Full-Repo Project Build

Whole-repository builds, signal vectors, and visual targets.

Retained generated outputs and their historical evaluators, not a matched or currently qualified experiment. Failures, incomplete coverage, replay scoring and different model/backend bindings remain explicit. A live flag does not establish containment or independent replication.

One brief and five harnesses, with in-place historical rescoring. Compare per-harness values: the saved advantage card uses the first Cortex role, which is not necessarily the same base provider.

Highlights & plots Dataset · 1 items Experiments · 20 runs Original report ↗All result tables ↗

Models & settings — click to expand

Signal vector by harness

Separate objective, similarity and judge signals; the saved composite is one dimension, not a new sum of these bars

Composite by harness

build-gated weighted sum of all signals

VERTEX trajectory similarity

capability + architecture cross-similarity vs the hidden reference (baseline-normalized)

Build prompt (what the harness receives)

Reference frames (visual target)

Reference frames requested by the brief. Unavailable retained image files are marked below.

Captured app renders (per harness)

Retained app captures per harness. Images are evidence to inspect, not a visual ranking.

Build explorer

Per harness: the full signal vector + the captured repo.

Harness	Build	Functional	VERTEX	Visual	Code/Arch	Security	PWA/a11y	Robust	Composite

Full-Repo Project Build

Signal vector by harness

Composite by harness

VERTEX trajectory similarity

Build prompt (what the harness receives)

Reference frames (visual target)

Captured app renders (per harness)

Build explorer

Inside the dataset

Explore experiments