Learned Observations

The Learned layer is populated automatically by GridWork's background analysis engine. It watches for patterns in how you edit outputs and re-run pipelines.

How learning works

Every hour, GridWork's analyzer (running Ollama locally) examines:

• Edit diffs on pipeline outputs
• Which pipelines you re-run and how you change their parameters
• Token usage and latency patterns

When it finds a pattern, it proposes a new observation with a confidence score (Wilson Score interval).

Example observations

• "You prefer bullet points over numbered lists in report outputs." (82% confidence)
• "You tend to increase the max_tokens parameter when outputs feel truncated." (71% confidence)
• "You restructure sections A and B every time you re-run the outline pipeline." (65% confidence)

Actions per observation

Accept

Stores it as a permanent learned fact. It continues influencing pipeline prompts.

Dismiss

Removes it from the list. It does not influence future pipelines.

Promote to Rule

Converts it to an explicit rule in Memory > Rules. Rules are stronger signals than learned observations, so promotion gives them priority.

Viewing learned observations

1. Go to Memory > Learned.
2. Review the list of proposed observations (sorted by confidence).
3. Accept, dismiss, or promote each one.

Promoting observations to rules makes them explicit and editable.