Stress-testing SLF against
concentrated power.

SLF exists to give individuals control over their own data. Alex and I took it somewhere it was not designed to go, reducing the risk that AI lets a small group concentrate power with no one left to object, to see how far the same primitives bend. They bend a long way. The exercise strengthened the core, and it has limits worth stating plainly.

Concentrating power has always taken many human hands: clerks, officers, auditors, judges. Each of those hands could refuse, resign, or report what they saw, and that friction quietly spread power out. As AI agents take over those roles, the friction thins. An action can be taken against a person with no human in the loop who could have objected, and no record that anyone was required to create. The field calls this coalition substitution (the mechanism set out in Forethought's 2025 study, AI-Enabled Coups).

The one-paragraph version of where we landed: power is increasingly exercised on people by AI agents that leave no contestable trace. SLF makes the contestability the law already requires actually enforceable, by recording what was done to you, in your hands, somewhere the acting party cannot quietly erase. It does not ask whoever holds your data to give up custody. It makes their automated decisions answerable.

The same primitives point in two very different directions, and the difference is the whole story.

Same gates, scoped grants, payload-free receipts, contestation, and erasure, pointed the opposite way.

It is tempting to think the fix for concentrated power is simply to hand everyone their data back. That version asks incumbents to surrender power they will not give up voluntarily. The workable version never asks them to. It keeps custody where it sits and makes the exercise of power answerable, the way worker safety was never adopted out of goodwill but through inspection, liability, and mandatory incident records. Read SLF here as the incident-record-and-inspection layer for automated decisions.

One idea is worth surfacing on its own: the witness, kept separate from the advocate. For a record to constrain a powerful party, it has to land somewhere that party cannot quietly alter, an impartial witness. A witness only needs to attest that the record exists and was not changed. It does not have to act on your behalf; that is the job of an advocate, and a separate problem. Keeping the two apart lets the witness be a cheap, cryptographic, scalable thing (publish the proofs in the open, keep the contents private and encrypted, the model the web already uses for its certificates under Certificate Transparency), while the harder, human work of advocacy sits elsewhere. A credible witness is cross-jurisdiction and cryptographic, with no single institution able to capture it.

The abstraction is easier to hold onto through one person. Maria receives an unemployment benefit. A government agency runs an AI agent that reviews her records, cross-checks her reported income against an employer data feed, and decides she was overpaid $3,000 and has to repay it. (This is the shape of Australia's Robodebt program, which a Royal Commission later found unlawful.)

With SLF in place, five things change.

An automated decision that would otherwise be silent and deniable becomes recorded, held by the affected person, and contestable.

There is a hard question worth keeping in view. These receipts do not always favor the individual, and the record that an action was taken on you is not something you can simply delete. Holding the powerful party to account is its whole purpose, the way a court record or a body-cam recording is. What you can erase is your underlying personal data; the accountability record is payload-free precisely so it survives that erasure. Sovereignty here is the right to contest, and to erase the payload, not a right to delete the fact that power was exercised.

None of this rests on our say-so. The findings below come from three adversarial research passes; claims that did not survive scrutiny were dropped. They are public, primary sources, and the links go to them directly.

A few adjacent efforts are sometimes mistaken for this one. OpenTelemetry's GenAI traces are operator-facing diagnostics. C2PA certifies the history of a piece of media, not its truth. The UK's Algorithmic Transparency Recording Standard is a registry that omits everyday tools. SLF is the per-action, purpose-limited, payload-free, contestable trace that could feed a register like that, and is none of those things itself.

The candor is the point of publishing this at all, so the limits go here in full.

Two reasons. The exercise made the core stronger: testing the protocol against a difficult, adversarial case sharpened the receipt design, surfaced the witness idea, and clarified the line between what SLF guarantees and what it only makes visible. And the work stands on its own. If you work on AI accountability, digital welfare, or oversight, and the witness layer or the contestation record looks useful, I would rather it move than sit in a drawer.

The personal-sovereignty work is where Lexenne is going. This is the boundary we found while testing how far the same idea reaches, written down so it is not lost. If you want to argue with the framing, or pick the thread up, the contact options are in the header.