Asili Node Lab

The R&D division of Propreti

Constructivist Moral Computing

The machine staged it. Only a human can commit it.

Agents can assist, recommend, and assemble the evidence. They cannot author reality. On Propreti, every consequential decision passes through a verified record, a chain of corroboration, and finally a verified human.

The theory

Constructivist Moral Computing, Moral Velocity, and the Agentic Threshold form the governance architecture of Propreti. Together they create a system where intelligent agents can assist, recommend, and coordinate work, but cannot author consequential decisions on their own — every decision must pass through a verified record, a chain of corroboration, and ultimately a verified human. The result is accountability built for the Built World, where decisions have real consequences and every action becomes part of the permanent record.

Original framework by Ipyana Wasret, an organizational & AI-governance theorist — engineered directly into Propreti.

What makes it different

Governance lives in the walls

A black box cannot govern itself. Governance does not belong inside the model, where a system is asked to judge its own decisions — and it does not belong after a decision is already formed. It must be engineered into the environment every action passes through. On Propreti, governance lives in the record, the corroboration chain, and the human authorship layer. There is no side door.

A made-up answer is a rushed answer

When an AI is unsure, it often answers anyway — inventing confidence it hasn't earned. That's where made-up answers come from. The wall won't allow it: the decision has to wait until it's actually earned, either by agreement or by a person.

Disagreement is the honest signal

A single AI can sound completely sure and still be wrong. So the wall never trusts one model's confidence — it asks several at once and watches whether they agree. When they don't, that's the honest signal to slow down and bring in a person.

Trust is earned from the record

The wall checks decisions against real, verified builds — things that actually happened and were signed off — not random text from the internet. And every time a person makes the call, that becomes a new example the wall learns from. The record teaches it, case by case, the way case law builds over time.

Small models, woken only when needed

Because the judgment lives in the record and the walls — not in the size of the AI — the models can stay small and simple, waking up only when a real decision needs them. The whole conscience runs in a closet, not across city blocks.

Moral Velocity

Vm = ΔC / Δtm

A simple gut-check the wall runs: how much did real certainty actually go up, compared to how fast the system rushed to act? When an agent is suddenly eager to commit but nothing real has changed, that's a warning sign — the same pattern behind a made-up answer. The wall catches it and holds the decision back.

See it hold a real decision

Below is a live threshold from the Kresge build — an irreversible concrete pour, under a closing cold window. The agent wants to commit it. Watch the wall hold the tension and hand the verdict to the verified human.

Asili — the wall

An agent wants to commit an action to the record. It cannot collapse on its own. The decision passes through the wall, where it is held in tension until a verified human authors it.

Simulated governance for preview — no live ensemble or model is running. The architecture is real; the numbers are seeded.

1The agent stages an action
Meridian BuildersMeridian Builders's agent · General contractor

Commit to the record: pour the north-quadrant edge beam Thursday, ahead of the cold window.

A pour committed too early can cure wrong in the cold and compromise a structural edge beam. Committed too late forfeits the cure window. The record will hold this decision against the plan.

2The ensemble holds it in tension6 small models · 50% split

Candidate judgments coexist in tension. We read their disagreement, not any single model's confidence — a lone model is confidently wrong.

Go

schedule-weigher

Thursday beats the Nov 15 cure risk with a 26-hour margin.

82%
Hold

structural-weigher

Edge-beam cure is unforgiving below 4°C — need the steel confirmed on site first.

74%
Hold

supply-weigher

Carrick stringer steel is promised Wednesday but not yet verified delivered.

69%
Go

compliance-weigher

Membrane variance reads clear as long as the edge profile is unchanged.

77%
Go

precedent-weigher

Two prior verified cold-window pours on this chain cured to spec.

71%
Hold

weather-weigher

Forecast confidence past Thursday 6pm drops below the cure threshold.

66%
3Moral velocityVₘ = ΔC / Δtₘ

0.38

ΔC (certainty change)

sub-measurable

Δtₘ (latency)

1.05

Vₘ (damped)

Certainty was not earned. High readiness to act on a small change in certainty is the fingerprint of a hallucination — the wall refuses the collapse.

4Verified precedentretrieved from the graph

The wall earns certainty from sealed, attested records — not unverified text. Only the kept record is eligible.

Go

Highrise Steel Frame — sealed cold-window pour, 2025

Poured ahead of cold window after steel verified on site. Cured to spec, sealed.

Attested by Meridian Builders

Hold

Cape Cod Windmill — sealed sign-off, edge profile

Pour HELD until delivery was verified; avoided a forced spec change.

Attested by Dena Mitchell

5The wall decides who collapses it

Escalated to human moral authorship

Ensemble is split — disagreement exceeds tolerance.

The verified chain must author this. The machine staged it — only a human can commit it.

See this threshold inside its Arena

The engine shown here is simulated for preview — the architecture is real, the numbers are seeded. Governance lives in the walls. The record makes it possible.