AGI — Proper Definition, Approach, Our Work

1. The Problem With Existing Definitions

“AGI” is an overloaded term with dozens of competing definitions. The research community cycles through Turing Test variants, benchmark scores (ARC-AGI, MMLU, HumanEval), and human-level thresholds. None is operationally useful: each is either unfalsifiable, gameable, or defined against a moving target. The OVRIN ecosystem cannot build toward a slogan; it needs a precise, testable, architecturally actionable target.

The rejection criteria are sharp. Turing-Test framings collapse into subjective judgement and are susceptible to stylistic gaming. Benchmark-only definitions substitute a narrow proxy for the general property they purport to measure. Human-level-as-threshold conflates a cognitive property with a contingent species-specific baseline that varies by domain and shifts with education, tooling, and time. None of these guides hardware procurement, research direction, or inference architecture.

The definition we adopt answers a different question: what target, if achieved, would make the ecosystem's architectural decisions structurally inevitable? The answer needs to generate hardware requirements, research priorities, and evaluation protocols as derivatives. It also needs to survive the harder scrutiny of its own logical closure: when the target is met, the definition should not need to be reframed to recognize success.

2. Definition

AGI is peak intelligence on any question, at any time, in an instant.

The definition decomposes cleanly into two conditions. Either alone is insufficient. Both together are necessary and, we argue, approximately sufficient for the practical realisation sovereign deployment requires.

Condition 1 — Peak Intelligence

The system produces K-minimal completions across arbitrary domains. For a well-formed problem P, the delivered response R minimises the Kolmogorov complexity K(R | P) — no alternative completion has equal or lower K-complexity given P. This is structural inevitability (SI) applied universally: at peak intelligence, the solution is not the best available option; it is the only option the problem's constraints permit.

A benchmark proxy is necessary because K-complexity is not directly observable in finite-precision hardware. HorizonMath — a curated corpus of hard-discovery, cheap-verification mathematical problems — is the proxy we use. A system's K-minimality can be checked problem-by-problem against known SI solutions, and the aggregate trend against the frontier estimates its progress on Condition 1. The community's ARC-AGI and FrontierMath benchmarks serve as adjacent proxies; none are sufficient in isolation.

Condition 2 — Instant Delivery

The system delivers at a floor of 3,000 tokens per second on sovereign hardware. This is not a convenience target. It is a cognitive architecture requirement. Below roughly 3,000 t/s, extended reasoning chains — the kind required for multi-step derivation, long-horizon planning, and real-time collaborative cognition — are effectively blocked by latency. The architecture that drops below this threshold cannot support the cognitive workloads Condition 1 is meant to unlock, regardless of how K-minimal its completions are.

Sovereign hardware is a deliberate clause. Cloud-hosted inference is disqualified by construction: it couples the cognitive substrate to an external authority whose availability, pricing, and content policies the operator does not control. A system whose peak intelligence vanishes when an API key is revoked is not peak intelligence. The condition terminates on hardware the operator owns, runs, and can replace.

3. Alternatives Considered and Rejected

Turing-style definitions. Rejected: subjective, gaming-susceptible, not architecturally actionable. The Loebner Prize history demonstrates how the framing collapses into style rather than substance.

Benchmark-only definitions (ARC-AGI, MMLU, GPQA). Rejected individually: each benchmark is narrow. The two-condition framework subsumes them as proxy measurements rather than replacing the definition with them.

Human-level-as-threshold. Rejected: human-level is a moving target and varies by domain. K-minimality is domain-independent and formally defined. A system exceeding human-level on chess did not “become AGI” and the boundary between specialized and general capability is better drawn by structural conditions than by comparative performance.

Scaling laws as a definition. Rejected: a scaling law is a prediction about trajectory, not a definition of arrival. A system whose loss curve projects favourably is not thereby a general intelligence.

4. Consequences

The two-condition framework generates the ecosystem's research and procurement posture as derivatives.

• Research direction targets K-minimal completion across domains. This directly motivates the Charter Protocol research architecture for approaching Condition 1 within bounded domains by pre-amortizing structural inevitability.
• Hardware procurement targets sustained 3,000+ t/s sovereign inference. This frames the Enzo inference infrastructure roadmap and the downstream hardware choices (memory bandwidth, PCIe topology, cooling envelope) that fall out of the throughput floor.
• Evaluation uses HorizonMath (K-minimality proxy) and throughput telemetry (Condition 2) as complementary signals. A system that passes one condition and fails the other is partially aligned — the partiality is the information.
• Zero Ultra — the Exocortex horizon — represents the extension of both conditions to sovereign cognitive augmentation. The same K-minimality and instant-delivery constraints that define AGI define the substrate the operator interfaces with as a cognitive prosthesis.

5. Alignment to Manifesto

The definition is not an engineering convenience. It is load-bearing against the ecosystem's constitutional principles.

Principle I — The Seed Is the System. The AGI definition specifies what the seed is building toward. Every architectural decision in the ecosystem must not preclude the target. The Zero Ultra test (DL-020) is the falsifiable application: if a proposed architecture closes the door to either K-minimal completion or sovereign instant delivery, it is rejected regardless of near-term convenience.

Principle V — Intelligence Compounds. Condition 1 (peak intelligence) is the output the system compounds. Condition 2 (instant delivery) is the speed at which compounding occurs. A system that meets Condition 1 but fails Condition 2 compounds slowly enough that the operator's own rate of thought exceeds the system's — the substrate becomes a bottleneck rather than an augmentation.

6. Where Citium's Work Sits

Against the two conditions, three active research programmes in the OVRIN ecosystem map cleanly to components of the target.

Charter Protocol is the proposed mechanism for approaching Condition 1 within bounded domains. Rather than chasing K-minimality globally, Charter Protocol decomposes intelligence into domain-scoped pre-amortizations, each pursued by writing the domain specification until queries within it can resolve toward structurally inevitable territory. The AGI companion preprint gives the public treatment of the computable charter layer.

Enzo is the inference infrastructure targeting Condition 2. The work is concentrated on throughput under sovereign constraint: local hardware, local models, and the scheduling and memory-bandwidth choices that permit sustained 3,000+ t/s rather than peak-but-bursty throughput.

Lorengine is the retrieval substrate underneath both. A Charter can only pre-amortize inevitability over a corpus the inference stack can surface; Lorengine provides the sovereign semantic-context layer that makes the Charter's canonical authority queryable at inference time without leaving the machine.

7. Closing

The two-condition framework is intentionally narrow. It discards philosophical flourishes that do not generate architectural constraints and keeps only the properties that drive the ecosystem's design. We acknowledge the framework leaves several legitimate concerns unresolved: safety under peak intelligence, the alignment of the canonical authority a Charter encodes, and the governance of the sovereign inference stack. These live in companion work (the manifesto, the Charter Protocol spec, Zero Ultra) and are not redundantly restated here.

The value of the framework is that it closes. Given Conditions 1 and 2, the ecosystem knows what to build, what to measure, and what to refuse. That is the standard the definition has to meet to be worth adopting.