Part II. Why Current Approaches Plateau

Where Generality Breaks

Foundational labs continue to bet that if they stretch generality far enough, domain sufficiency will emerge for free. That works in domains with wide acceptance regions, plentiful training signal, and low consequence for missteps. You can stumble through ecommerce support tickets or casual conversation without destroying anything. But the physics flips in hard, high-risk problems: the trajectories are long, failure probabilities compound, admissible regions are narrow, and a single wrong move invalidates the entire rollout.

Generality layers mountains of knowledge onto a tight cognitive core, harnessing universal reasoning patterns to pick high-probability paths. It vastly outperforms random search because it recognizes familiar surface patterns and leans on the shared heuristics of humanity. This is compression at work: we store fuzzy, overlapping templates that cover many situations and trust the core to interpolate the rest. Yet in critical domains the relevant patterns are rare, highly specific, and often look like noise in the aggregate. Worse, the compressed template can actively mislead; a maximal-likelihood step under the noisy match might be the one move that invalidates the arc contract for this cohort. The borrowed knowledge becomes poison because it keeps firing transitions whose sufficiency was never measured. When you must hit sufficiency on every step, a chain of best guesses guarantees eventual failure. Multiplying even tiny error rates across hundreds of decisions drives the success probability toward zero.

The escape route is not more generality; it is compositional causality. Use general reasoners as search primitives to generate hypotheses, but immediately squeeze out correlations, retain only the causal pathways, and encode the resulting habits into guarded arcs. Hard problems demand systems where measurement proves every link of the causal chain, replay regenerates statistics under updated blueprints, and orchestration refuses to enter an arc unless the sufficient-statistic contract is airtight for that cohort. Only then does domain sufficiency stay intact under high risk.

Consider acute sepsis management. Traditional decision support might spot “possible sepsis” from a few vitals and escalate automatically. A compositional system keeps the dimensional blueprint for that patient current—tracking lactate trajectories, fluid responsiveness, ventilation status, and consultant availability—before allowing the resuscitation arc to fire. Without that patient-specific blueprint, the same escalation pattern can trigger inappropriately, exhausting ICU capacity or delaying care for the cohort that actually matches the validated contract.

The Energy Investment Problem

The current era's dominant approach focuses on scaling generality—broader data mixtures, longer reasoning trajectories, denser models—but seldom rewrites the blueprint that grounds those capabilities. That blueprint is the measurement plan for the patient or asset we are optimizing; when it stagnates, we saturate the watermark of a domain by chance, not because the system actually measures the object's decisive dimensions.

Energy investment and compute requirements therefore grow exponentially while the blueprint remains static. The added capacity keeps every arc warm, regardless of whether the object's state justifies it, so we pay multiplicative costs to push linear gains.

The alternative is to rein in that combinatorial explosion: penalize unnecessary reasoning tokens, quantize long trajectories into ledgered arcs, and reward only the compositions that demonstrably move the object's coordinates within the blueprint-defined sufficient-statistic space toward their target sets. That is where compositional approaches recover efficiency.