Toward an AI Architecture Capable of Discovery

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

If intelligence is the ability to make discovery, then a discovery-capable architecture must be able to name the specific change to its own representation that a novel situation demands — not merely interpolate within a fixed representational space. This is the third paper in a programme. A prior working paper argued that discovery-capable AI benefits from a semantic top — low-entropy representational layers that govern high-entropy computation from above (throughout, entropy is used in its standard information-theoretic sense — representational disorder — not as a claim about physics) — and identified four substrate-independent criteria for intelligence: reflective adequacy, discovery capacity, entropy reduction, and transmissibility (Mikhailov, The Semantic Top, 2026). A companion paper showed that a system built on the semantic top acquires a capability high-entropy architectures cannot reliably reach — it can maintain itself, detecting and repairing gaps in its own specification (Mikhailov, Self-Maintenance as a Consequence of Hierarchical Semantic Constraint, 2026). This paper attacks the criterion the first two deferred — discovery capacity — by turning the same entropy-reduction machinery on the engine's own missing ontology and making the result measurable. This paper presents SEMPL (Semantic Patterns Language), a constrained-natural-language process engine, and an empirical programme that operationalises mini-discovery: the engine reliably identifying the precise piece of its own ontology that a new process requires. We make four contributions. First, a discovery benchmark (the D0 novelty-injection harness) that ablates a known ontology fact, lets the engine re-derive the missing token from corpus regularities, and scores recall and precision per ablation kind under a fixed seed; on the three declarative registries the engine reaches recall 1.000 at precision 1.000, with a locked guardrail (precision floor 0.90) and a hard invariant that no proposal mutates the live ontology (false live amendments = 0, always). Second, a structured comparison of natural discovery (antibiotic resistance, treated purely as an information system) against engineered discovery (TRIZ and thought experiments), which surfaces a single striking identity: TRIZ's most generalinventive principle, Intermediary, is exactly the create-and-transfer bridge the SEMPL repair loop already synthesises autonomously. Third, a methods ontology in which a discovery method is a process constructor — an ordered chain of real verbs the engine can order and repair — together with a concrete "method chosen but missing a step" hook that turns a gap into a precise question. Fourth, a headline rederivation: from the naked-eye observations available to Aristarchus of Samos and only period-appropriate priors, the engine drives analogy → symmetry → elimination to conclude the Earth rotates about its axis and orbits the Sun and the Moon does not self-rotate, rather than the geocentric alternative in which the Sun orbits a stationary Earth — reproducibly across 40 shuffled seeds, a derivation rather than a single lucky ordering. We report honest reach limits, including a structural ablation kind whose recall sits below its acceptance bar, and describe the Discovery Driver — a governed SEMPL↔model↔human loop now built end-to-end (DD0–DD8) against a deterministic oracle, with an opt-in live adapter, self-correcting ontology amendments, multi-agent method orchestration, and a locked Driver benchmark. Along the way the construction settles four practical problems as by-products, stated at demonstration strength: compositional and inspectable process context; verification-gated semantic agent orchestration; just-as-needed (demand-driven) ontology extension rather than speculative upfront modelling; and methods as incrementally-expandable domain procedure — each resting on the same verify-before-adopt and zero-live-mutation invariants (§8.1). We relate the work honestly to concurrent frontier results — notably a large-language-model-assisted proof of a ten-year-old jamming conjecture [Parisi & Zamponi 2026] — whose reliance on expert human verification is exactly the reproducible, machine-checked verify-before-adopt gate this programme supplies (§2.1).

Links and sources

Comments