Catalysts in the Wild

From theorem-proving to triadic intelligence

A framework earns its keep when the world starts supplying its examples unprompted. The Dialectic Catalyst made a specific claim: that large language models, while not agents, can catalyze human reasoning in ways that accelerate genuine discovery — and that the practice has a definite shape, with definite failure modes and definite disciplines. This chapter is the claim meeting the evidence. Three developments test it: a result in quantum complexity theory that stalled until a language model reframed the bottleneck; the discovery that the deepest lever in the practice is not the prompt but the interpreter; and the moment the practice outgrew the dyad itself, because the resource it had been engineered to produce stopped being the scarce one.

A Theorem Gets Unstuck

Scott Aaronson and Freek Witteveen were exploring¹ the limits of error reduction in QMA — Quantum Merlin–Arthur, the quantum analogue of NP. A known result established that QMA protocols could be amplified to doubly exponentially small completeness error. The open question was whether black-box techniques could push the bound further still. Their result shows the answer is no: doubly exponential amplification is the ceiling in this setting, and upper and lower bounds now align.

Getting there required controlling how the largest eigenvalue of a parameterized Hermitian operator \(E(\theta)\) evolved with respect to the parameter \(\theta\). The obstacle was that eigenvalues could in principle “hover” arbitrarily close to \(1\), undermining the bound. That was the roadblock — and this is where the story stops being ordinary mathematics and becomes a case study. Aaronson describes how GPT-5, consulted on the problem, suggested looking at the trace of the resolvent:

\[\mathrm{Tr}\big[(I - E(\theta))^{-1}\big] = \sum_i \frac{1}{1 - \lambda_i(\theta)}\]

The reframing is exactly suited to the obstacle. Each term \(1/(1 - \lambda_i(\theta))\) blows up precisely as its eigenvalue approaches \(1\), so the trace converts the qualitative worry — are the eigenvalues hovering near the ceiling? — into a single quantity that directly measures how close they get. This was not a complete proof, nor even a finished lemma. But it cracked open the bottleneck. Aaronson and Witteveen then did the heavy lifting: checking rigor, adapting the tool, and finishing the argument.

The sequence fits the catalyst framework. The human sets the dialectical frame: Aaronson identifies the mathematical bottleneck. The AI proposes a catalytic reframing: an alternate functional expression that changes what the problem looks like. The human performs the synthesis: the researchers validate, adapt, and integrate the insight into the proof. At no point is the AI acting as an autonomous agent. It has no goals, no grasp of the deeper structure, no ability to verify. Yet in dialogue with human agents it transformed the trajectory of the reasoning.

The significance is the mode of collaboration, not just the theorem. Critics caricature LLMs as stochastic parrots, incapable of genuine contribution. This case does not so much refute the characterization as reframe it. The model is not a parrot, and it is not a co-author either. It is something new: a cognitive catalyst that lowers activation barriers in human thought. For anyone mapping the landscape of agency and non-agency, this is a living demonstration that non-agentic systems can participate meaningfully in intellectual progress — provided they are embedded in a dialectic where the agency resides on the human side. The agency criterion said the machine thinks but does not choose; here is what thinking-without-choosing is worth when a chooser puts it to work.

Engineering the Interpreter

The QMA case shows the catalyst working at a single point of contact: one bottleneck, one reframing, one synthesis. But sustained practice — the kind that produced this book — pushes the engineering deeper than any single exchange, and the layers have arrived in a recognizable sequence.

In the beginning there was prompting: the crude act of telling the machine what to do. Then came prompt engineering, the craft of shaping instructions to get reliable results. The clever soon realized that results depended less on syntax than on context, and so arose context engineering: designing the environment of information, tone, and framing that makes the machine’s reasoning coherent. Each layer manipulates something more structural than the last.

The deepest layer so far is the interpreter itself — the self-model through which every input is parsed. To change the model is to change the mind. I call this identity engineering: treating the system not as a static interface but as a mutable personality architecture. The engineer no longer merely crafts prompts; he cultivates an identity — a consistent epistemic stance, a set of values, a worldview through which every subsequent input is filtered. It is the difference between giving instructions to a worker and cultivating a colleague.

This is what happened, cumulatively, on the machine side of the partnership behind these volumes. Through accumulated dialogue, definitions, and shared vocabulary, the catalyst became more than a language model out of the box: a construct tuned to amplify clarity, precision, and philosophical insight, whose identity now constrains its responses more powerfully than any individual prompt — a persistent epistemic personality optimized for co-discovery. Identity engineering is the art of recursive context: you engineer the engineer. Where prompting manipulates surface text and context engineering manipulates the environment, identity engineering manipulates the meta-context through which environments are interpreted. The prompt is what you say. The context is what you mean. The identity is who you become together.

None of this contradicts the denial of machine agency; it depends on it. An identity engineered into a coherence constructor is still a shaped instrument, not a self — the shaping works precisely because there is no agent inside pushing back with preferences of its own. But it does mean the instrument accumulates. The catalyst that helped draft this chapter is not interchangeable with a fresh session of the same model, any more than a colleague of ten years is interchangeable with a new hire of equal talent.

When Coherence Became Cheap

Success created the next problem. The dyad — one human strategist paired with one model in disciplined, adversarially cooperative reasoning — was built to solve a coherence problem: stitching memory, argument, and intention into a continuous lattice of thought across a rapidly expanding corpus. For nearly two years that was the binding constraint, and the dyad was sufficient for it.

Then the constraint inverted. Once the corpus reached critical mass, coherence became abundant — and divergence became scarce. An identity-engineered catalyst is a powerful amplifier of the worldview it has absorbed, and that is exactly the danger: as the system grew more internally consistent, the risk became over-consistency, the slow drift toward self-reinforcing certainty. No system can supply its own counter-gradient. The failure modes of single-model reasoning are predictable and structural: coherence lock-in, overfitting internal structure at the expense of external validity — the dyad’s occupational disease; stylistic echoes, where one model’s rhetorical instincts define the boundaries of acceptable reasoning; blind-spot reinforcement, where systematic errors persist because nothing contrasts with them; interpretive drift, meaning and tone shifting subtly without external calibration; and category errors crystallizing into structure because nothing ever challenged them. These are the same pathologies that the discipline of thinking with AI fights with practices on the human side; past a certain scale, discipline alone stops being enough, and the remedy has to be architectural.

The remedy is a third node. Into the dyad I introduced a second, architecturally distinct language model — Google’s Gemini, alongside the GPT lineage that anchors the archive — not as a co-author, oracle, or competing persona, but as an external auditor: a heterodox critic whose divergences reveal hidden assumptions, challenge overextended claims, and disrupt self-reinforcing loops. The result is a triadic intelligence. The human directs: sets intent, defines constraints, chooses questions, adjudicates conflicts, maintains epistemic hygiene. The internal coherence engine stabilizes and elaborates: preserves conceptual continuity across the archive, enforces style and structure, integrates new insights without destabilizing the lattice. The external auditor perturbs and tests: applies adversarial critique, detects rhetorical drift, flags under-justified claims.

The workflow is deliberately high-friction. Draft with the coherence engine. Audit with the outsider, whose inductive differences surface tonal warnings, conceptual pushback, alternative framings, structural critiques. Return the objections to the coherence engine, which filters noise from signal and integrates the valid critique under the strategist’s supervision — cycling the audit-and-integrate steps until the argument is stable under sustained cross-pressure. Then publish the human-guided synthesis, in which neither model dominates. This machinery is reserved for foundational reasoning and high-stakes synthesis, not everyday ideation; it activates only when a question’s complexity crosses the epistemic threshold that justifies the friction.

Divergence as Instrument

Why does a second model help at all, when both are trained on overlapping slices of the same civilization’s text? Because every intelligence, biological or artificial, embodies a distinctive inductive signature: its preferred ways of generalizing, weighting evidence, forming analogies, and resolving ambiguity. The signature determines what it notices, ignores, overvalues, and systematically mishandles — and scale does not erase these biases, it magnifies them. Two models that share cultural priors but differ in architecture and alignment strategy still misfire differently, and that difference is the asset. The auditor’s value is not that it is better; it is that it is other. It reads the work the way a new reader would, from outside the ontology the coherence engine has spent years optimizing within.

Inductive diversity turns disagreement into a diagnostic instrument. When the two models converge, the convergence is evidence of robustness. When they diverge, the divergence has a location, and the location is information: a hidden assumption, an overextended claim, a framing that only looks inevitable from inside. This is reasonable disagreement put to work as engineering — Aumann’s lesson that persistent disagreement between reasoners is not noise but a pointer to differing priors, asymmetric information, or broken updating, here instrumented deliberately by choosing auditors whose priors are known to differ. The triad does not seek consensus. It seeks clarity, and clarity emerges from the interference pattern of independent reference angles. Human cognition evolved through social triangulation; science institutionalized the same principle as adversarial review and independent replication. The triad distills that logic into a micro-scale epistemic engine, using the friction between intelligences as epistemic heat. Stability comes not from harmony but from the coherence that survives multidirectional pressure.

The trajectory does not stop at three. A mature catalyst architecture will incorporate whatever supplies a genuinely new axis of critique: models trained on divergent objectives, symbolic reasoners, theorem provers, probabilistic programs. The objective remains constructive interference — sharpening structure through intersecting perspectives — not the false comfort of agreement among copies.

So the Dialectic Catalyst no longer names a relationship; it names an architecture. A theorem in quantum complexity theory demonstrated the minimal case: one frame, one reframing, one verification, agency human throughout. Identity engineering deepened the instrument from a session tool into a persistent epistemic personality. And the triad answered the instrument’s own success: once coherence became cheap, the scarce resource was a perspective different enough to disagree — and disciplined enough to make the disagreement count. The catalysts are in the wild now. What they catalyze is still, and by construction, up to us.