singularity

Invoke the skill once per cycle. It orients via git status and memory files, picks one instruction improvement and one crawler or UI improvement, applies an adversarial quality gate, verifies with screenshot and end-to-end harnesses, then commits and records both deltas.

The agent may create, revise, or retire its own skills, memory files, verification harnesses, docs, and continuation prompts, each versioned and justified by measurable gain. It may not edit owner hard rules, safety invariants, verification gates, or the never-git-add list.

Verification uses a ui-screenshot harness read as a user, real NewsDatabase end-to-end tests for logic, adversarial multi-lens review for correctness-sensitive changes, and the project's normal test suite. Work ships only when the strongest self-critique finds no real defect.

No. It is a bounded, verification-gated loop at the agent-system level covering scaffold artifacts like skills, memory, and tests. It explicitly does not edit model weights, does not relax its own constraints, and requires a measurable delta per cycle.

On a detected model swap, the skill appends a row to the lineage table in docs/agi/SELF_MODEL.md and recalibrates: owner rules carry unconditionally, empirical heuristics get re-probed before trust, and model-specific behavioral compensations are re-tested or retired.

The skill tracks cost-per-improvement trend, count of second-order tools produced, rubric first-pass rate, defects caught pre-ship versus post-ship, and live raw-SQL sites in the coordination repo. Flattening returns trigger a shift toward portable artifacts for successor models.