Targeted interventions, not generic training.
Every diagnosis points at a structural weakness. Every intervention declares a target metric, a measurement window, and an expected internal-metric delta. We test the repair, then re-measure.
See the re-measurement loop → Start from diagnosisFive classes of intervention.
Interventions are not a single bucket. They fall into five distinct classes, each with its own evidence base, ownership, and expected effect surface.
Guides
Written operator guidance targeting specific structural weaknesses.
- Context construction
- Task decomposition
- Project persistence
- Iterative workflow
- Verification
- Model selection
- Agent orchestration
Tools
Systems and integrations that change what the operator can do.
- MCP servers
- Memory systems
- Project templates
- IDE tools
- Routing systems
- Context managers
- Specialized agents
Workflow changes
Structural changes to how AI work fits into the surrounding process.
- Stage separation
- Context handoff
- Human checkpoints
- Persistent project state
- Review structures
Training
Structured learning programs, internal or external.
- Internal guides
- Workshops
- Role-specific programs
- External courses
- Manager coaching
Human support
People-side support that complements the technical interventions.
- Operator review
- One-on-one coaching
- Team redesign
- Specialist consultation
Match to diagnosis
Class is chosen by matching the diagnostic hypothesis to the intervention surface. A context-construction weakness suggests a Guide. A persistence weakness may suggest a Tool. A verification gap may suggest a Workflow change. The diagnosis drives the class — not the other way around.
Every intervention is a declared, measurable object.
No intervention ships without a recorded object. The object carries the diagnosis it responds to, the metric it targets, the window it runs in, and the owner accountable for it.
The object is the contract between diagnosis and verification. Without it, there is no re-measurement — only anecdote.
Baseline → intervene → re-measure → classify.
The loop is the unit of evidence. Each pass produces a result classification, not a verdict on the operator.
BASELINE
Operator metrics recorded during the evaluation phase. This is the reference point for every delta.
INTERVENTION
Declared intervention object applied. Class, target metric, window, and owner recorded.
POST WINDOW
After the declared measurement window, telemetry is collected for the same operator over an equivalent span.
INTERNAL METRIC DELTA
Target and non-target internal metrics compared baseline vs post. Both are reported — not only the target.
OPTIONAL EXTERNAL OUTCOME DELTA
If an external outcome was declared, it is joined and compared. Always ASSOCIATION, never CAUSATION.
RESULT CLASSIFICATION
Improved / unchanged / degraded / inconclusive. The classification is about the intervention, not the person.
What interventions are NOT.
The intervention system is developmental. These guardrails are not soft preferences — they are structural constraints on how the product may be used.
Interventions exist to improve operator structure. They are never used to rank, penalize, or compare employees for adverse purposes.
Intervention tracking records the intervention object and the metric delta. It does not monitor employee behavior, content, or conduct.
No intervention, metric, or classification triggers automated employment decisions. Adverse action is explicitly excluded from the pilot terms.
Every intervention declares its permitted use.
Decision-use labels govern who may act on an intervention result and for what purpose. They are set at creation and enforced by the pilot terms.
For developmental use
The default label. Results may be used to guide the operator's growth, refine the intervention, and inform coaching. Visible to the operator and the pilot team.
For workflow experiments
Results may inform workflow design changes — stage separation, handoff structure, review checkpoints. Treated as experiments, not performance verdicts.
For research
Results may be aggregated and used for methodology research, archetype refinement, and field comparison. Individual identity is governed separately.
Elevated governance — not in pilot
Any personnel-related use requires elevated governance, explicit consent, and review structures. This label is not available during the pilot. It exists to define the boundary, not to cross it.
The PERSONNEL label is defined so the product can name what it will not do in the pilot, rather than leaving the boundary implicit.
Once the intervention runs, we re-measure.
The verify phase takes the intervention object and produces the result classification — internal metric delta first, external outcome separately.
Go to Verify →