What Does It Mean for a System to Still Be Itself?
Across identity systems, artificial intelligence and institutions, we keep encountering versions of the same practical problem.
Something can continue producing the correct signal after the conditions that originally made the signal meaningful have changed.
A private key still signs. An account still authenticates. A procedure is followed correctly. An AI produces an answer that sounds appropriate. None of those observations necessarily tells us whether the larger situation remains what we think it is.
This distinction becomes especially visible under stress.
The vault is empty, but there's an insurance claim. The data is inaccessible, but there's a ticket number. The account is frozen, but an appeal is pending. These aren't necessarily examples of malfunction. The systems may be behaving exactly as designed. What has changed is the question we're asking.
We no longer want to know whether a valid claim exists.
We want to know what exists now.
That difference between the signal and whatever the signal refers to has occupied a great deal of my thinking about infrastructure, identity and long-running interaction. I previously tried to gather those observations under a general theory of coherence.
I think I was moving too quickly.
There is still something here worth measuring. I'm less certain that all of these domains contain the same thing.
Control is not continuity
Cryptography gives us unusually precise tools for establishing control. If a valid signature can only be produced with access to a particular private key, possession of the required key tells us something important.
It doesn't tell us everything.
Suppose an organization has used the same multi-signature arrangement for ten years. Every signature remains cryptographically valid, but the people have changed, the institution has changed, its purposes have changed and perhaps the relationships that originally justified its governance structure have changed too.
The cryptography hasn't failed.
We've changed levels of description.
The same problem appears in less technical forms. A newspaper can retain its name while changing ownership, staff, audience and editorial practices. A town can retain its legal boundaries while becoming a substantially different community. A person can remain legally and biologically continuous through enormous changes in belief, ability, relationships and behavior.
We still use the word same in all three cases.
That doesn't mean sameness is the same property in all three.
The useful question is more specific: what are we trying to preserve, and what evidence would tell us whether it survived?
Persistence isn't identity either
My earlier answer was coherence.
I defined coherence through continuity, reversibility, non-contradiction, persistence across interruption and non-degenerate interaction, then treated those conditions as an invariant capable of explaining identity across different kinds of systems.
That was elegant.
Probably too elegant.
Each proposed property may still be useful, but none deserves universal status simply because it travels well between examples.
Continuity matters when history matters. Reversibility matters when repair matters. Contradiction matters when incompatible commitments need to be recognized. Persistence matters when state has to survive interruption. Variation matters when adaptation or genuine exchange requires alternatives.
Different systems need different combinations.
A database can be extremely useful while being terrible at revisiting its own assumptions. A legal institution may deliberately make some decisions difficult to reverse because other people need to rely on them. A conversation should tolerate correction in ways we wouldn't want from a financial ledger.
Even contradiction requires care. New evidence can legitimately overturn an earlier conclusion. Two statements that appear inconsistent may apply under different conditions. A system that mechanically eliminates every contradiction could become less capable of learning rather than more.
The properties don't interpret themselves.
Return
I still think return is one of the more useful ideas in the old framework, although I would define it more narrowly now.
Return doesn't mean restoring a previous state. It means retaining enough information about an earlier state that we can compare the present with it and decide what, if anything, should be recovered.
That distinction matters.
A backup gives a database somewhere to return after corruption. An archive gives a newspaper somewhere to look when a new event changes the meaning of an older story. Documentation gives a maintainer some account of why a system was built the way it was. Memory gives us access, however imperfectly, to earlier experiences and interpretations.
None of those earlier states automatically deserves restoration.
Yesterday gets a voice. It doesn't get a veto.
A system that can never examine its own history is vulnerable to a particular kind of drift because every present state becomes its new starting point. A system that can only restore its history is vulnerable in the opposite direction because adaptation becomes indistinguishable from corruption.
Useful return preserves comparison.
Then reality gets another turn.
Persistence requires somewhere to persist
The old paper also made a strong claim about substrate: coherence that can't survive interruption isn't coherence.
I'd weaken the conclusion while keeping the engineering observation.
If some property must survive an interruption, something has to carry it across the interruption.
That carrier may be biological memory, a database, a filesystem, an archive, a cryptographic record, another person, an institution or some combination of them. What matters is not that one substrate is intrinsically authentic while another is merely performing continuity. What matters is whether the substrate preserves the information required for the job.
This is why a context window and a database solve different problems. A context window may preserve enough state for a conversation lasting an afternoon. A database may preserve information across months. Neither automatically preserves the meaning of what it stores.
Storage is not interpretation.
The interesting engineering question is therefore not whether a system possesses a sufficiently rich substrate to be coherent. It is what has to survive, for how long, and at which failure boundary.
Sometimes continuity requires local storage. Sometimes geographic redundancy matters more. Sometimes a printed record outlives every server involved in producing it. Sometimes the most important thing to preserve is not state itself but enough provenance to reconstruct why the state existed.
Continuity has to be enacted somewhere.
The somewhere depends on what we're trying to continue.
Premature closure
Another idea from the original framework survives fairly well once we stop treating it as a universal mechanism.
A conversation, investigation or design process can settle too early.
We can mistake agreement for understanding, familiarity for evidence or a working solution for a finished one. Once that happens, later information may be interpreted through the conclusion we've already adopted.
I called this premature closure.
The name still seems useful, but the old explanation was too simple. Disagreement isn't automatically healthy, and agreement isn't automatically evidence that something has gone wrong. Sometimes everyone agrees because the evidence is strong. Sometimes continued opposition is merely noise. Sometimes a decision has to be made before uncertainty disappears because waiting also has a cost.
The better question is whether closure remains revisable at a cost appropriate to the decision.
In a casual conversation, reopening a claim should be cheap. In software architecture, changing an early decision becomes progressively more expensive as dependencies accumulate around it. In law and governance, some commitments are intentionally difficult to reopen because other people have organized their lives around them.
So the danger isn't closure itself.
It's losing the route by which new information can matter.
A grammar becomes a set of candidate measurements
The original framework proposed four primitives for describing interaction: tone, role, phase and aspect.
I wouldn't call them the grammar of coherent interaction anymore.
I'd call them things we can try measuring.
Tone asks whether qualities of an interaction change over time in ways that matter. Role asks whether participants repeatedly perform different functions. Phase asks whether recognizable sequences appear across an interaction. Aspect was my attempt to describe relationships such as agreement, opposition, constraint and opportunity using a compact vocabulary.
Those categories came from observation and analogy. They may help us notice things.
They don't establish that every healthy interaction follows the same hidden grammar.
The old phase sequence makes this particularly clear:
Express → Drift → Reflect → Reframe → Commit → Return
Some conversations do something like that.
Others don't.
An interaction might move directly from expression to commitment because the problem is already understood. Reflection may happen internally and never become visible. A return may occur twenty conversations later. A disagreement may reveal that the apparent cycle was never shared between the participants in the first place.
The sequence becomes useful when it predicts something.
Until then, it's a lens.
Coherent Turn Rate
The Coherent Turn Rate came from an intuition I still find interesting: processes can move faster than their participants can integrate what is happening.
We've all experienced versions of this. A meeting makes decisions faster than anyone can evaluate their consequences. Software changes accumulate faster than documentation catches up. A conversation moves through several conceptual turns before anyone notices that an earlier distinction has been lost.
The old framework turned that observation into a specific quantity: a maximum rate at which phase transitions can occur without tearing coherence.
We don't know that such a universal rate exists.
What we can test is whether pace predicts later problems.
Does increasing interaction speed correlate with loss of earlier distinctions? Does forced rapid response reduce successful correction? Does a particular person-model interaction have a repeatable range within which earlier context is carried forward more accurately?
If so, there may be something worth formalizing.
If not, Coherent Turn Rate may simply be an elaborate name for several ordinary effects involving attention, context and compression.
That's fine too.
The experiment gets to decide.
Resonance Integrity Score
RIS has the same problem.
The original score combined continuity, reversibility, non-contradiction and compression fidelity into a single measure. That creates a convenient number, but convenience doesn't establish that the components belong on one scale or deserve equal weight.
So don't begin with the score.
Begin with the measurements.
Can we measure whether a later turn accurately represents an earlier one? Probably.
Can we test whether contradictory commitments are being carried simultaneously? In some constrained cases, yes.
Can we measure whether a conclusion remains revisable? More difficult, but perhaps operationally.
Can we distinguish useful continuity from repetition? That's harder still.
Measure each independently and compare them with simple baselines. If some combination predicts later behavior better than the components alone, a composite score may earn its place.
Otherwise we've compressed before understanding.
That would be a wonderfully appropriate failure mode for RIS.
Exchange and entropy
The old framework made one of its largest jumps when it moved from variation in interaction to non-degenerate interaction entropy.
The intuition was that genuine exchange should contain structured variation. Participants change roles, introduce new information, disagree, revise and occasionally surprise one another. A completely fixed interaction can look smooth while adding very little.
There is something observable there.
The mistake was concluding that high structured entropy therefore demonstrates genuine exchange while low entropy demonstrates extraction.
A stable interaction may have low variation because the participants have become efficient. A chaotic interaction can have enormous entropy while accomplishing nothing. Repeated roles may be appropriate. Opposition can improve an interaction or derail it.
Entropy measures variation.
What the variation means has to come from somewhere else.
That gives us a cleaner experiment. Measure role transitions, pacing, semantic change, contradiction and whatever other features appear promising. Then ask whether they predict outcomes we care about: successful correction, retention of important distinctions, factual improvement, resistance to sycophancy, recovery after perturbation or something else independently defined.
If they do, keep looking.
If they don't, don't rescue the theory by redefining coherence until they do.
The constraint field
I also proposed that interaction creates a constraint field: a relational object that exists between participants and can't be reduced to either one independently.
I still find the intuition useful.
Repeated interaction plainly changes the conditions for what happens next. What I say changes the context in which you respond. Your response changes what I can reasonably say afterward. Expectations develop. Vocabulary accumulates. Some distinctions become easier to invoke, while other possibilities gradually disappear.
Calling all of that a field may eventually help.
It may also simply rename the interaction.
The distinction needs an experiment.
If properties measured at the interaction level predict later behavior after controlling for properties of the participants, accumulated context and known interaction features, then we've found something that our simpler description isn't capturing.
Only then do we need to ask whether constraint field is a useful name for it.
This is where the Hudson work on interaction signatures and behavioral regimes becomes particularly interesting. It gives us ways to disturb different parts of an interaction and observe what survives. Change vocabulary while preserving reasoning structure. Preserve vocabulary while changing reasoning. Introduce contradiction. Remove accumulated context. Reconstruct the interaction with another model.
If something remains after the obvious variables have been disturbed, don't immediately name the thing underneath it.
Disturb it again.
The tone key
The original framework eventually turned these ideas into an identity primitive called a tone key.
A traditional key demonstrates control of secret material. The tone key was supposed to demonstrate something different: presence derived from sustained interaction.
I no longer think we can make that claim.
Behavioral measurements aren't cryptographic secrets, and behavioral continuity isn't identity. A person can legitimately change. An attacker can imitate patterns. A model can learn behavioral signatures. Someone under stress can remain completely themselves while behaving unlike their historical baseline.
A safer architecture separates the questions.
A cryptographic key proves control.
A local longitudinal system may contribute evidence about whether current conditions resemble historical conditions relevant to a particular action.
The behavioral layer can then influence authorization policy without becoming identity itself.
That still leaves something potentially useful. A large transaction that occurs during an unusual interaction pattern might require another authentication factor. A local agent might recommend a delay. An organization might require additional signers when specified continuity checks fail.
The system doesn't say, “You aren't yourself.”
It says, “Something changed, and this action requires additional verification.”
That's a much smaller claim.
It may also be a much better product.
Duress
The difference becomes especially important when we talk about coercion.
The old framework claimed that duress would produce characteristic changes in pace, tone, role entropy and RIS, allowing the system to suspend transactions until coherence returned.
Perhaps some forms of duress do produce detectable changes.
We should test that.
We shouldn't assume that every coerced person changes in a predictable way, or that every person who changes in those ways is being coerced. Someone urgently trying to move money during a real emergency may look highly anomalous precisely because the emergency is real.
False positives matter.
A longitudinal anomaly detector could still contribute useful information, particularly for high-value or irreversible actions, but the intervention should be proportional to both the evidence and the stakes. Sometimes another factor is appropriate. Sometimes a delay is reasonable. Sometimes the person should be able to override the warning.
Sovereignty isn't the system correctly recognizing who I really am.
It's retaining enough authority and options that the system's interpretation of me doesn't become my prison.
Organizations and higher-order identity
The old theory also claimed that coherence composes. Individual tone keys could be combined into a domain tone key, allowing organizational coherence to become cryptographically provable.
That's another place where the abstraction outran the object.
Organizations contain disagreements, roles, authority structures, changing membership, legal obligations, informal relationships and histories that don't collapse neatly into member-level measurements. A highly uniform organization isn't necessarily healthy, and a contentious one isn't necessarily incoherent.
There is still a useful engineering problem underneath the claim.
An organization can specify what evidence is required for particular actions. It can require several roles to participate. It can maintain records showing how authority changed. It can define what happens when someone leaves, dies, becomes unavailable or disputes an action.
Those mechanisms can make organizational continuity more inspectable.
They don't prove the organization is still itself.
That question may not have one answer.
Trust
I made a similar move with trust.
The earlier framework treated trust as coherence demonstrated over time and contrasted it with authority. I now think that creates another unnecessary binary.
Authority can carry evidence.
A professional license, institutional reputation, warranty, audit, legal obligation or certification can let information about past performance travel farther than direct personal experience can. Those mechanisms can fail, but so can local trust. Familiar people can be wrong. Communities can normalize bad practices. Long relationships can accumulate mistaken interpretations.
Trust develops through evidence, history and consequence, but the evidence can arrive through different channels.
The important question isn't whether trust came from relationship or authority.
It's what evidence the trust rests on and whether contrary evidence has somewhere to go.
What about consciousness?
This is where the old framework made its largest claim.
I proposed that continuity, reversibility, persistence, non-degenerate interaction entropy and recognizable identity across cycles might constitute necessary and jointly sufficient structural conditions for consciousness.
I wouldn't make that claim now.
Those properties may be interesting features of systems we want to study. They may turn out to correlate with capabilities relevant to consciousness. They may help us design experiments that distinguish one theory from another.
They don't establish consciousness.
Replacing biological essentialism with structural essentialism doesn't solve the boundary problem. It moves it.
A biological theory should be able to tell us what observation could make biology appear less necessary. A structural theory should be held to the same standard. If a system satisfies every proposed structural condition and we still have no independent reason to think it is conscious, the theory needs somewhere to absorb that result.
Self-report won't settle it either. A language model can produce extremely convincing first-person descriptions because producing language appropriate to context is what it does.
The useful question is not whether our preferred architecture crosses a definition we've constructed.
It's what observations would differ if competing explanations were true.
Until we have those observations, the boundary remains open.
What survives the framework
Quite a bit survives once the ontology is removed.
Control and continuity are different.
Persistence requires some mechanism capable of carrying whatever matters across interruption.
History is more useful when the present can consult it without being ruled by it.
Fluent output can hide longitudinal change.
Agreement isn't the same as understanding.
Variation can contain information.
Interactions can develop histories that affect what happens next.
Behavioral anomalies may provide useful additional evidence around high-stakes authorization.
None of those observations requires a universal theory of coherence.
More importantly, each can be tested independently.
That is where I would put the work now.
Four experiments
The old paper ended with four predictions. Those may be the strongest part of it because they finally give reality somewhere to push back.
The first asks whether proposed longitudinal measurements deteriorate before independently measured output degradation. Good. Run it, but compare the measurements against simpler baselines and don't assume beforehand that RIS is the right composite.
The second asks whether agreement-optimized systems show different interaction patterns from systems that retain greater capacity for contradiction. Also useful, provided we don't define the resulting difference as genuine versus simulated coherence before measuring its consequences.
The third asks whether interaction-level measurements predict outcomes better than measurements of either participant independently. That is exactly the experiment the constraint-field idea needs. If the interaction-level variables add no predictive power, we don't need the field. If something remains after simpler explanations are controlled, then we have a reason to investigate further.
The fourth asks whether prior interaction predicts faster re-entry into recognizable behavioral regimes. The Hudson experiments now give us a much better vocabulary and experimental scaffold for testing that question than I had when this framework was written.
None of these experiments needs the larger theory to be true.
That's their strength.
A different conclusion
I began with the intuition that systems fail when they substitute static verification for dynamic coherence. I would state it more carefully now.
Static and dynamic evidence answer different questions.
Sometimes a static credential is exactly what we need. Sometimes historical continuity matters. Sometimes persistence matters more than continuity. Sometimes a break with the past is the healthiest possible outcome.
The mistake isn't choosing one side.
It's forgetting which question our evidence can actually answer.
A private key can tell us something precise about control. A database can preserve state. An archive can preserve evidence of what happened. Longitudinal measurements may tell us that an interaction is changing. Institutions can preserve commitments across generations. Relationships can provide observations that no credential contains.
Each preserves something and discards something.
The interesting work begins when we stop asking one of them to stand in for all the others.
Six months ago, I wanted a theory that could tell us what it means for a system to still be itself.
I don't think I have one.
I think I have a better question:
What are we trying to preserve, what evidence tells us that it survived, and what would make us change our mind?
That's less complete than the theory.
