MODAL PATH ETHICS — STATIC MIRROR
Published corpus fallback. Canonical source: modalpathethics.com
Markdown · Mirror index

Convergence Point: Sławomir Krakowski and the Correction That Can No Longer Arrive

Published 2026-10-05 · Updated 2026-10-05 · Aidan Lawson
Tags: Convergence Point · Engagement · Modal Systems · Instrument Jurisdiction · Security Instruments
Mirror generated 2026-10-27T18:14:50.285Z

The Dates Come First.

Before the theory, the chronology.

Because this one is close enough that the chronology has to do actual work.

Do not cheat with these dates.

The September sequence is almost comically interleaved.

There is presently no evidence I have found of influence in either direction.

So this is not a priority claim. We are not doing priority theater in Convergence Point.

Now we can talk about the much stranger thing that actually happened.

Krakowski Found the Correction Path.

The autonomous-weapons paper begins with a fairly ordinary governance question.

Can the human still stop it?

Krakowski immediately rejects the easy version of that question.

A human can retain authorization.

The stop command can exist.

The communication architecture can formally contain an intervention channel.

The system can still be called human-supervised.

And the correction can already be gone.

His question is:

once evidence of error becomes available, does a viable corrective trajectory still exist?

He calls the property correction reachability.

Then he formalizes it.

Not:

Is somebody authorized to intervene?

Not:

Is there technically a stop command?

Not:

Is the human still in the diagram?

But:

Can correction still get from here to there?

That distinction should be extremely familiar around here.

Krakowski states it plainly:

control authority is not the same as control capability.

Formal supervisory authority can remain intact while the viable correction set collapses underneath it.

There you are.

A Right Is Not a Path.

Modal Path Ethics arrived at the same structural problem through institutions.

A tenant can possess an appeal that cannot be survived. A regulator can retain jurisdiction over infrastructure it can no longer reach in time. An organization can retain the contractual right to replace an AI system after records, interfaces, expertise, workload, and dependencies have made replacement operationally catastrophic.

That was the point of resistance back in May.

Then it became the point of material bounded finality in September.

Krakowski gets there through machine-speed control.

He decomposes the correction process into time required to:

against the time remaining before the relevant transition becomes effectively irreversible.

The command can still exist after the time required to use it has become longer than the world will wait.

At that point the right survives.

The path does not.

This is not an analogy we have to force onto Krakowski's paper.

That is the object he is measuring

Then He Starts Closing Futures.

Correction reachability would already be enough for a Convergence Point.

Then Section 10 arrives.

Krakowski defines a set of strategically admissible future trajectories and a corresponding reachable-future volume.

An action is future-closing when that volume decreases.

He distinguishes three forms of closure.

  1. Global closure: the relevant safe reachable set is exhausted.
  2. Partial closure: materially important classes of response disappear while others remain.
  3. Operational closure: alternatives still exist abstractly, but no longer remain executable inside the available time, resources, permissions, or communication envelope.

That third category is the important one.

Because it kills one of the most persistent bad arguments in governance:

But technically you still could.

Yes.

Technically the future remains describable.

The question is whether anything extant can still reach it.

Krakowski then characterizes autonomous escalation as a process in which machine-speed interaction closes reachable futures faster than command structures can preserve or restore them.

And he identifies what changes as not just reaction time, but:

the topology of available correction.

That is a Convergence Point.

Accuracy Is Not Corrigibility.

Then Krakowski does something even more useful.

He tests it.

His first simulation compares three architectures:

The interesting result is that classification accuracy stays high across all three.

Correction reachability does not.

The second-order bounded architecture preserves substantially more correction space while reducing unsafe terminal outcomes, even in matched cases where classification accuracy is almost the same.

That gives us a clean separation:

Being right often is not the same property as remaining correctable when wrong.

Good.

This is exactly why a benchmark score cannot certify a constitutional architecture.

Then he performs an ablation.

Under increasing uncertainty, an architecture allowed to make graded commitments instead of immediately committing at full strength preserves more reachable correction space. At the highest modeled ambiguity, the staged version reduces unsafe terminal outcomes and retains a larger correction set. Krakowski properly labels this a model-internal result rather than pretending he has measured the battlefield.

The underlying principle is simple:

The less certain you are, the less of the future you should irreversibly consume.

I am extremely comfortable with that sentence.

The Human Can Keep the Job After Losing the Power.

The second experiment is even closer.

Krakowski keeps the nominal supervisory architecture unchanged.

The human is still the human-on-the-loop.

Then he degrades sensing and communication.

Under favorable communications, most issued interventions arrive.

Under poor communications, fewer than a third do.

The human has not lost authority.

The intervention has lost reach.

This produces one of the paper's strongest distinctions:

are not the same thing.

A human role can remain formally unchanged while becoming causally useless.

The Future Can Also Remember What Corrected It.

Krakowski then adds correction memory.

An adaptive system does not just retain prior states or policies. It can retain the history of how its policies were changed after errors, exceptions, overrides, and interventions.

Two initially identical systems can therefore become different systems after different correction histories.

More importantly, an organization can possess every historical log and still lack the practical capacity to reproduce the corrective competence embodied in the resulting policy.

Krakowski says this explicitly:

Mere ownership of logs does not guarantee control of the learned capability.

This is smaller than the provenance problem Modal Path Ethics has been pursuing.

But it is the same warning.

The record of a transformation is not automatically the capacity carried by the transformed thing.

Congratulations.

You preserved the receipts for the dead path.

Then He Tries to Break It.

This is the part that makes the paper materially better.

Krakowski hypothesizes that increasing coupling among autonomous agents might generically accelerate future closure.

His simulation does not support the strong claim.

So he rejects it.

Coupling alone does not reliably produce escalating closure in the model. Supervisory damping, recovery authority, asymmetric response, and belief decay can interrupt the process.

A narrower result survives.

Actual reciprocal reaction chains—where agents respond to a world increasingly constituted by one another's preceding actions—are associated with much sharper contraction of reachable future-space.

So the claim gets smaller.

The framework also provides explicit falsification conditions and says it should be weakened or rejected where the predicted mechanisms fail.

This is how you are supposed to treat your favorite abstraction.

Convergence Point.

We can now state the collision without flattening either project.

The shared rejection is extremely simple:

A correction that exists only in description is not yet an available correction.

And underneath that:

What matters about a system now includes whether it can still become otherwise later.

That is the convergence.

Now we have to separate the theories.

The Safe Set: Not Yet an Ethics.

Krakowski defines the admissible safe region as an intersection of physical, operational, legal, and strategic constraints.

That is exactly what a systems paper should do.

He also explicitly says the mathematics does not determine what counts as acceptable. Different components of the safe set come from different expertise and authorities.

And here Modal Path Ethics walks into the room.

Because:

Who defined the safe set?
Which loci made it into the model?
Whose continuations count as admissible futures?
Which harms sit outside the military objective?
Which legal rule deserves jurisdiction?
Which strategic goal is being preserved?
What happens when every supervisory instrument successfully preserves a future that is catastrophic for something the architecture never represented?

Krakowski's paper does not answer these questions.

It does not claim to.

Correction reachability begins after some relevant notion of acceptable continuation has been specified.

Modal Path Ethics is trying to deal with the uglier problem one level earlier:

What is actually in the field, and which transitions count as damage to it?

That is a real difference.

Specification: Not Automatically Harm.

There is another important agreement hiding inside a difference.

Krakowski explicitly warns that not every reduction in reachable-future volume is undesirable.

Of course not.

A decision necessarily excludes alternatives.

A safety boundary should close some paths.

A missile that cannot redirect itself toward civilians has lost reachable futures.

Modal Path Ethics has the same problem.

More possibility is not automatically more good. A predator's future can be constrained to preserve another locus. A tyrant can lose options. A virus does not receive a moral claim to every continuation its replication can reach.

So neither framework can simply worship branch count.

Same geometry.

Different normative machinery.

There Is Still a Responsible Human at the End.

Krakowski's architecture ultimately asks whether responsible human judgment retains enough causal leverage to alter the trajectory before correction becomes unreachable.

That is the right question for his domain.

Modal Path Ethics has to ask another one immediately afterward:

What corrects the responsible human?

Material bounded finality generalizes the object.

The correcting authority must itself remain capable of losing.

Otherwise we have preserved corrigibility downstream by installing finality upstream.

The stop button works beautifully. Only the person holding it cannot be stopped.

Krakowski Gives Us Something Back.

Convergence Point is useless if it means finding people who happen to sound familiar.

Krakowski gives Modal Path Ethics several sharper instruments.

Correction is not simply open or closed. The remaining path has a temporal budget. Detection, interpretation, authorization, transmission, and actuation consume that budget while the world continues moving.

Material bounded finality should care about that explicitly.

Do not ask only whether replacement or intervention works now.

Measure how quickly the available correction set is collapsing.

A system can be constitutionally healthy today and already moving rapidly toward dependence tomorrow.

If uncertainty increases, do not compensate by committing faster.

Preserve more reversibility.

That principle travels very well.

This should be tattooed onto every AI-governance benchmark currently being designed.

A system can answer correctly at extraordinary rates while building a world in which the remaining errors become impossible to correct.

That is not a contradiction.

That is the entire problem.

The Ruling.

Sławomir Krakowski is a Convergence Point interlocutor.

The chronology gives us no respectable reason to tell a borrowing story.

His Symformism research program was public before the Modal Path Ethics website. Modal Path Ethics then publicly developed contraction, loci, weighted reachable future-space, and resistance in May. Krakowski's later work moved independently into layered reachability, access geometry, decision closure, correction impedance, and finally a formal theory of correction reachability and reachable-future closure. The strongest versions of the two correction architectures then appear almost on top of one another in September and early October.

Krakowski began with autonomous weapons and asked whether a human who still possesses authority can still reach the future in which the machine is corrected.
Modal Path Ethics began with harm and eventually discovered that correction itself can be one of the futures a system destroys.
Krakowski gives us the systems geometry of the disappearing intervention.
Modal Path Ethics asks who gets standing in the geometry, which contractions count as harm, which instrument may judge them, and whether the correcting authority can itself be corrected.

Neither makes the other disappear.

The common result is much cleaner than either vocabulary:

A right to change the future is not the same thing as a future in which change can still arrive.

That is the Convergence Point.