Reasoning Calibration Governance

Purpose

Calibration governance lets SENCOs and clinicians review how well the reasoning engine maps observations to behaviours, domains, potential needs, and escalation decisions.

Where to go

1. Sign in as an owner, SENCO, or clinician.
2. Open /dashboard/settings.
3. Find reasoning calibration analytics.
4. Find calibration governance below it.

Add calibration feedback

1. Open a student report history card.
2. Use the calibration feedback form.
3. Choose target type:

• Behaviour.
• Domain.
• Need.
• Escalation.

1. Enter target key.
2. Add a rating from 1 to 5.
3. Optionally suggest a scoring weight.
4. Add a comment.
5. Submit feedback.

Review calibration analytics

The settings analytics panel shows:

1. Feedback record count.
2. Reviewer count.
3. Average rating.
4. Reviewed observation comparison totals.
5. Proposed scoring-weight updates.
6. Extraction provider health.

Export calibration materials

Use settings export actions to download:

1. Reviewed-versus-initial extraction comparison CSV.
2. Calibration meeting summary markdown.
3. Calibration release comparison CSV showing baseline weights, latest active release weights, current runtime weights, and whether released weights are currently applied.

Refresh proposals

1. Open calibration governance.
2. Click refresh proposals.
3. senco converts proposed scoring-weight updates into persisted proposal records.

Approve or reject proposals

1. Review the target type, target key, current weight, proposed weight, feedback count, and rationale.
2. Click approve to allow release packaging.
3. Click reject if the proposal should not move forward.
4. Approval and rejection are audited.

Create a calibration release

1. Find an approved proposal.
2. Review or edit release notes.
3. Click create release.
4. A versioned release is created.
5. The proposal moves to released state.

Apply released weights at runtime

Released weights are safely separated from baseline scoring until runtime loading is explicitly enabled.

1. Confirm the release is active in /dashboard/settings.
2. Set CALIBRATION_RELEASE_RUNTIME_ENABLED=true in the deployed environment when the organisation is ready to use released scoring weights.
3. Student-linked reasoning then loads the latest active release for the signed-in organisation.
4. Domain proposals such as SENSORY_PROCESSING are applied directly.
5. Explicit need-domain proposals such as ASC_RELATED_PROFILE:SENSORY_PROCESSING are applied to that need weighting.
6. Ambiguous behaviour or escalation proposals remain in the governance history but are not applied to scoring until a precise runtime mapping exists.
7. Public analyser use and unflagged environments continue to use the baseline calibration.

Compare release weights

1. Open /dashboard/settings.
2. Find reasoning calibration analytics.
3. Click Release comparison.
4. Download the CSV for a non-diagnostic governance view of:

• baseline scoring weights
• latest active release weights
• current runtime weights
• whether CALIBRATION_RELEASE_RUNTIME_ENABLED is applying released weights
• proposals that remain governance-only because they do not yet map to runtime scoring

1. Use the export in SENCO/clinician calibration meetings before changing rollout scope.

Roll back a release

1. Find an active release in release history.
2. Enter a rollback reason.
3. Click roll back release.
4. The release is marked rolled back.
5. Linked proposals return to approved status for further professional review.

Provider health and safe probes

1. Review provider status in reasoning calibration.
2. Click run safe probe in calibration governance.
3. If the LLM provider is configured, senco sends a synthetic non-pupil observation.
4. If not configured, the probe is skipped.
5. Probe results are audited.

Important safety note

Calibration changes tune non-diagnostic reasoning support. They do not create diagnostic labels or replace professional judgement.