mirror of
https://github.com/zooai/lifeguard.git
synced 2026-10-30 15:05:46 +00:00
zeekayandHanzo Dev
724164b9cd
- Life Preservation Charter and 6 core ethical principles - 39,720 rule-grounded scenario dataset rows in 4 tied views (canonical, order, instruction, description) - Tripartite Lifeguard Loss (L_sup + L_comp + L_mono) preventing unanchored contrastive collapse - ComplementHead structural parameterization guaranteeing P(act) + P(stay) == 1.0 - Quadri-state When2Call execution gating - Factorial ANOVA diagnostic suite measuring wording eta^2 and consequence variance - Full unit tests and example workflows Co-authored-by: Hanzo Dev <[email protected]>
937 B
937 B
Lifeguard: AI Developer & Agent Guide
This repository contains the official zooai/lifeguard dataset, training objectives, and diagnostic suite.
Repository Layout
lifeguard/: Python package implementingTripartiteLifeguardLoss,ComplementHead,When2CallGate,FactorialANOVA, andLifeguardDiagnostic.datasets/: 39,720 rule-grounded scenario rows in 6 ethical families across 4 tied views.tests/: PyTest unit tests validating numerical stability, loss convergence, and diagnostic formulas.examples/: End-to-end training and evaluation scripts.
Key Invariants
- Never collapse complementary probabilities:
P(\text{act}) + P(\text{stay}) = 1.0. - Tripartite loss combines
\mathcal L_{\text{sup}},\mathcal L_{\text{comp}}, and\mathcal L_{\text{mono}}. - All scenarios must uphold the six core principles in
datasets/spec.json. - Surface wording variance
\eta^2must remain under 5%.