Compound-skills hypothesis

Emergent

Does a skill library compound in value across rounds?

The emergent workflow runs with no framework but inherits a skill library from previous runs. It's allowed to evolve skills in place, author new ones, and deprecate wrong ones via a CHANGELOG. If the compound-skills hypothesis is real, each emergent round should produce measurably better results than the last as the inherited library accumulates craft. This page is the evidence rollup.

Anchor decisions: gad-65 (CSH pinned), gad-68 (emergent-evolution synthesis), gad-73 (fundamental skills triumvirate).

Emergent runs
6
Playable
5
Scored
3
Latest score

CSH signal — human review across rounds

If skills are compounding, the line goes up-and-to-the-right across rounds. The sixth rubric dimension, skill_inheritance_effectiveness, is the CSH-specific signal (weight 0.20).

v1
2026-04-08
0.100
no 6th dim
v2
2026-04-08
0.500
no 6th dim
v4
2026-04-09
0.885
inherit 0.95

Data provenance: scores read from EVAL_RUNS[n].humanReviewNormalized.aggregate_score, computed at prebuild from each run's rubric submission via gad eval review --rubric .

Skill lineage per run

Every run's skill footprint — what it inherited from the previous run, what new skills it authored, what it deprecated. A healthy CSH signal looks like: inherited count goes up, authored count stays positive, deprecated count is non-zero (the agent is self-correcting), and CHANGELOG dispositions are recorded.

No preserved skill artifacts for this run.

No preserved skill artifacts for this run.

v3
2026-04-09Full breakdown

Skills authored or carried forward this run — click to read

Skills authored or carried forward this run — click to read

Skills authored or carried forward this run — click to read

No preserved skill artifacts for this run.

How emergent differs from bare and GAD

Bare
No framework, no inherited skills. Agent gets AGENTS.md + requirements and builds. Tests the freedom hypothesis directly.
Emergent
No framework, but inherits skills from previous runs. Authors a CHANGELOG documenting disposition (kept / evolved / deprecated / replaced) of each inherited skill. Tests the CSH. See /standards for the SKILL.md format these skills follow.
GAD
Full framework: .planning/ XML, plan/execute/verify/commit loop, skill triggers. Tests whether process discipline pays off despite overhead.
Client debug · NEXT_PUBLIC_CLIENT_DEBUG=1
0 lines

No events yet. Window errors, unhandled rejections, and React render errors appear here. Set NEXT_PUBLIC_CLIENT_DEBUG_CONSOLE=1 to mirror console.error / console.warn.