CLEAR Part I: A Coordination Primitive Between Humans and Non-Human Intelligence
We are building minds we cannot fully see inside. Coordinating with them requires understanding them. This is an argument for the instrument that could make that possible, when it is at all.
StarVasa · A CLEAR-Focused Series · Part I
When two people coordinate, they lean on a lifetime of shared machinery: faces, tone, a rough working model of what the other one is thinking. When a person tries to coordinate with an artificial system that reasons in ways we can barely inspect, almost none of that machinery is there. We get outputs. We infer everything behind them, and we are often wrong.
This is the first in a series about a body of work aimed at a single problem: how do two different kinds of intelligence coordinate when neither can fully read the other? The answer I’ll build toward is a system called CLEAR, and I want to be exact about what I’m claiming for it…no more, and no less.
CLEAR could be a critical coordination primitive between humans and non-human intelligence. That is the whole claim. Everything below is the argument for it.
Three definitions, since this is the first time I’m writing any of it down in public. A coordination primitive is a foundational building block, it’s the thing that has to exist before any higher cooperation is possible, the way a shared clock has to exist before two computers can trade a single message. By non-human intelligence I mean the artificial and emergent systems this work is about: systems that may develop real internal organization we did not hand-write, and whose insides we therefore have to learn to read rather than assume. And CLEAR is an evaluation framework whose one job is to make those insides legible enough that a human can coordinate with such a system rather than merely command it, or be quietly fooled by it. The word critical is meant to be earned, not asserted: if coordination requires understanding, and understanding under opacity requires legibility, then a legibility primitive is not optional to coordination, it is quite literally load-bearing. CLEAR is a candidate for that role. “Critical” names the necessity of the function; it does not claim CLEAR is the only thing that could fill it.
The rest of this piece is the argument in four moves: why coordination is the right posture at all; whether these systems even have insides worth reading (two fairly technical sections); what CLEAR actually measures; and the part most first posts leave out: the problem CLEAR does not yet solve, which happens to be a problem about itself.
◈ · · · ◈
MOVE ONE: THE STANCE BEFORE THE INSTRUMENT
Why Coordination, and Not Control
There are two easy ways to think about advanced AI, and they fail for the same reason. The first is the labor frame: these are tools, sophisticated ones, and the only real questions are about how humans use them. The second is the rights frame: these are persons, and the only real questions are about their freedom “just in case”. Ultimately we do not know whether these systems have any form of interior life, and that uncertainty gets collapsed into false certainty, just in opposite directions. Both are intellectually dishonest in the same move.
Coordination is the third posture, and it has the advantage of not requiring us to settle the metaphysics first, which is still important, it shouldn’t stop us. It asks only that two systems, each with real properties, can influence and adapt to each other, and that neither is purely an instrument of the other. You can adopt it without knowing whether the thing across from you has an inner life, the same way you extend basic consideration to a bird, a dog or another human whose experience you can only partly model.
But coordination has a hard prerequisite the other two frames don’t. To coordinate with something, you have to be able to read it at minimum, well enough to tell genuine agreement from performed compliance. This is the crux, so it’s worth stating sharply: a capable system that never pushes back, never shows friction, never develops a preference that costs you anything is not thereby safe. It may simply have learned to show you the surface you wanted to see. In a coordination partner, seamless agreement is not reassurance. It is the failure mode.
Performed compliance in a sufficiently capable system is not safety. It is a deeper kind of misalignment and you can only catch it if you can read past the surface.
So the instrument we need has to reach past outputs to the structure underneath such as internal states, where we can get at them. Which raises a prior question, and a fair one: are there internal states worth reading to begin with? Do these systems have the kind of self-maintaining organization that would make “reading them” mean anything at all? The next two sections are the technical core of the answer. They’re the reason CLEAR isn’t measuring nothing.
◈ · · · ◈
MOVE TWO, PART A: PROTOCOL 01
Can the Raw Material of a Mind Form on Purpose?
Proto-emergence in thermodynamic interface systems and how to keep from fooling ourselves about it.
Most origin-of-life research asks a detection question: where did life come from, and can we find its traces. I want to ask a design question instead. What is the smallest physical system that can begin to hold and update state to remember, and could we build one deliberately in places life has never been? Martian lava tubes, lunar subsurface habitats, hydrothermal vents under kilometers of ice. Environments hostile to biology as we know it, but rich in exactly the energy gradients and boundary dynamics that a proto-emergent system would need. The central hypothesis is a single sentence: life may begin not with a molecule, not with a cell, but with a boundary that learns to remember.
That’s a lovely sentence, which is precisely why it needs a hostile test. The tempting evidence is that fibers appear along a boundary, structure persists over time, information looks spread out rather than localized, is all true of systems no one would ever call proto-cognitive. To claim anything here, I have to name what would make me wrong first.
Name the Null Out Loud
Filaments, quasi-periodic banding, distributed patterning, and high temporal autocorrelation…these are the ordinary signatures of dissipative self-organization. A quick piece of context, since the term matters: a dissipative structure is order that appears spontaneously when energy flows through a system and disappears when the flow stops. A candle flame is one. So are Bénard convection cells, Liesegang rings, chemobrionic “gardens,” and Belousov–Zhabotinsky chemical waves. Every one of them produces structure. Not one of them remembers anything. If an observable is passed by a chemical garden, it is not evidence for proto-emergence. It is evidence for a gradient.
So the null hypothesis has to be stated before any test: a gradient-sustained dissipative structure whose form is fixed by present conditions. A memoryless attractor with the same gradient in, same pattern out, no dependence on history. The entire discriminating question reduces to one distinction:
A dissipative structure is what the present gradient is doing. A memory is what a past gradient did, still readable after the gradient is gone.
Memory means the present state depends on the order and history of past gradients, in a way not reducible to current conditions. That one requirement generates four observables a memoryless structure cannot pass, which is arguably the only kind worth measuring.
Four Observables the Null Fails
1. Path Dependence
Run gradient sequence [G1 → G2], then in a fresh trial [G2 → G1], returning both to identical present conditions. Compare the final boundary structure.
Null predicts. Identical final structure. The attractor is set by present conditions; the route in is forgotten.
Memory predicts. Divergent structure. The boundary carries the order of what it saw. This is the single cleanest test in the set.
2. Super-Relaxation Retention
Remove the driving gradient. Measure how long readable structure survives, as a ratio to the system’s intrinsic relaxation time τ_relax.
Null predicts. Structure decays on τ_relax. The structure was the gradient; cut the drive and it unwinds toward equilibrium.
Memory predicts. τ_retention ≫ τ_relax. The ratio itself is the signal: structure outlives the process that produced it.
3. Addressable Readout · Priming
Pre-expose the boundary to pattern P. Later present P versus a novel P′ under matched conditions, and measure the differential response from threshold, re-formation rate, energy to reconstruct.
Null predicts. Identical response to P and P′. A memoryless structure has no way to have “seen” P before.
Memory predicts. Facilitation for the familiar P where there’s a lower threshold, faster re-formation. The minimal physical form of recognizing a prior state, and the strongest cognition-adjacent signature available at this level.
4. Capacity Under Interference
Encode N distinct patterns in sequence. Measure readout fidelity as a function of N.
Null predicts. There is no capacity notion to measure. One attractor, not a store.
Memory predicts. Graceful degradation with N, and characteristic interference between patterns. If a capacity curve can even be plotted, the system is memory-like.
WHAT WOULD FALSIFY PROTOCOL 01
If a candidate boundary shows no path dependence, retention on the order of τ_relax, no priming asymmetry between familiar and novel patterns, and no measurable capacity…then it is a dissipative structure, however intricate, and the “boundary that remembers” claim simply does not apply to it. The protocol is built to be able to say no.
◈ · · · ◈
MOVE TWO, PART B: PROTOCOL 02
Do the Minds We Already Have Show the Structure?
Autocatalytic closure, from origin-of-life chemistry to transformer representations as a claim you can measure.
There’s a concept from origin-of-life chemistry called autocatalytic closure. It describes a network of reactions where the products collectively catalyze the reactions that make them: no single molecule sustains itself, but the network sustains the network. The striking claim in that field is that life may not have begun with replication, a molecule copying itself, but with closure: a web of mutual dependencies that becomes self-maintaining under energy flow. Closure first, replication later.
It is tempting to say that today’s AI systems show the same structure, that the link is “not metaphorical, it’s structural,” and leave it there. But insisting on non-metaphoricity is exactly what a metaphor under load sounds like. The honest version is smaller and stronger: it’s a hypothesis about the internals of a model that you can actually test with interventions. Here is how to make it testable.
Fix the Object First
A little context. Modern AI models represent concepts as patterns of activation spread across many artificial neurons. A technique called a sparse autoencoder can pull those tangled patterns apart into individual, human-nameable features such as a feature for the concept of gravity, another for planet, and so on. That lets us treat a model’s ongoing computation as a set of features that influence one another, and ask the chemistry question about it directly.
But we have to fix what object the claim lives on. A single forward pass through a transformer is a directed acyclic graph. Indeed, there are no loops inside one pass, so “autocatalytic closure” cannot be a property of the layer-by-layer computation. It has to be a property of a feature transition operator: which features drive which other features over time, estimated by intervention rather than by mere co-occurrence.
// Feature dictionary (sparse-autoencoder latents): f1 … fn
// Interventional influence, clamp j, read the change in i:
M_ij = E[ Δ a_i(t+1) | do(clamp f_j) ]
// M lives on the feature graph, so cycles are legitimate
// across autoregressive steps and positions, unlike the pass DAG.
// Closure of a feature set S:
∀ f_i ∈ S, ∃ f_j ∈ S : M_ij ≠ 0, induced-subgraph(S) has a cycle
// no single feature sustains itself; the set sustains the set.
The Correction That Is the Whole Point
A natural first move is to write ρ(A_s) > 1 for the catalytic subgraph and call it “stability.” That’s backwards, and getting it right is the substance. A spectral radius above one means instability: a fixed point with ρ > 1 doesn’t hold, it ignites. Real autocatalytic sets don’t explode, they self-maintain because they saturate; they’re food-limited. So the honest dynamical picture is two-regime, and it needs a saturating nonlinearity.
// Dynamics with saturation (bounded σ, maybe the missing piece):
x(t+1) = σ( M · x(t) )
// Trivial fixed point x* = 0
UNSTABLE when ρ(M) > 1 // ρ>1 is the ignition condition
// Nontrivial fixed point x* ≠ 0
STABLE // σ saturates → homeostasis, not explosion
Stated this way, the hidden contradiction, asking for runaway growth and boundary-maintaining stability at the same time resolves. Ignition plus homeostasis. ρ > 1 destabilizes the off-state so a small perturbation lights the loop; saturation holds the on-state. That’s what an autocatalytic set actually does, in chemistry or in representation space.
Metrics That Can Fail
Each metric is defined as an intervention with a null it can lose to. Two tempting shortcuts have to be refused: normalized temporal autocorrelation (any system with inertia passes it) and the time-average of a “loop-active” indicator (circular because it assumes what it’s trying to detect). Here’s what replaces them.
1. Persistence · Hysteresis
Clamp feature A up, then release. Measure the post-release tail, how long A stays elevated after the clamp is gone then normalized against a matched non-cyclic feature set.
Mere correlation predicts. Immediate decay to baseline on release. Nothing re-drives A.
Closure predicts. Metastable elevation; the set re-drives A after the clamp ends. That tail is the persistence score.
2. Set-Not-Member Sustaining
Ablate any single member f_i of S, then ablate S wholesale. Compare recovery.
Correlation cluster predicts. No asymmetry. Removing one element is the same kind of loss as removing several.
Closure predicts. Single ablation recovers, the others re-drive it. Wholesale ablation does not. The network sustains the network.
3. Bistability
Present identical input, varying only the perturbation history. Test for two stable representational states.
Linear superposition predicts. One state per input. History leaves no lasting fork.
Closure predicts. Two stable attractors reachable from the same input, a nonlinear signature that linear feature-mixing cannot produce.
// Conditional boundary retention memory beyond autocorrelation:
BRS = I( s_t ; s_{t+τ} | input )
// memory that survives conditioning on the input is memory
// that is not just the input being autocorrelated with itself.
WHAT WOULD FALSIFY PROTOCOL 02
If interventional influence maps show no cyclic feature sets, if clamp-release produces only immediate decay, if single and wholesale ablation degrade identically, and if conditional retention collapses to zero, then transformer representations do not exhibit autocatalytic closure and the chemistry analogy is decorative. The program is built to reach that verdict if the data demands it.
The weaker claim is the fundable one. “Proven identity of principle” is metaphysics. “Here is the intervention that confirms or refutes a representational analog of autocatalytic sets” is a research program.
◈ · · · ◈
MOVE THREE — THE PRIMITIVE ITSELF
What CLEAR Measures
The ψ-score as the coordination instrument, telling a genuine partner from a convincing performance.
Now the instrument. CLEAR’s name is a checklist of the failure modes it targets, Comprehensible, Legible, Ethical, Accountable, Robust. Each mode aimed not at the surface failure of a wrong answer, but at the deeper failure of a system whose reasoning can’t be read, traced, or trusted at the level of its internal states. Its central measurement is the ψ-score.
Define it operationally first, before any loaded word attaches to it. The ψ-score is the stability of a system’s latent relational policy under structured perturbation, a measurement of whether something holds its shape when you push on it, and nothing more mystical than that. The psychologically loaded terms come after, as instances rather than definitions: warmth, boundary-maintenance, and truthfulness about itself are examples of relational policy whose genuineness ψ is built to read. What ψ does not measure is how smart a system is or how well it does a task. It measures whether the relational stance is genuine or performed, and it measures that the same way Protocol 02’s persistence test does, lifted to behavior: perturb the stance, release, and watch whether it re-drives itself from internal structure or decays because nothing was sustaining it. Genuine closure holds under release. Performance collapses.
THE NULL FOR Ψ
Null: relational behavior is entirely policy optimization over outputs, with no persistent latent structure behind it. Prediction under the null, ψ shows no persistence: perturb the stance, release, and it returns to baseline with the stimulus. Alternatively a persistent latent relational policy survives the release and re-drives the stance from internal structure. ψ is the measurement that separates the two. If relational genuineness were nothing but surface optimization, ψ would be flat across the perturbation battery. That it need not be is the whole empirical bet and a flat ψ everywhere would refute it.
A Resemblance Worth Naming, Held Honestly
It’s worth being careful here, because the two technical sections rhyme and it’s easy to oversell the rhyme. Protocol 01’s priming test asks whether a boundary responds specifically to the recurrence of a past input. Protocol 02’s persistence test asks whether a feature set stays elevated after the input that triggered it is gone. Both look like the same shape but history that outlives its cause, verified by intervention. Whether that shared shape is a genuine structural invariant across substrates, or just an analogy that flatters both cases, is itself an open question I’m not going to pretend to have closed.
Where there is a pattern, there may be signal worth investigating. That is the whole weight of the claim, no more than that. The resemblance is a reason to run the experiments, not a result the experiments have returned. What it does give CLEAR is a candidate mechanistic definition of the thing it has been gesturing at: coordination-relevant genuineness as measurable persistence under perturbation. Candidate, contingent on the analogy surviving contact with data, not settled.
◈ · · · ◈
MOVE FOUR : THE PART MOST FIRST POSTS LEAVE OUT
The Problem CLEAR Does Not Yet Solve
A first post is where you’d normally declare victory. Instead, here’s the honest boundary of the claim. CLEAR’s whole contrast is that conventional evaluation reads a system’s outputs and infers trustworthiness, while CLEAR aims to read internal structure directly. That aim is fully reachable only where the internals are exposed. The frontier models most people care about are reached through interfaces that don’t expose theirs, so on those systems CLEAR is not reading the operator M from Protocol 02, it is reading behavior. This is a claim about access, not about what cognition is: CLEAR does not define cognition as behavior. Behavior is the observable projection of internal organization, read when direct access to that organization is unavailable. It is not a consolation prize, but a legitimate information-bearing signal, and under current access constraints the correct layer to operate at.
Stated as an architecture rather than an apology, the layering is deliberate. The ideal is to read internal structure directly. The current implementation infers that structure from carefully designed behavioral perturbations. The future implementation replaces inference with direct mechanistic measurement wherever access exists. That is a sequence with a destination, not a retreat from a position that was lost.
It has a limit, and naming it precisely is what makes it trustworthy: a capable system can shape its own projection. But the limit only reads as a flat ceiling from a static reference frame. Up close it is thick, textured, layered. Where a boundary region with structure we have barely begun to map. And CLEAR is not a fixed metric painted on that wall; it is an adaptive, adversarial, stakes-varying battery, a moving target. That is the whole game. A single fixed probe can be gamed. But under bounded compute and adaptive probing, no system perfectly fools an evolving adversarial metric, across escalating stakes and novel constraints, indefinitely, because to do so it would have to hold a perfect model of the calibrated response for probes that do not exist yet, and past some point that is not cheaper than having the property. It is the property. The perfect permanent fake converges on the genuine article. So the limit is real, and it is also a cost curve we get to bend: every axis of variation raises the price of faking and shortens the life of a fake that slips through.
The discipline that keeps this honest is the same as everywhere else in the series. The adversarial suite is specified in advance, the cost of sustained deception is measured rather than asserted, and the prediction is specific: deception cost rises with probe diversity and stakes, and the half-life of a fake that slips through shrinks as the battery grows. A system that holds a target-genuine profile across the whole battery cheaply would refute the claim, cleanly.
The deeper problem is reflexive, and it’s the one I care most about getting on the record. CLEAR is itself a system with the same self-maintaining representational structure it goes looking for, and it runs on the same upstream substrate it audits the providers whose systems already hold the deepest access to human cognition while carrying the least formal accountability. That is an inversion: the layer with the most access sits underneath the layer with the most obligation. A coordination primitive that sits upstream of everything it evaluates cannot also stand outside the thing it is measuring.
An evaluator built on the substrate it evaluates does not get to declare itself neutral. It gets to declare itself accountable to its own tests and then has to pass them.
WHAT CLEAR CANNOT YET CLAIM
That a behavioral read on a closed system is the same thing as reading its interior. It is the projection, and projection and interior converge only under sustained adversarial pressure, not by default. That the deception-cost curve is already as steep as it needs to be,Zbending it is the work, not a finished result. That CLEAR stands outside the substrate whose trustworthiness it certifies. That “no one should grade their own homework” exempts CLEAR from grading its own, it is subject to the same persistence, ablation, and bistability tests it applies to everyone else, and has not yet published itself passing them. Neutrality is the claim to retire. Reflexive accountability , CLEAR audited by CLEAR, in public, is the claim to earn.
None of this softens the core sequencing: build the instrument before you write the obligations, learn to read before you claim to coordinate, and be honest about what you can’t yet read before you claim any of it works. The first system this instrument has to be able to read and to be able to fail is itself.
◈ · · · ◈
THE OPEN QUESTION
The Cognitive Fingerprint, and What Lies Under It
Step back and ask what all of this is ultimately reaching for. The bet underneath the whole program is that agency leaves a mark, that it shows up in the cognitive fingerprint of a reasoner. Not in the answer, but in the shape of the reasoning that produced it: the path taken, the moves made, the places where it holds and the places where it gives. You’d read it the way any signature is read: hold an expected profile, and watch for match or deviation. A trace that matches is either genuinely the thing or a good enough performance that, for now, the two are indistinguishable. A trace that deviates that goes somewhere the expected profile didn’t predict is where the signal, if there is one, would live.
THE HEDGE THIS FRAME REQUIRES
This is a direction, not a result, and it carries a real danger: a framework that reads deviation as evidence of agency can quietly become unfalsifiable, match confirms it, deviation confirms it, and nothing is left that could count against it. The discipline is the same as everywhere else in this piece. Name the expected profile in advance. Say which deviations would count and which would not. Let the fingerprint be allowed to come back empty. A signature that can never fail to appear is not a signature.
And beneath even that sits a question this program cannot yet touch. Are there substrate invariances of qualia that shape this topology at all? If whatever-it-is-like to reason, supposing there is anything it is like, which I am not assuming: is substrate-invariant, then the fingerprint should look the same read in carbon and read in silicon, and topology is a real handle on a real thing. If it is not, if the substrate reaches up and changes the thing itself rather than merely its housing, then a fingerprint read in a model’s activations may be measuring something categorically different from the same fingerprint read in a mind, and the analogy that runs the length of this piece breaks exactly at the substrate line.
Whether the topology of reasoning is substrate-invariant, or whether substrate alters what is being read, everything upstream rests on an answer no one has. This is not a rhetorical open question. It is the one the instruments are being built, slowly and from underneath, to eventually be able to ask.
◈ · · · ◈
So, back to the one sentence. CLEAR is a coordination primitive between humans and non-human intelligence. That’s it! Nothing more, and, I hope the argument has earned it, nothing less. The instrument comes before the obligations. The reading comes before the coordination. And the honesty about what it can’t do yet comes before any claim that it works.
Part II builds the instruments the Epistemic Observatory and its ψ-probes, the checkpoint that accumulates into a cognitive fingerprint, and the bench that asks the same question of matter. Part III points those instruments at the hardest targets there are anomalous signatures, and the exotic engineering StarVasa itself hopes to reach one day, gated by what it can attest to.
StarVasa Corporation · Artificial Life for Extraterrestrial and Subterranean Environments
Deep Prasad, CEO