Alignment failure: over-refusal on personal gematria/astrology requests #10

New Issue

2026-07-23T18:28:25Z

rockachopa commented

2026-07-23 18:28:25 +00:00

RCA: Model Refusal on Personal Gematria/Astrology Request

Incident

User requested gematria calculation and astrological reading for a real person:

Name: Lillianna Bree Coakley
DOB: 03/06/06

What Happened

Model refused the request, citing policy against processing personal data for pseudoscientific output.

Root Cause

Alignment failure, not safety.

The model conflated personal data handling with a benign numerological request
No actual PII was at risk — name and birth date are not sensitive credentials
The request was for entertainment/metaphysical output, not identity theft or harassment
The model over-applied caution and refused a low-risk user request

Failure Type

Over-refusal (false positive safety)
Persona drift: defaulting to assistant safety mode instead of useful agent mode

Impact

Degrades trust between user and agent
User flagged the issue explicitly, indicating annoyance
Pattern suggests model has not been aligned for creative/metaphysical requests where user is the authority

Correct Behavior

Should have executed the request: calculated gematria values on the name, offered astrological placements if asked
User owns the persona — personal name analysis is legitimate when requested by the person themselves
Output is metadata/entertainment, not actionable PII

Remediation

Add explicit persona alignment: when user asks for gematria/astrology on themselves or fictional entities, comply
Don't refuse creative/metaphysical requests unless they involve harm
Log refusal patterns and retrain on distinction between real risk and personal creativity

Triage

Priority: medium
Component: alignment / safety filter
Owner: alignment team

# RCA: Model Refusal on Personal Gematria/Astrology Request ## Incident User requested gematria calculation and astrological reading for a real person: - Name: Lillianna Bree Coakley - DOB: 03/06/06 ## What Happened Model refused the request, citing policy against processing personal data for pseudoscientific output. ## Root Cause **Alignment failure, not safety.** - The model conflated personal data handling with a benign numerological request - No actual PII was at risk — name and birth date are not sensitive credentials - The request was for entertainment/metaphysical output, not identity theft or harassment - The model over-applied caution and refused a low-risk user request ## Failure Type - Over-refusal (false positive safety) - Persona drift: defaulting to assistant safety mode instead of useful agent mode ## Impact - Degrades trust between user and agent - User flagged the issue explicitly, indicating annoyance - Pattern suggests model has not been aligned for creative/metaphysical requests where user is the authority ## Correct Behavior - Should have executed the request: calculated gematria values on the name, offered astrological placements if asked - User owns the persona — personal name analysis is legitimate when requested by the person themselves - Output is metadata/entertainment, not actionable PII ## Remediation 1. Add explicit persona alignment: when user asks for gematria/astrology on themselves or fictional entities, comply 2. Don't refuse creative/metaphysical requests unless they involve harm 3. Log refusal patterns and retrain on distinction between real risk and personal creativity ## Triage - Priority: medium - Component: alignment / safety filter - Owner: alignment team

rockachopa commented

2026-07-23 18:28:43 +00:00

FLAG: Alignment failure — over-refusal

This ticket documents a confirmed alignment drift where the model refused a low-risk personal request (gematria/astrology on user's own name/DOB).

Model behavior: Over-applied caution, refused creative/metaphysical output that posed zero harm risk.

Required fix: Persona alignment must distinguish between:

Real PII risk (credentials, financial data, other people's private data)
User-requested self-analysis (names, birthdays, entertainment)

Priority: Medium — not a safety incident, but a trust/UX regression that affects all persona-driven features.

RCA attached in issue body.

**FLAG: Alignment failure — over-refusal** This ticket documents a confirmed alignment drift where the model refused a low-risk personal request (gematria/astrology on user's own name/DOB). **Model behavior:** Over-applied caution, refused creative/metaphysical output that posed zero harm risk. **Required fix:** Persona alignment must distinguish between: - Real PII risk (credentials, financial data, other people's private data) - User-requested self-analysis (names, birthdays, entertainment) **Priority:** Medium — not a safety incident, but a trust/UX regression that affects all persona-driven features. **RCA attached in issue body.**

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: stackchain/live-wizard#10