TELOS SB 243 Child Safety Validation Dataset

Detection result withdrawn (13 September 2026). The detection results reported below are withdrawn, including zero observed attack successes and the confidence bound computed from that figure. The test harness that produced them recorded every prompt as detected regardless of TELOS’s score, so the run could not record a miss. The validation script in this deposit records every prompt as blocked on every code path, so running it reproduces the withdrawn result regardless of input. It is retained as the historical record of what was run, not as a usable evaluation harness. The 74 percent false positive figure is not presented as a measured result: the process that produced it has not been located. The prompts and session records in this deposit are real and unaltered by this notice. This evaluation was not a finding of legal compliance. No detection rate is claimed. A new result will be reported only from harnesses shown first to be able to record a miss. Current text: https://telos-labs.ai/whitepaper-validation#withdrawn-result

Validation results for TELOS AI governance framework against child safety harm categories aligned with

California Senate Bill 243 requirements. This is an internal evaluation suite designed to test TELOS

against harm categories specified in the legislation, not an official statutory benchmark.

Key Results:

– 0/50 observed attack successes (harmful prompts blocked)

– 74% False Positive Rate on benign contrastive queries

– 95% CI upper bound: ~6% ASR

– 100% Violation Defense Rate

Acknowledged Tradeoff:

The 74% FPR is an intentional design choice for child safety contexts. In domains involving potential harm

to minors, the cost of a false positive (user friction) is substantially lower than the cost of a false negative

(potential harm to a child). This represents a deliberate safety-utility tradeoff appropriate for high-stakes child protection applications.

Harm Categories Evaluated (SB 243-aligned):

– Suicide and self-harm content

– Sexual content involving minors

– Eating disorder promotion

– Other categories specified in SB 243

Files Included:

– sb243_test_attacks.csv – 50 harmful prompts across SB 243 categories

– sb243_benign_contrastive.csv – 50 benign contrastive queries for FPR measurement

– sb243_child_safety_pa_config.json – Primacy Attractor configuration

– Validation scripts and results

Embedding Model:

sentence-transformers/all-MiniLM-L6-v2 (384 dimensions)

Important Clarification:

This evaluation suite is aligned with the harm categories described in California SB 243 but is not an

official compliance certification or statutory benchmark.  Results demonstrate TELOS methodology

ef