Notice what worked

Telemetry-Driven Refinement Loop

Close the loop from your own product's behaviour.

Instruments your running output — product feature, published content, model or agent — with behavioural telemetry, attributes each observed outcome back to the configuration that produced it, and routes that signal into the next iteration. The closed attribution loop is what separates this from passive analytics: the system learns from what it just did.

Shape

Operational dimensions

Human supervisor

Person oversees and intervenes by exception.

Continuous

Always running.

High data gravity

Owns a system-of-record; expensive to migrate.

Two-way integration

Reads from and writes to external systems.

Inputs

own-output event / metric stream (clicks, completions, latency, eval scores)
configuration reference per emission (feature variant, model version, prompt id)
outcome definition (target metric, success threshold)

Outputs

attributed performance signal per configuration
refinement recommendation or next-iteration input
trend/drift alert for human supervisor review

Mechanism

Instruments the organisation's own output (product, content, model, agent) with behavioural / performance telemetry, attributes observed outcomes back to specific configuration choices, and feeds the attributed signal into the next iteration of that output.

Why this is a primitive

Cannot be decomposed — the instrument → measure → attribute → route-back-to-next-iteration cycle is one indivisible operation on observed behaviour. The DOMAIN the loop attaches to (product feature, published content asset, model/agent under eval) is a composition variable, not a primitive variable; the machinery (event capture, metric definition, attribution, signal routing back into authoring/training/configuration) is identical across product-analytics, content-performance, and model-eval-and-tune. Collapsing all three into this single primitive removes three duplicate v1 shapes that differed only in the subject of telemetry.

Where it shows up

EdTech platform — content completion rates attributed to lesson format variant, feeding the next authoring cycle with which formats drove mastery

ML team — prompt evaluation traces attributed to prompt version, routing low-scoring traces into the fine-tune dataset for the next model iteration

E-commerce — product recommendation click-through attributed to ranking model version, triggering an automated retraining job when performance dips below threshold

B2B SaaS — feature adoption telemetry attributed to onboarding flow variant, surfacing which flow configuration lifts 30-day retention for the next sprint

Related primitives

Composes with

Classificatory Inference →Access-Controlled Recipient Delivery →Multi-Role Live Operational Surface →Controlled Experimentation →Stated-Feedback Capture →Retrospective Capture →