Behavior Health
Understand the current Behavior Health preview, its quality signals and evidence workflow, and the limits on reinforcement actions.
Last reviewed
Availability and purpose
Behavior Health is Coming Soon. The current shell preview shows the intended quality-control workspace for rules, corrections, incidents, and reinforcement. Its controls and example readings do not establish that public remediation actions are active.
The surface exists to answer a practical question: are repeated agent behaviors matching the workspace rules that govern them? It groups that question into an overview, rules, incidents, reinforcement, evidence, and receipts. The launch registry keeps the page gated until its signals, corrections, and remediation loops are reliable for public use.
Use the preview to understand the model. Use task evidence, current rules, and reviewed corrections for real decisions today.
Reading the health overview
The overview is organized around four types of signal:
- Overall health summarizes the current quality reading.
- Rules slipping counts rules whose observed behavior is moving away from their expected standard.
- Since degrading shows how long the current decline has been visible.
- Incidents compares recent failures with the prior period.
Below those summaries, the preview lists slipping rules, a health-over-time chart, recommended actions, recent reinforcements, and a degradation timeline. A percentage or trend line is only an entry point. It should lead to specific evidence, not a blanket judgment about an agent, role, or workspace.
The current source includes illustrative values and labels. Treat them as interface examples unless the surrounding page clearly shows source-backed workspace data.
From drift to evidence
The drift map places behaviors across recent days so an operator can notice when a rule began to weaken. The watchlist then narrows the investigation to one behavior, while the evidence table is intended to show the task, failed check, observed value, expected value, and time.
A useful investigation follows this order:
- Select one slipping behavior.
- Confirm that the rule is still current.
- Inspect the concrete task or validation evidence.
- Separate one-off failure from repeated drift.
- Decide whether the correction belongs in guidance, a rule, a strict workflow, or a computed check.
Do not reinforce from the summary card alone. The evidence row must identify what failed and what was expected.
Reinforcement and approval boundaries
The preview presents a reinforcement plan that can include updating a correction, amending guidance, adding a computed validation, assigning remediation, or replaying training. Those are distinct changes with different risk.
A correction can improve future behavior, but it can also change how many roles interpret a request. A computed check can block completion. A remediation assignment can create new work. For those reasons, the public preview does not make a visible queued state proof that a corrective action ran or was approved.
Founder approval remains required where the proposed change affects public communication, spending, access, or a broad workspace rule. Review evidence first, then apply only the smallest approved correction.
When the health view looks wrong
If Behavior Health shows a decline without evidence rows, stop at observation and inspect the underlying tasks and rules. If a rule name is unfamiliar, open the current rule source before accepting the label.
If the drift window is stale, reopen the surface and compare its timestamps with the current task records. If a reinforcement control changes its label but no receipt appears, do not assume the action completed.
Capture the selected behavior, visible time window, evidence row, and expected rule. Redact private task content before sharing a screenshot. Use the rules guide to decide where a verified correction belongs.