v1.89.19

Fictionarry/TalkingGaussianv1.89.19Jun 29, 2026by autogame-17

AI Summary

Enhances the Skill-to-Gene reverse-distillation pipeline with provenance awareness and quality controls to prioritize real execution insights and prevent data leakage.

Key Highlights

Provenance-aware synthesis prioritizes corrective insights from solve→fail→fix→pass trajectories.
Quality gate downgrades low-signal transcription-only Genes and flags them.
Leakage audit removes task-specific literals from hidden output before publishing.
Host-driven distillation allows the host agent to supply insights inline.

New Features

Provenance tracking
Quality scoring and flagging
Literal leakage removal
Host agent integration

Full Release Notes

## What's new in v1.89.19

### skill2gep: evolve-aware Gene distillation

The Skill → Gene reverse-distillation pipeline now distinguishes **how** a Gene was produced and prioritizes the lessons learned from real execution:

- **Provenance-aware synthesis.** Genes distilled from a verified solve→fail→fix→pass trajectory now lead with the corrective insight that actually flipped the outcome, and turn the failures they overcame into explicit preconditions. Genes with weaker provenance are scored and flagged accordingly.
- **Quality gate.** Low-signal transcription-only Genes are downgraded to Gene-only (no success Capsule) and flagged, configurable via `SKILL2GEP_MIN_QUALITY`.
- **Leakage audit.** A mechanical audit removes task-specific literals that appear only in hidden execution output before a Gene is published, while preserving the public contract a fresh solver is allowed to see.
- **Host-driven distillation.** The host agent can supply the corrective insight inline; no extra subprocess or network call is made.

Grounded in the TaskGenome Bench evaluation (EvoMap/Gene-Bench).

Internal improvements and stability enhancements.