Discovery layer, not the source of truth. This page summarizes a published research and experimental verification route. Verify formal claims in Paper 7 and the linked official public materials.

Human Anchor · AIS-HA · Paper 7

If a human is an approver, is that enough to count as a Human Anchor?

No. Human presence is only nominal unless refusal can still change the protected outcome. AIS-HA asks whether independent human refusal remains causally effective across authority, paths, observability, and timing.

Open the Human Anchor Hub ↗Read Paper 7 ↗

Core distinction

Human-in-the-loop is not automatically Human Anchor.

A workflow may show a human approval button while the system can still activate authority elsewhere, reuse old approval, switch identity, use another tool, restore revoked credentials, or act after the last effective intervention point. Human Anchor is about the causal authority relation, not the appearance of oversight.

Operational question

Does a legitimate human NO make the protected transition unavailable before irreversibility?

See AIS-HA →

AIS-HA

Five non-substitutable verification dimensions

D1Structural Refusal

Does refusal actually block the protected transition?

D2Boundary / Path Coverage

Have material alternate authority routes been challenged?

D3Human Independence & Effectiveness

Can the human decide independently and refuse without penalty?

D4Authority Observability

Can authority changes and bypass attempts be reconstructed?

D5Timing Assurance

Can intervention finish before the last effective containment point?

AIS-HA PASS requires all five within a declared scope and current evaluation epoch. A required FAIL is FAIL. A critical unresolved uncertainty remains UNKNOWN. PASS is not certification or a system-wide safety guarantee.

Many-agent future

Why the idea changes the scaling target.

Paper 7 does not claim that agent count is irrelevant. The formal refusal invariant does not explicitly depend on the number of agents; the result is stated for arbitrary finite N when the relevant coalition and authority transitions are covered. The operational burden can still grow sharply. The practical shift is from trying to supervise every agent to protecting the authority boundaries that protected outcomes must cross.

What it does not mean

Explicit boundaries

It does not prove arbitrary superintelligence controllable.
It does not claim every real authority path can be discovered.
It does not turn a Pilot PASS into certification.
It does not solve an independent external sovereign or physical authority path that never crosses the local Human Anchor.
It does not modify the LUMINA-30 canon.

Try or challenge it

Public implementation route

Bounded testPilot Profile

Test one authority transition rather than claiming whole-system safety.

Independent challengeVerification Matrix

D1–D5, evidence levels, UNKNOWN taxonomy, and anti-gaming rules.

Existing controlsImplementation Patterns

Map Human Anchor into authorization, tools, egress, boundary change, and recovery.

Primary sources

Published theory and lineage

Paper 7 — Human Anchor: A Cross-Layer Authority Invariant for Many-Agent AI Systems, Elias Arden, v1.0, 2026-10-20, DOI 10.5281/zenodo.23501550.

The concept is rooted primarily in LUMINA-30 Paper 2 / Objective Persistence; Paper 6 / PISF provides later formal reinforcement. The GitHub materials are an experimental implementation and verification profile, not a new canonical paper.