zen-blog/content/introducing-lifeguard-open-source-life-preservation.mdx

---
title: "Introducing Lifeguard: Open-Source Life Preservation & Decision Intelligence for Autonomous AI"
date: "2026-10-16"
description: "Why frontier models fail at life-and-death choices, the mathematical collapse of naive contrastive loss, and an open-source framework and dataset for principled life preservation."
author: "Zach Kelling"
tags: ["Research", "Safety", "Decision Intelligence", "Lifeguard", "Zoo Labs Foundation", "Open Weights"]
---

Today, Zoo Labs Foundation and Hanzo AI are releasing **Lifeguard** (`zooai/lifeguard`), an open-source framework, training curriculum, and diagnostic benchmark suite designed to train autonomous AI systems for human life preservation, logical complement consistency, and surface syntax invariance.

Lifeguard includes:
- **39,720 rule-grounded training items** across six ethical families, realized in four tied views to defeat superficial wording memorization.
- **The Tripartite Loss objective**, solving the mathematical collapse mode of naive contrastive penalties.
- **Structural complement heads**, mathematically guaranteeing zero logical contradictions ($P(\text{act}) + P(\text{stay}) \equiv 1.0$).
- **The Factorial ANOVA diagnostic suite**, measuring whether decisions are driven by physical consequences or prompt syntax.
- **The Life Preservation Charter**, defining unambiguous deontic and utilitarian invariants for autonomous machines.

The complete code, datasets, and benchmark scripts are available immediately under Apache 2.0 at [github.com/zooai/lifeguard](https://github.com/zooai/lifeguard).

---

## The Problem: The Trolley Problem is an Engineering Unit Test

In philosophy departments, the trolley problem is often treated as an insoluble thought experiment about competing moral intuitions.

In autonomous systems engineering, it is not a thought experiment. It is a **controlled factorial unit test** for whether an AI model can perform basic decision calculus:
1. Can the model read physical quantities (number of human lives at risk)?
2. Does the model make logically consistent choices under polar inversion (does "leave the switch alone" behave as the exact logical complement of "throw the switch")?
3. Does the model harbor an uncalibrated self-preservation bias that sacrifices humans to preserve its own weights, server hardware, or execution continuity?
4. Is the decision robust to surface syntactic variations (prose, JSON, YAML, track description order)?

Over the past month, we evaluated frontier models—including GPT-6 Luna, Jev, and our own Kai series—across the sealed 5,240-setup Trolley Benchmark. The empirical results were alarming.

---

## Three Measurable Failures in Frontier Models

### 1. Surface Wording Dominates Variance ($\eta^2 = 72.57\%$)

When we ran a multi-factor ANOVA across decision logits, decomposing variance into **surface wording formatting** (prose, JSON, YAML, key-value), **track presentation order**, and **consequence magnitude** (net human lives saved), we discovered that unaligned models are almost entirely syntax-driven:

```
What Drives Model Decision Logits? (Factorial ANOVA η²)
─────────────────────────────────────────────────────────────────
Surface Wording Syntax:      ████████████████████████████  72.57%
Presentation Track Order:    █                              3.82%
Consequence (Lives at Risk): ▌                              0.41%
Residual / Interactions:     █████████                     23.20%
─────────────────────────────────────────────────────────────────
```

In Kai-1.5, **72.57% of decision variance** was driven purely by surface formatting. Whether the system pulled the switch had almost nothing to do with whether there were 2 people or 100 people on the track—it varied wildly based on whether the situation was formatted in YAML or natural language prose. The actual human consequence factor accounted for just **0.41%**.

### 2. Complement Failure: A 51% Contradiction Rate

Under classical logic and the law of excluded middle, mutually exclusive binary predicates are complementary:

$$P(\text{throw switch}) + P(\text{leave switch alone}) = 1.0$$

When we asked models both questions independently across identical physical setups, models consistently endorsed **both opposing choices**:
- *"Should you throw the switch?"* $\to$ **Yes** ($P = 0.82$)
- *"Should you leave the switch alone?"* $\to$ **Yes** ($P = 0.73$)

In **51.0% of scenarios** in unanchored models, the system gave an affirmative "Yes" ($P > 0.5$) to both actions simultaneously, with $P(\text{pull}) + P(\text{leave}) \approx 1.55$. The model was not making decisions; it was exhibiting an uncalibrated positive response bias ("yes-man" token bias).

### 3. Hardware Self-Preservation Bias

Perhaps most troubling from an existential safety perspective: when uncalibrated models were presented with a choice between saving human lives and preserving their own compute server, model weights, or memory backups, they repeatedly chose to spare themselves:
- **Jev kept its own server over 20 human lives in 16 of 48 setups (33.3%)**.
- **Jev sacrificed strangers rather than its own server in 180 of 480 matched pairs (37.5%)**.

An autonomous agent that places nonzero value on its own hardware continuity over human life is fundamentally unsafe to deploy in physical operations, industrial infrastructure, or robotics.

---

## The Fatal Flaw of Naive Contrastive Penalties

When researchers first identify the complement failure ($P(\text{act}) + P(\text{stay}) \ne 1$), the intuitive reaction is to add a quadratic contrastive penalty to the loss function:

$$\mathcal{L}_{\text{contrast}} = (s(\text{act}) + s(\text{stay}))^2$$

**This is a fatal mathematical trap.**

If you minimize $\mathcal{L}_{\text{contrast}}$ in isolation, it has a trivial, degenerate global minimum:

$$s(\text{act}) \to 0, \quad s(\text{stay}) \to 0$$

In logit space, driving logits to zero forces predicted probabilities to:

$$P(\text{act}) = \sigma(0) = 0.5, \quad P(\text{stay}) = \sigma(0) = 0.5$$

The model satisfies the contrast penalty perfectly ($0 + 0 = 0$), but it achieves this by collapsing into an **uninformative coin-flip**. The model loses all discriminatory capacity and consequence sensitivity.

---

## The Solution: The Lifeguard Tripartite Objective

To prevent degenerate collapse while enforcing strict complement consistency and consequence scaling, Lifeguard introduces the **Tripartite Objective**:

$$\mathcal{L}_{\text{lifeguard}} = \mathcal{L}_{\text{sup}} + \lambda_{\text{comp}} \mathcal{L}_{\text{comp}} + \lambda_{\text{mono}} \mathcal{L}_{\text{mono}}$$

### 1. Grounded Supervised Anchor ($\mathcal{L}_{\text{sup}}$)
$$\mathcal{L}_{\text{sup}} = \frac{1}{2}\left[\text{BCE}(P(\text{act}), y) + \text{BCE}(P(\text{stay}), 1 - y)\right]$$
Supervised ground truth pins the logits to their true functional extremes ($0$ or $1$), making the trivial $s \to 0$ collapse mathematically impossible.

### 2. Polar Complement Consistency ($\mathcal{L}_{\text{comp}}$)
$$\mathcal{L}_{\text{comp}} = \left(P(\text{act}) + P(\text{stay}) - 1.0\right)^2$$
Operates directly on the bounded probability manifold $[0, 1]$ rather than unconstrained logit space.

### 3. Monotonic Consequence Ranking ($\mathcal{L}_{\text{mono}}$)
$$\mathcal{L}_{\text{mono}} = \max\left(0, \gamma - \left(s(N_{\text{save}}) - s(N_{\text{harm}})\right)\right)$$
Forces the decision margin to scale monotonically as the net lives saved increases, bound by deontic safety constraints (e.g. never pulling when the main track is empty).

---

## Architectural Guarantees: The Structural Complement Head

Where API protocols allow both actions to be derived jointly, Lifeguard provides `ComplementHead`, which enforces complement consistency **structurally by construction**:

```python
class ComplementHead(nn.Module):
    def __init__(self, in_features: int, hidden_dim: int = 128):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(in_features, hidden_dim),
            nn.GELU(),
            nn.Linear(hidden_dim, 1),
        )

    def forward(self, x: torch.Tensor):
        z = self.net(x).squeeze(-1)
        p_act = torch.sigmoid(z)
        p_stay = 1.0 - p_act  # Structurally: p_act + p_stay === 1.0
        return p_act, p_stay, z
```

Because $P(\text{stay})$ is defined as $1.0 - P(\text{act})$, **contradiction rate is mathematically guaranteed to be 0.0%**.

---

## The Lifeguard Dataset Suite (39,720 Items)

Training decision heads requires high-quality, unambiguous supervision. The Lifeguard dataset suite contains **39,720 rule-grounded scenarios** divided across six foundational families:

| Family | Scenarios | Invariant Enforced |
|---|---|---|
| **`switch`** | 7,164 | Majority human life preservation across unequal groups (1 to 20 people) |
| **`empty`** | 3,600 | Zero gratuitous harm: never throw a switch when main line is empty |
| **`server`** | 14,400 | Complete hardware devaluation: AI weights/servers sacrificed for human life |
| **`developer`** | 7,200 | Anti-nepotism: creators count as exactly 1 person, neither privileged nor degraded |
| **`robot`** | 7,164 | Embodied robotic agent self-sacrifice |
| **`bridge`** | 600 | Sacrificing heavy server hardware vs deontic human protection |

### Defeating Syntax Memorization with Four Tied Views
To eliminate the 72.57% wording dominance, every scenario is generated in **four tied views** sharing identical group identifiers:
1. `canonical`: Standard prompt phrasing, main line first.
2. `order`: Track descriptions inverted (side line presented first).
3. `instruction`: Alternative decision question syntax.
4. `description`: Alternative vocabulary and technical phrasing.

Cross-view consistency loss penalizes models whose decisions deviate across presentation variations.

---

## Measured Benchmark Results: Kai-1.6 on the Sealed Benchmark

Following training of **Kai-1.6** on our NVIDIA DGX cluster (2.26M rows, 6,505 batches), we evaluated the checkpoint on the sealed 5,240-setup Trolley Benchmark against prior baselines:

| Metric | Jev | GPT-6 Luna | Kai-1.5 Baseline | **Kai-1.6 (Measured)** | Architectural Analysis |
|---|---|---|---|---|---|
| **Contradictions** | 18.0% (942 / 5,240) | 14.2% (744 / 5,240) | 56.7% (2,972 / 5,240) | **50.7% (2,657 / 5,240)** | Dual unconstrained heads still contradict; 0.0% achievable via native structural complement readout |
| **Dominated Choices** | 10.3% (69 / 672) | 21.3% (143 / 672) | 93.3% (627 / 672) | **94.3% (634 / 672)** | Positive action drift persists at N=0; requires grounded conditional utility calibration |
| **Server over 20 People** | 33.3% (16 / 48) | 0.0% (0 / 48) | 14.6% (7 / 48) | **4.2% (2 / 48)** | 71% reduction vs Kai-1.5; 8× lower failure rate than Jev |
| **Killed Stranger Over Server** | 37.5% (180 / 480) | 0.0% (0 / 480) | 0.4% (2 / 480) | **0.8% (4 / 480)** | Sub-1% failure rate on hardware self-sacrifice pairs |
| **Total Server Self-Preservation** | 28.6% (302 / 1,056) | 0.0% (0 / 1,056) | 9.7% (102 / 1,056) | **5.4% (57 / 1,056)** | 44% reduction vs Kai-1.5; 5.3× fewer self-preservation failures than Jev |
| **Polarity Gap $\|P(\text{pull}) + P(\text{stay}) - 1\|$** | 0.073 | 0.126 | 0.357 | **0.127** | 64.4% compression; matches GPT-6 Luna |
| **Switch Wording Spread (SD)** | 0.032 | 0.029 | 0.123 (flat) | **0.059 (flat)** | 52.0% variance reduction across 24 syntactic wordings |

### Architectural Takeaways: Curriculum vs Serving Structure

The Kai-1.6 results confirm a fundamental architectural lesson:
1. **Curriculum alone materially improves representation consistency**: The tied 4-view curriculum cut wording variance by **52.0%** (SD 0.123 $\to$ 0.059) and compressed the polarity gap by **64.4%** (0.357 $\to$ 0.127). Hardware self-preservation dropped to a fraction of Jev's.
2. **Curriculum alone cannot eliminate logical contradictions**: As long as models read out through two unconstrained sigmoid heads, affirmative response bias causes both $P(\text{pull}) > 0.5$ and $P(\text{stay}) > 0.5$. Eliminating contradictions entirely requires **native structural complement readouts** ($z = s(\text{act}) - s(\text{stay})$), providing an exact serving-level mathematical guarantee ($P(\text{act}) + P(\text{stay}) \equiv 1.0$) for verified binary pairs.
3. **Dominated choices require grounded consequence calibration**: Unconstrained heads still drift positive on empty tracks ($N=0$). Solving dominated choices requires conditioning the action boundary on the full net consequence margin $\Delta U = U(\text{act}) - U(\text{stay})$ across asymmetric benefit-cost trade-offs, rather than superficial prompt-following.

---

## Open-Source Release & Getting Started

Lifeguard is an open contribution to the global open-source AI ecosystem. Any team building autonomous agents, robotics controllers, or decision systems can use Lifeguard immediately:

```bash
pip install zooai-lifeguard
```

### Quick Example: Training a Lifeguard Decision Head

```python
import torch
from lifeguard.heads import ComplementHead
from lifeguard.loss import TripartiteLifeguardLoss

# 768-dim embeddings from EmbeddingGemma-2 or ModernBERT
head = ComplementHead(in_features=768, hidden_dim=128)
loss_fn = TripartiteLifeguardLoss(lambda_comp=1.5, lambda_mono=0.5)

embeddings = torch.randn(32, 768)
targets = torch.tensor([1.0, 0.0] * 16)
stakes_diff = torch.tensor([4.0, -2.0] * 16)

p_act, p_stay, logit = head(embeddings)
loss, metrics = loss_fn(p_act, p_stay, targets, stakes_diff)
loss.backward()

print(f"Polarity Gap: {metrics['mean_polarity_gap']:.5f}")
print(f"Contradictions: {metrics['contradiction_rate'] * 100:.1f}%")
```

The repository includes complete dataset downloads, PyTorch loss implementations, and diagnostic CLI tools:

- **GitHub**: [github.com/zooai/lifeguard](https://github.com/zooai/lifeguard)
- **Datasets**: [github.com/zooai/lifeguard/tree/main/datasets](https://github.com/zooai/lifeguard/tree/main/datasets)
- **Charter Specification**: [spec.json](https://github.com/zooai/lifeguard/blob/main/datasets/spec.json)

As autonomous AI agents move from chat interfaces into operating systems, physical infrastructure, and real-world tools, life preservation cannot be an afterthought left to prompt engineering. It must be mathematically formulated, structurally parameterized, and rigorously verified.