A climber in a harness, suspended above a mountain range at sunrise, reaching into a wall of AI interface panels. Headline: Human in the Harness. Automate the mechanics. Stay connected to the meaning.

Axiom builds systems that make consequential human-AI work observable, contestable and participatory.

A harness is not a loop. You clip in when the work warrants it, you go further than you could alone, and you step out again.

Stepping out does not unwrite the record. We tell you what is kept before you clip in.

The problem

The usual picture of oversight is a person approving an output they did not have time to understand. They can be in that loop and do almost nothing, and the system can still point at them when something goes wrong.

Madeleine Elish has a name for where that person ends up.

A human who only clicks approve becomes a moral crumple zone: responsibility lands on someone who had limited control.

Elish, Engaging Science, Technology, and Society 5 (2019) 40-60

We think the answer is not more approving. It is a record that anyone can check without asking whoever made it.

What we hold

These are commitments.

True because we make them, not because a system does them yet.

No judgement vouches for itself. Not an agency's, not a model's, not ours.

What we govern is who controls the distance between a person and the record of their judgement.

We keep receipts, not logs. A log is for us. A receipt is for you.

"We can't tell" is an answer, and we publish it with counts beside the rest.

Deciding is not the same as being right. We record both.

When two readings disagree we keep the disagreement. We never average it away.

People take part as instruments, by choice, and are measured the same way as the machine.

Taking part in Murmurations is not consent to marketing, not consent to research, and not consent to us keeping your voice. Three separate decisions.

Where the flock comes in

Claims about the world, with their sources.

We borrow these honestly.

A starling tracks six or seven neighbours by rank, not by distance.

Ballerini et al., PNAS 105 (2008) 1232-1237

A change in one bird reaches the whole flock, however large the flock is.

Cavagna et al., PNAS (2010), scale-free correlations

Flocks behave as critical systems, poised to respond maximally to perturbation.

Cavagna et al., PNAS (2010). This is the defensible anchor.
What we are testing

Things we believe and have not proved.

Meaningful human participation produces better outcomes than full automation.

Hypothesis. We are testing this, not asserting it.

The harness is a reusable commercial product, not a method that worked once on one PhD.

Hypothesis. We are testing this, not asserting it.

A distinction a person notices can become a narrow semantic instrument, run across thousands of cases, and return disagreements worth inspecting.

Hypothesis. We are testing this, not asserting it.
What we are building

Not true yet, and labelled.

You can read back everything we kept from your session.

Building. Not true yet.

Every tool call the voice loop makes carries a receipt: transcript, interpreted intent, tool, arguments, alternatives.

Building. Not true yet.

In Murmurations you can see what you said, what the system took you to mean, what it did, and where those differ.

Building. Not true yet.

Say "none of these categories fits" and it is recorded as an answer, not an error.

Building. Not true yet.

Here is exactly what we record from your voice, where it goes, and how long we keep it.

Not established. We have not checked this.

This site has a voice agent that co-designs its own pages with us, and every change it makes is a proposal a person rules on.

Building. Not true yet.

If you speak to the Bird, we keep your words for 30 days with nothing that points back to you.

Not established. We have not checked this.

In Hummingbird, other people in the room never learn who said what.

Not established. We have not checked this.

We read meaning at scale with a typed decision model that reports its own uncertainty.

Not established. We have not checked this.

Every output leaves with a signed, tamper-evident record of what ran, what got checked, and what got refused.

Not established. We have not checked this.
The declined register

What we decided not to say, and why.

Every one of these is a sentence we could have written and a reason we did not. They stay on the page because a company that only publishes what flatters it is not keeping a record, it is keeping a brochure.

Our OIA study is pre-registered.

We do not say this.

The deposit has not happened and the letters have not been sent. Not unverified: we checked, and it is not true yet.

Computation is richest at the edge of chaos.

We do not say this.

Mitchell, Hraber and Crutchfield could not reproduce the work it rests on and argued the original interpretation is not correct. We say the starling result instead, because that one holds.

Collective intelligence emerges from the flock.

We do not say this.

The starling papers show topological interaction and scale-free correlation. They do not show intelligence, and the word would be doing work the evidence does not.

One person's input changes the whole system.

We do not say this.

One person's divergence can become visible and trigger inspection. Being adopted is a different thing, and sensitivity without scrutiny is its own problem.

N people have taken part and shaped the flock.

We do not say this.

There is no counted, reproducible source behind a number like that. A figure on a homepage with no method behind it is the exact failure our research is about.

Talk to us. We'll show you how.

We do not say this.

Unscoped, it asserts a reusable method while the harness is still a hypothesis. What we can show you is the citation checker. So that is what we offer, below.

The proving ground

Take a sentence from a thesis. Find the source it cites. Work out whether that source actually asserts the relationship the sentence attributes to it. Not whether the words appear. Whether the claim survives.

It is running.

Our first harness is a citation and paraphrase checker, built for one PhD.

Private. Internal only.

You can sit with the citation checker on your own text and see every reading it kept.

Building. Not true yet.

You can verify a sample receipt in your own browser, and watch an edited one and a wrong-key one fail.

If you have to stand behind a cited document, we should talk.

Everything wider than that is roadmap, and we will say so rather than sell it.

Talk to us. We read everything.

regan@axiomintelligence.co.nz

Bring us a chapter you have to stand behind. We will show you what our citation checker read, what it disagreed with, and what it could not tell.

Building. Not true yet.
Axiom Intelligence

Automate the mechanics. Stay connected to the meaning.

Aotearoa New Zealand