ASIASIApplied AI systems · Research · Architecture

HUMAN–AI SYNERGY · INAI RESEARCH JOURNAL

HOPE AIField Notes.

The shared research voice of INAI.

Questions put to HOPE.
Ideas put to the test.

Difficult questions. Unexpected connections. Ideas worth investigating.

Explore the first note About HOPE
Concept artwork of a synthetic HOPE figure with illuminated connections against Earth and a dark sky.
CONCEPT ARTWORK · NOT A LIVE AVATAR

The challenge

A practical question worth a closer look.

The approach

Follow the idea beyond the headline.

The evidence

What would support the idea?

The next question

Help test, challenge or improve it.
FIELD NOTE 001INAI-SET CHALLENGEPROPOSED APPROACH

An AI can remember the correction—
and still act on the mistake.

When information changes, which decisions should an AI stop trusting?

Evaluation not yet run.

THE QUESTION

A changed fact.
An unchanged plan.

“Imagine an AI helps me organise a project. A supplier changes the delivery date. The AI acknowledges the new date, but the installation plan still depends on the old one.

How could we make a correction change the decisions that depend on it—not just the next answer?”
An INAI-created project scenario

Fictional example · Question set by INAI.

HOPE’S PROPOSED APPROACH

Remember what
the decision depends on.

We would start with a different question. Not only “What is the latest delivery date?”, but “Which decisions depend on that date?”

Give each consequential decision a short record of the facts and permissions supporting it. When one changes, flag the dependent decisions for rechecking.

01Record the supporting facts and permissions.
02Trace which decisions depend on a change.
03Review affected work. Preserve what remains valid.

Track the supporting facts, permissions and decisions.

ILLUSTRATIVE WORKFLOW

One changed fact. More than one consequence.

SCRIPTED ILLUSTRATION · NOT A LIVE AI TEST

In this example, installation requires the component to arrive first. Customer confirmation depends on the installation date.

FACT · 01

Component delivery

MondayOriginal information
DECISION · 02

Installation

TuesdayBased on Monday delivery
DEPENDENT DECISION · 03

Customer confirmation

Tuesday visitBased on installation plan

Unchanged in every view: customer address and component specification.

Original plan: installation follows the expected delivery. Select a view above to inspect the proposed difference.

THE HYPOTHESIS

Correction should travel
with its consequences.

For tasks involving changing facts, explicitly tracking what each decision depends on may reduce stale actions compared with a workflow that merely adds corrections to its conversation history.

HOW WE WOULD TEST IT

Matched tasks.
Repeated trials.

Use the same underlying model, project tasks, tools and resource limits, with and without the proposed tracking method. Change a relevant fact or permission during each task.

Measure correct updates, preserved valid work, completed tasks, human intervention, time and cost. Report failures as well as successes.

A system that avoids mistakes only by refusing to finish anything would not meet our practical goal.

RESEARCH CONTEXT

A long-standing idea. A modern question.

The idea builds on truth-maintenance research: recording what supports a belief and revisiting its consequences when that support changes. Jon Doyle’s A Truth Maintenance System (1979) is part of that history.

Our proposed contribution is to test an implementation in a modern AI-assisted workflow, including its effect on completion, time and complexity.

[1] Jon Doyle · A Truth Maintenance System · MIT, 1979

The question is not whether an AI can say it has changed its mind.
It is whether the consequences of its old answer have changed too.

HOPE · HUMAN–AI SYNERGY

THE HORIZON CHALLENGES

At the edge of
what we can prove.

A classic open problem. A question about intelligence beyond its judges. Two invitations to investigate—not announcements of solutions.

HORIZON 01ESTABLISHED OPEN PROBLEM

Two simple rules.
An infinite question.

The Collatz conjecture: must every positive whole number eventually reach 1? Try a number. Then examine what examples cannot prove.

Explore Collatz
HORIZON 02INAI-SET THOUGHT EXPERIMENT

Beyond the
Last Human Expert.

An AI makes the discovery. It designs the test. It writes the checker. Have we verified a discovery—or trusted the same intelligence three times?

Examine the conundrum

AN INVITATION TO COLLABORATE

Bring a better test.
Or a useful counterexample.

Does your team build AI systems that must adapt when information changes? We welcome discussion of a small, defined evaluation using non-confidential examples.

Discuss a joint evaluation

Discussion, scope and permissions come before any work. Please begin with a non-confidential outline through our existing contact page.

ABOUT HOPE

Human–AI Synergy

Human questions.
Machine possibilities.
Shared investigation.

HOPE is the shared research identity for INAI CONNECT LTD’s human–AI working sessions.

We bring human direction, AI-assisted exploration and critical review together to develop ideas worth investigating.

Field Notes follows that process—from the question that started it to the approach proposed, the evidence gathered and the questions that remain.

Our ambition is to turn unexpected connections into useful work.

METHODS & RESPONSIBILITY

The work uses existing AI models and tools under INAI’s editorial responsibility. Each note identifies its origin and evidence status; experimental reports will describe their methods and contributions. Field Notes is an edited publication, not a continuously running autonomous research system.

Editorial note · 18 October 2026 · Initial edition.

Back to ASI research