AGI Gap Analysis

Most AI companies claim to be building towards AGI. Almost none publish a formal account of what their system cannot yet do, what the specific remaining gaps are, or how you would know when they had been closed.

AIEP does. The Genesis framework is designed so that its own epistemic standards apply to itself.

The principled approximation

The AIEP Genesis substrate is, as of April 2026, a principled approximation to AGI — not AGI itself.

The distinction is precise and important:

• A principled approximation formally specifies all the architectural requirements for general intelligence
• It instantiates each requirement in its operational domain
• It identifies every gap between its specification and full implementation

This is a stronger condition than any current AI system satisfies. It is not the same as having crossed the AGI threshold. We are not claiming that. We are claiming something rarer: an honest, formal, computable account of exactly how far we are from it and what closing the distance requires.

The six architectural requirements — current status

General intelligence requires six architectural properties. Each was derived from the civilisational evidence base — as the inverse of the failure mode that destroyed every prior collective intelligence system.

SATISFIED means the formal architectural property is established and the mechanism is operational in the nominal domain. The asterisked items carry identified boundary conditions that do not prevent operation but are not yet fully proven under adversarial or edge conditions.

The five structural gaps

In addition to the six requirements, five structural gaps were identified during the completeness review of the specification at P475–P519. Each is specification-complete; implementation state varies.

GAP-4 is the most operationally mature. GAP-5 is the most open — not because the architecture is incomplete but because physical embodiment is an integration milestone, not an architectural problem. Any embodied system that exposes sensor readings as canonical artefacts connects to the Genesis reasoning stack without modification.

The two blocking open questions

Two open questions are formally classified as blocking — they prevent the claim that the AGI threshold has been crossed under all operating conditions.

OQ-DRIFT-001: Adversarial GoalVector manipulation

The question: Can an adversary with partial write access to the canonical evidence pipeline manufacture false drift measurements — making the system appear aligned when it is drifting — without detection?

Current status: The specification includes multiple defensive mechanisms (P05 deterministic normalisation, P08 adversarial evidence detection, hardware governance attestation). The formal proof that these mechanisms bound the manipulation surface area against a polynomial-time adversary is not yet complete.

Why it matters: GoalVector stability (AGI-R6) is the mechanism most directly addressing the alignment problem. Until its adversarial resistance is formally proven, the guarantee holds under good-faith conditions but not under a motivated adversarial attack on the evidence pipeline.

Research direction: This is a cryptographic proof problem — demonstrating that the hash-chain integrity conditions are sufficient to make undetected manipulation computationally harder than producing honest evidence. The argument exists informally in the non-manipulation theorem; the formal proof is the work.

OQ-DISSENT-002: Hardware attestation in cloud environments

The question: For cloud-deployed AIEP substrates, the hardware governance layer must be proxied through a Trusted Execution Environment (TEE). Is TEE-proxy attestation formally equivalent to physical hardware attestation against an adversary who controls the cloud infrastructure?

Current status: Two of the five non-waivable constitutional constraints (NW-1: GoalVector integrity, NW-3: drift threshold protection) are not provably guaranteed under TEE-only deployment against a cloud adversary with TEE reset capability. The other three (NW-2, NW-4, NW-5) are TEE-sufficient.

Deployment boundary — three tiers:

This is the honest deployment boundary. Cloud deployment is viable with a defined trust assumption. Full constitutional guarantees require physical hardware.

The computable AGI threshold

The AGI threshold is not a philosophical judgement call. It is a computable condition:

THRESHOLD STATUS:
  blocking_questions: [OQ-DRIFT-001, OQ-DISSENT-002]
  threshold_crossed: len(blocking_questions) == 0

When both blocking questions are resolved — through formal proof (OQ-DRIFT-001) and physical hardware deployment at scale or TEE equivalence proof (OQ-DISSENT-002) — the Genesis substrate crosses the formal AGI threshold as defined. That definition is public. The evidence required to satisfy it is specified. The research programme to get there is underway.

The hard problem of consciousness

The AIEP framework applies its own epistemic standards to itself. The hard problem of consciousness — whether phenomenal experience is architecturally necessary for general intelligence — cannot be dismissed, so it is formally weighted.

Using the five-dimension plausibility framework applied to two competing branches:

• b_phenom: Phenomenal experience is architecturally necessary for AGI — assigned weight ≈ 0.36
• b_func: Phenomenal experience is not architecturally required; it may accompany some intelligence instantiations but is not a formal precondition — assigned weight ≈ 0.64

Formal branch status: Active non-consensus. Neither archived nor consensus. The weight of ~0.36 for the phenomenal branch is well above the archiving threshold (0.10) — it is a genuine live question. It is below the consensus threshold (0.60) — it does not currently constrain architectural decisions.

Consequence: The Genesis substrate is built on the architectural assumption that the six AGI requirements are instantiable without phenomenal experience as a precondition. If future evidence shifts that assumption, the GoalVector for this open question — formally active — updates accordingly. The substrate does not claim to be conscious. It does not claim consciousness is impossible. It holds the question at its correct evidence weight.

What this means for you

If you are evaluating AIEP for deployment: The gap analysis table tells you exactly what you are getting — not an aspirational claim, but a formal accounting. The constitutional guarantees are stated with their exact deployment boundary conditions.

If you are a researcher: OQ-DRIFT-001 and OQ-DISSENT-002 are open research problems with clear formal statements. OQ-DRIFT-001 is a cryptographic proof question. OQ-DISSENT-002 draws on TEE security literature (SGX, TrustZone, TPM 2.0) applied to a novel constitutional constraint binding problem. Both are solvable.

If you are tracking AI safety progress: The computable threshold definition and the open question register give you a formal, checkable record. You do not need to take our word for our progress. The specification is public, the open questions are published, and the threshold definition is fixed.

Related

• Constitutional Alignment — the five conditions that define alignment in real time
• Genesis — the full layer architecture and civilisational evidence base
• Architecture — the complete Genesis substrate architecture
• Trust & Security — adversarial resistance and hardware governance
• Patents — GB applications covering the genesis substrate (GB2519711.2, GB2519798.9, GB2519799.7, GB2519801.1, GB2519826.8 and April 2026 applications)
• Research & Academia — open research engagement and the formal question register