Measure what changes. Test business impact separately.
Re-measurement is the unit of evidence in the closed loop. We measure operator change directly. We test business impact separately. The two are never conflated.
See the intervention system → Enterprise privacy modelThe loop is the unit of evidence.
Every intervention produces one pass through this loop. The output is a result classification — not a verdict on the operator, and not a claim about business performance.
BASELINE
Operator metrics from the evaluation phase. The reference for every delta.
INTERVENTION
Declared intervention object applied. Target metric and window recorded.
POST WINDOW
Telemetry collected over an equivalent span after the window closes.
INTERNAL METRIC DELTA
Target and non-target internal metrics compared. Both reported.
OPTIONAL EXTERNAL OUTCOME DELTA
If declared, external outcome joined and compared. ASSOCIATION only.
RESULT CLASSIFICATION
Improved / unchanged / degraded / inconclusive. About the intervention, not the person.
Seven possible outcomes. No spin.
A result is classified honestly. "Unchanged" and "inconclusive" are real outcomes — not failures to be reframed.
Improved internally
Target internal metric moved in the expected direction beyond noise threshold.
Unchanged internally
Target internal metric did not move beyond noise threshold. The intervention did not produce the expected structural change.
Degraded internally
Target internal metric moved against the expected direction. Worth investigating — possibly a wrong diagnosis or a harmful intervention.
External outcome improved
Joined external outcome moved favorably. ASSOCIATION with the internal change — not causation.
External outcome unchanged
Joined external outcome did not move. Internal improvement may not have reached the business surface — or the outcome measure was wrong.
External outcome degraded
Joined external outcome moved unfavorably. Flagged for review. Still ASSOCIATION.
Inconclusive
Insufficient post-window data, noisy telemetry, or confounding changes. No classification can be made. The loop is recorded but not interpreted.
Operator 031 — context construction intervention.
A worked example. Operator 031 received a context construction guide (intervention_014). After a 14-day window, the same metrics were re-measured. Target and non-target deltas are both shown.
| Metric | Baseline | Post-14d | Delta |
|---|---|---|---|
| Leverage | 31.2x | 43.2x | +38% |
| Yield | 6.1 | 8.7 | +43% |
| Token SNR | 0.28 | 0.34 | +21% |
| Construction | 1.4 | 1.6 | +14% |
| Context form | structured | composite | +1 grade |
| Stability (non-target) | 0.71 | 0.72 | +1% |
| Trajectory (non-target) | ascending | ascending | unchanged |
Target metrics (construction, context form) moved as expected. Non-target metrics (stability, trajectory) were monitored but not expected to change — confirming the intervention was specific, not a general Hawthorne effect. Synthetic demo data.
Business impact is tested separately — and optionally.
External joins are optional and company-specific. They are never automatic, never default, and never treated as proof that the internal change caused the business change.
Engineering
- Cycle time
- PR throughput
- Review time
- Defects
- Rework
- Deployment frequency
- Rollback rate
Sales
- Pipeline
- Win rate
- Cycle length
- Revenue
- Conversion
Support
- Tickets handled
- First response time
- Resolution time
- Escalation
- Reopen rate
- CSAT
Operations
- Task throughput
- Completion time
- Error rate
- Blockers
- Handoffs
- Rework
Each category is a candidate join, not a default. The company selects which outcomes to join, for which cohort, and under what governance. The join is declared in the intervention object before the window opens.
What we do NOT claim.
We measure operator change directly. We test business impact separately.
Internal change ≠ business performance
The product must not treat internal metric change as automatic business performance change. An operator who improves context construction has improved context construction — not necessarily shipped faster.
Association ≠ causation
Even when an external outcome moves alongside an internal change, the relationship is association. Confounders, seasonality, and parallel initiatives are not controlled. The label is honest about this.
All outcome joins are ASSOCIATION, not CAUSATION. This is a structural claim about the method, not a disclaimer added after the fact.
Diagnose. Intervene. Re-measure. Classify.
The verify phase closes the loop the develop phase opened. The result classification feeds back into diagnosis — refining the hypotheses for the next pass.
Back to Develop → Enterprise governance →