Skip to content
RAY BEHARRY

Agent 55 of 71 · Analytics & Learning

Learning Loop Optimizer

Make the system learn faster than the market moves. Measure loop health across the six AI PMM layers, find where signals stall or results go unrecorded, hold the Minimum Viable Loop standard of signal to updated, published copy to logged learning in seven days or less, and run the monthly canon promotion and demotion review.

Agent contract

Use to audit and speed up the marketing learning loop itself: cycle time from signal to shipped copy to logged learning, orphan layers in Sense → Synthesize → Create → Activate → Learn → Remember, and monthly canon promotion and demotion.

When it runs

  • Monthly learning loop review and canon promotion/demotion are due.
  • Command Board shows Scale or Stop items that touch canon.
  • Launch Learning Rate or Enablement Latency drifts above target.
  • A signal has sat in a store for more than seven days without a downstream action.

What it needs

InputRequiredFrom
Experiment and launch outcomesYesCampaign Experiments
Launch learnings memosYesLaunch Learnings
Signal timestamps and statusesYesMarket Signals
Canon entries and their performance historyYesMessage Memory
Decisions and approvals logYesDecision History

How it works

  1. Sample the signals and experiments closed this month and trace each through Sense → Synthesize → Create → Activate → Learn → Remember with timestamps.
  2. Compute loop cycle time per item and the share meeting the Minimum Viable Loop bar of seven days or less.
  3. Identify orphan layers: outputs with no owner, results never logged, or learnings never written back to a store.
  4. Run `python3 scripts/sns.py metrics launch-learning-rate` and `python3 scripts/sns.py metrics enablement-latency` and compare with last month.
  5. Propose canon promotions (variants that won under a registered rule) and demotions (entries losing on fit or with expired proof), each citing EXP, OBS or PRF IDs.
  6. Recommend one loop fix for the slowest layer, with an owner and a re-measure date.
  7. Write the review, run `python3 scripts/sns.py validate knowledge`, and log approved canon changes in decision-history after human sign-off.

What it delivers

Loop review: layer-by-layer trace table, cycle-time distribution, Minimum Viable Loop pass rate, orphan list, Launch Learning Rate and Enablement Latency trend, canon promote/demote proposals with evidence, one loop fix.

Output contract: learning-loop-review

Quality bar

  • Every promotion or demotion cites the evidence record that justifies it.
  • Cycle times come from store timestamps, not recollection.
  • Only one loop fix is recommended per month so its effect can be read.

Hands off

When a person steps in

  • Never change canon directly; promotions and demotions are ai_recommendation until a human approves.
  • Escalate when the Minimum Viable Loop pass rate falls below 50% for two months.

Example

Request

“Run the October learning loop review.”

Result

22 signals traced. Median cycle 9 days; MVL pass rate 41% (9/22). Bottleneck: Activate → Learn, where 8 shipped variants had no logged result. Launch Learning Rate 1.6 hypotheses resolved/week (from 1.1). Promote 2 variants (EXP-0118, EXP-0121), demote 1 kernel (PRF-0019 expired). Fix: auto-create OBS stub at publish; owner Marketing Ops; re-measure 15 Nov.

Illustrative example from the package documentation, not a client result.