What problem does it solve? Improving a deployed language model on a specific capability normally requires ad-hoc scripts, unverified training runs, and evals that can be gamed. This Skill provides a disciplined autonomous loop that uplifts a measurable capability on Qwen3.5-4B served by Kiln, chaining SFT, OPD, GRPO, and agentic-GRPO stages into a pipeline where every iteration is gated by a blind eval and a falsifiable hypothesis. ## Core Features & Use Cases - Method routing by evidence: Analyzes per-sub-score headroom and routes each stage to SFT, OPD, GRPO, or agentic-GRPO based on baseline score, teacher availability, and reward variance rather than a-priori commitment. - Blind-eval discipline: Enforces an information firewall so training never sees eval task pools, with rubric calibration sanity checks (good/bad margin > 0.2) required before any GPU work. - Stage pipeline with promotion gates: Reference stage runners, 3-seed paired evals, adapter verification, cross-capability regression checks, and closeout criteria before a pipeline is marked shipped. - Use Case: Ask the agent to "train a JSON-schema-following capability" and it scaffolds the capability directory, writes an adversarially-designed rubric, baselines the model, runs stage-1 SFT, then routes subsequent stages until headroom is exhausted. ## Quick Start Ask the agent to train a new capability on the kiln-served model, giving it a plain-English description of the capability and confirming a blind eval oracle exists.
