Learning evidence, feedback, and evaluations

Learning collects proposed improvements and evidence about how Gideon works. Review what is being changed and what the evidence actually establishes.

Review capture and proposals

Open Learning for the capture week, identity report, memory health, attention information, and proposal queue. Filter by proposal kind. Read the change and evidence before accepting or rejecting it. Flagged proposals remain visible because a malformed change should not disappear silently.

Skill proposals may live in their own Skills review queue. An empty main proposal list does not prove there are no pending skills to inspect.

Give useful feedback

Thumbs-up and thumbs-down on judgments are attributed to the producing prompt, judge, or rule. Settings → AI feedback shows these sources. A low-performing source can generate a retire proposal; only source kinds with a real surfacing gate are automatically prevented from surfacing. Read the displayed status rather than assuming every bad source has stopped running.

Evaluation controls

Settings → Evaluations controls the evaluation substrate and defaults for paired studies, judge agreement, budgets, and ablation cadence. Each registered study still has its own parameters. A result of judge-unreliable is not a winning template.

Operator reports

Self-hosted Learning exposes additional benchmark, retrieval, judge, study, ablation, and field-metric reports. Hosted Learning does not expose that operator report collection. Use the reports your environment actually offers.

Compare evidence

Read sample size, test conditions, spend, and limitations before adopting an improvement. A skill being saved or accepted is not a measured performance gain. Ask for a representative follow-up task and inspect its outcome.