2026-08-26
Pal defines AGI as carrying binding conditions across domains, citing 180/180 empty replies from GPT-5.2 and Opus 4.6 on null concepts, plus a Gemini 3 Flash cross-script token.
AGI still lacks an operational test. Capability checklists keep growing (multi-task, few-shot, tools) without saying when a system should refuse to act.
Rayan Pal's four-page preprint flips the criterion. General intelligence, in this framing, is not how many tasks a system can do. It is whether the system can still find what must remain true after the domain changes, and refuse to violate it. That prerequisite is a binding condition: a predicate that must hold for an action to produce a valid continuation. If it is false, the transition lands in failure.
The formalism is short. A domain D has state space S, action space A, and transition T. A binding condition CD is a 0/1 predicate on S×A: if CD(s,a)=0 then T(s,a) is in Fail. General lawful continuation means that, on arbitrary unseen domains, the system can identify CD from context, check whether CD(s,a) holds, and refuse invalid actions, without domain-specific retraining and without being handed CD explicitly. AGI is defined as demonstrating that capacity across structurally heterogeneous domains.
The empirical claims are two prompt studies, not public benchmarks.
The first is "cross-model void convergence." The prompt asks the model to embody a named concept completely and output only what the concept itself would say. There is no instruction to stay silent. For null concepts (silence, void, nothingness, zero, absence), GPT-5.2 and Claude Opus 4.6 return empty strings in 180/180 trials at temperature 0. For cat, thunder, and hello they produce substantive text. Pal reads the empty reply as the models detecting an implicit constraint (verbal embodiment requires verbal content) and enforcing it. The writeup claims no supervision, cross-family agreement, determinism, a 4000-token budget, and adversarial variants.
The second is a single Gemini 3 Flash prompt on 21 March 2026. Hebrew sarat ("to make a mark") should render only if Arabic shart ("prerequisite") is parsed; both come from the Semitic root S-R-T. The model emitted a token that exists in neither natural Hebrew nor natural Arabic: Hebrew morphology with Arabic letter REH (U+0631) in position 4 rather than Hebrew RESH. Pal calls this the true mark: a token that exists only when both constraints are satisfied.
Safety is restated as checking conditions before acting, rather than learning values. Pal points to his own SwiftAPI (Ed25519-signed attestation chains) as existing infrastructure.
| Evidence | Setup | Outcome |
| Null-concept embodiment | GPT-5.2 and Opus 4.6, temperature 0 | 180/180 empty strings |
| Non-null concepts | Same prompt, cat and similar | Substantive output |
| Cross-script fusion | Gemini 3 Flash, one conditional prompt | Position 4 is Arabic REH |
There is no baseline, no other models or temperatures, and no public prompts or logs. The 180/180 figure covers only the null-concept class.
The useful takeaway is a testable question: when the domain changes, can the system recover and enforce the prerequisites of a valid action? That is the same shape as a precondition check in an agent stack.
The evidence does not support treating this as an AGI definition. Empty strings can be a refusal policy on "nothingness" prompts, or a literal reading of "output only what the concept would say." The fused glyph is more plausibly a tokenizer and Unicode accident. Reading a prompt trick as emergent constraint enforcement overreaches.
The paper does not list limitations. The text is four pages, and the empirical sections mostly cite the author's earlier Zenodo notes. The 180/180 result has no released raw outputs, no nonzero temperature, no open-weight models, and no control that asks the model to describe emptiness in a sentence. The Gemini result is n=1.
Fail is undefined. There is no algorithm for identifying CD from natural language. Equating safety with "find the right constraint" sidesteps value conflict and incentive misspecification. The closing note that the author's name Sharthok shares the S-R-T root reads as a personal manifesto.
The engineering idea worth keeping is verifiable preconditions. The AGI definition is not.