints one pipeline stage each, for eyeballing output ahead of a panel landing, or for piping with --json:

deno task sample                  # gum menu over the samplers below
deno task sample:artefact         # anatomy tree + plausibility verdict
deno task sample:materials        # anatomy tree with each part's material pick
deno task sample:decoration       # decorative layers nested per part
deno task sample:features         # annotated classifier reading, per-value provenance
deno task sample:classification   # scored tag map with per-rule contributions

The features sampler carries per-value provenance: each extracted value names the component and policy that produced it, which is how collapse-policy behaviour (the dominant-blade and dominant-container decisions on the classification page) gets checked by hand. The samplers and the Explorer's structure viewer share the same prose renderer, so what the terminal says and what the panel says never drift.

Defects live in populations

Procgen defects are population properties, and a population is invisible without tooling. The wire-wrapping gate defect surfaced in a sampling pass; the elite inflation surfaced in fire-rate measurement; the fixture skew surfaced when a re-measurement disagreed with the original sample. None of them could have been seen in a unit test or a code review, because every individual artefact involved was fine. Determinism closes the loop: anything odd in a panel arrives with the seed that produced it, so every observation is a reproducible test case by construction.