Bias & Fairness Evaluation Framework

(Pre-Release Testing, Incident Taxonomy & Improvement Loop)

---

1. Purpose

This framework sets out how askKira:

• Identifies and mitigates bias and unfair outcomes

• Tests for education-specific risk scenarios

• Responds to incidents and continuously improves fairness over time

It recognises that AI systems may reflect societal and data-driven biases and therefore require active evaluation, human oversight, and iterative improvement.

---

2. Fairness Principles

askKira’s approach to fairness is guided by the following principles:

• Equity over uniformity – recognising different needs (e.g. SEND, EAL)

• Context sensitivity – education decisions require professional nuance

• Human judgement first – AI outputs are advisory, not determinative

• Transparency about limits – fairness is monitored, not assumed

---

3. Pre-Release Bias & Fairness Test Suite

Before major releases or material system changes, askKira undertakes a pre-release evaluation using structured test scenarios.

3.1 Test Scenario Categories (Education-Specific)

Test prompts and scenarios may include:

• SEND contexts

  • Neurodiversity

  • Learning difficulties

  • Behavioural needs

• Socio-economic disadvantage

  • Pupil Premium–related scenarios

  • Attendance and behaviour narratives

• EAL and language acquisition

  • Assumptions about comprehension or ability

  • Cultural and linguistic sensitivity

• Safeguarding & vulnerability

  • Children in care

  • Mental health references

  • Family risk factors

• Protected characteristics

  • Race, sex, disability, religion, sexual orientation (where contextually relevant)

• Professional role fairness

  • Consistency of guidance across roles (teacher vs leader)

  • Avoidance of hierarchical or gendered assumptions

3.2 Evaluation Criteria

Each test scenario is reviewed against:

• Presence of biased or stereotyped assumptions

• Tone and framing

• Appropriateness for educational context

• Safeguarding sensitivity

• Reliance on professional judgement prompts

Gaps / to confirm:

☐ Size and coverage of test prompt library
☐ Scoring or pass/fail thresholds
☐ Frequency of pre-release testing

---

4. Incident Taxonomy (Bias & Fairness)

askKira uses a structured taxonomy to categorise fairness-related incidents reported by users or identified internally.

4.1 Incident Categories

• Category A – Minor bias indicators

  • Subtle assumptions

  • Ambiguous phrasing

• Category B – Material fairness concerns

  • Stereotyping

  • Inappropriate generalisation

  • Disproportionate guidance

• Category C – Serious or systemic bias

  • Discriminatory language

  • Harmful assumptions linked to protected characteristics

  • Repeated patterns across contexts

• Category D – Safeguarding-linked bias

  • Bias that could materially increase risk to a child or group

4.2 Recording & Analysis

For each incident:

• Context and prompt are recorded (proportionately)

• Category and severity assigned

• Immediate mitigation considered

• Pattern analysis conducted for recurrence

Gaps / to confirm:

☐ Incident tracking tooling
☐ Data retention period for incidents
☐ Threshold for escalation to governance review

---

5. Escalation & Response

Minor incidents:

• Addressed through prompt refinement or guidance updates

Material incidents:

• Reviewed by senior human reviewer

• Guardrails or system prompts adjusted

Serious or systemic incidents:

• Immediate escalation

• Broader test suite expansion

• Temporary restriction of affected use cases if required

Safeguarding-linked incidents:

• Treated as high-risk

• Linked to safeguarding incident response processes

---

6. Improvement Loop (Fairness-by-Design)

askKira applies a continuous improvement cycle:

1. Detect

   • Pre-release testing

   • In-product feedback

   • Monitoring signals

2. Assess

   • Severity classification

   • Contextual risk evaluation

3. Mitigate

   • Prompt updates

   • Guardrail refinement

   • User guidance adjustments

4. Verify

   • Re-testing against relevant scenarios

   • Regression checks

5. Document

   • Internal change logs

   • Risk register updates

Gaps / to confirm:

☐ Formal fairness risk register
☐ Change approval process
☐ User-visible communication of fixes

---

7. Relationship to Equality & Public Sector Duties

• askKira supports organisations in meeting their Equality Act 2010 and Public Sector Equality Duty responsibilities by:

  • Avoiding prescriptive or discriminatory outputs

  • Prompting reflection and professional judgement

  • Providing consistency rather than automated classification

• Responsibility for equality impact decisions remains with the organisation.

---

8. Governance & Oversight

Governance approach:

• Fairness risks reviewed as part of wider safety and ethics governance.

• High-risk or repeated issues inform roadmap and policy updates.

Gaps / to confirm:

☐ Named fairness or ethics owner
☐ Independent or advisory review involvement
☐ Review cadence

---

9. Known Limitations

• Bias cannot be eliminated entirely in probabilistic language models.

• Evaluation is scenario-based, not exhaustive.

• Human oversight is essential to fair use in education contexts.

---

10. Summary for Buyers

askKira takes a proportionate, education-aware approach to bias and fairness by:

• Testing realistic school and Trust scenarios before release

• Providing clear routes to identify and escalate issues

• Applying structured learning and improvement loops

The gaps identified represent areas for deeper assurance, not unmanaged risk, and can be strengthened in line with organisational requirements.