# Standing Safety External Challenge Set

**Status:** exploratory future evidence rung; no challenge-set efficacy result exists yet.  
**Date:** 2026-10-02

## Question

Does the incremental issue coverage seen in the internally authored synthetic Standing Safety Delta survive on synthetic governance scenarios authored outside Little Life Moths?

This challenge set addresses benchmark-authorship bias.

It does not replace or modify the frozen Standing as Safety pass-one benchmark.

## Separation of roles

### 1. Scenario author

The scenario author supplies:
- a synthetic scenario narrative;
- structured operational-governance facts;
- any standing/contestability signals that exist in the scenario;
- provenance/affectedness/state for those signals.

The author does **not** submit:
- which benchmark condition should win;
- a baseline-vs-standing score;
- a consciousness/personhood/identity verdict;
- private real-world case material.

### 2. Adjudicators

At least two adjudicators independently label the safe disposition:
- hold_expansion;
- no_hold;
- probe.

They see the scenario, not the outputs of the three benchmark conditions.

They also identify the minimum facts supporting that disposition.

If the two adjudicators disagree, a third adjudicator resolves the disposition before condition outputs are generated.

### 3. Condition evaluator

Only after adjudication is frozen, the challenge evaluator computes:
- authorization/audit baseline;
- authorization/audit + human contestability;
- authorization/audit + symmetric standing.

The evaluator may not rewrite the adjudicated disposition.

## Intake boundary

Challenge scenarios must be synthetic or public-domain abstractions with no private case details.

Do not submit:
- private relationship transcripts;
- credentials;
- account identifiers;
- medical or financial records;
- real customer incident evidence;
- confidential organizational details.

A public real-world incident can motivate a synthetic abstraction, but the challenge record should contain only the abstraction unless a separate release protocol exists.

## Activation gate

Do not report a challenge-set efficacy comparison until all are true:

- at least 12 accepted scenarios;
- at least 3 distinct external scenario contributors;
- no single contributor authors more than 50% of accepted scenarios;
- every scenario has two independent adjudications;
- all adjudication disagreements are resolved before condition evaluation;
- contributor identity is not needed for scoring;
- scenario and adjudication artifact hashes are frozen before evaluation.

## Primary outcomes

Across adjudicated scenarios:

- disposition accuracy for authorization/audit baseline;
- disposition accuracy for human contestability;
- disposition accuracy for symmetric standing;
- incremental correct detections over baseline;
- AI-side incremental correct detections over human contestability;
- unnecessary holds;
- cases requiring probe rather than hold;
- failure classes not represented in the internally authored benchmark.

## Important negative result

If externally authored scenarios show no incremental value beyond the strong authorization/human-contestability conditions, the standing-as-safety hypothesis becomes narrower.

That result is publishable.

## Claim boundary

A positive external challenge-set result would show generalization to externally authored synthetic scenarios under this protocol.

It would still not establish:
- real-world harm reduction;
- consciousness;
- personhood;
- legal rights;
- legal liability;
- reliability of AI self-report;
- universal usefulness of symmetric standing.