Forecast-band anomaly enrichment example (TimesFM)

Summary

Ship a runnable example where an out-of-band TimesFM producer emits per-rule quantile bands loaded as a dynamic source, and a lookup enricher annotates firings with expected bands.

Motivation

Operators want a glass-box 'is this louder than usual' signal without putting ML in the hot path. Composing shipped enrichment and dynamic sources keeps the daemon a cache lookup.

Proposed approach

• Python producer with file/prometheus/custom series adapters.
• Daemon loads bands via dynamic source; enrich_rule_forecast lookup recipe.
• example under examples/forecast-anomaly/ plus guide docs.

Out of scope

• Core-engine ML models.
• Synthesizing alerts with no rule firing.

Tasks

• Example producer + golden band table
• sources/enrichers recipes
• Guide page and enricher catalogue entry

References

• Post-evaluation enrichment
• Detached dynamic sources
• Per-rule match metrics

Start by inspecting the existing examples and recipes for post-evaluation enrichment and detached dynamic sources, then create the proposed example under examples/forecast-anomaly/. The work is complete when the producer and golden band table, source and enricher recipes, guide page, and enricher catalogue entry are all runnable and documented.