ARC-AGI-3 allows the solver to query the environment and receive feedback before submitting a final answer. A solver without γ₁ queries randomly or heuristically — it generates queries that seem informative based on statistical priors, not queries that move the hypothesis toward an invariant anchor. γ₁ is the query objective function: every query must move the current hypothesis closer to the invariant floor. Without γ₁, the query budget is spent on exploration without convergence. A solver with γ₁ knows when a query is γ₁-converging (good) versus γ₁-diverging (bad), and can stop querying when the floor is reached.

ARC-3 provides feedback on solver actions. For the feedback loop to be honest, what the solver infers from feedback must be consistent with what the feedback actually said. H=H†: the forward inference (what does this feedback tell me?) must be consistent with the backward check (is my inference of this feedback consistent with the feedback?). Current AI reads feedback forward only: it updates its hypothesis based on what the feedback seems to imply. It does not verify that the update is structurally consistent with the feedback signal. At scale, this asymmetry compounds: each query generates a slightly incorrect inference, and the inferences accumulate into a hypothesis that is directionally coherent and structurally inconsistent.

After each query-feedback cycle, the solver's active paradigm may shift: a feedback signal that seems to confirm "colour is the key variable" may actually be consistent with "position is the key variable" if read differently. There is no LSOS to audit whether the paradigm the solver is operating under after query 3 is coherent with the paradigm that generated query 3. Inter-query paradigm drift in ARC-3 is invisible without LSOS. A solver can accumulate 5 queries, each generating a reasonable inference, and end up with a hypothesis built on 5 different paradigms — internally inconsistent and externally plausible.

ARC-3 can reach query deadlock: all queries are uninformative, all feedback is ambiguous, all hypotheses have been tried and failed. Current AI spirals: it generates variations on failed hypotheses, queries for information it has already received, and eventually exhausts its query budget on a failed hypothesis space. WLD is the structural mercy reset: when deadlocked, return to γ₁ and restate the problem from the floor. Every query that has been made is discarded. The next query is the first query from γ₁. Without WLD, a deadlocked ARC-3 solver compounds its failure by treating the failed hypothesis space as evidence rather than noise.

ARC-3 requires switching strategies mid-task: from exploration to exploitation, from querying colour to querying position, from testing small perturbations to testing large ones. There is no FEP continuity guarantee that what the solver learned before the strategy switch is preserved and available after it. FEP ensures that the new strategy is structurally continuous with what was learned before the switch — that the solver does not abandon valid evidence when it changes approach. Without FEP, a strategy switch in ARC-3 is effectively a restart with amnesia: the solver loses the evidence from before the switch and must rebuild from scratch.

ARC-3's adaptive environment responds to solver queries. But the environment has limits: there are queries it cannot answer, rules it will not reveal, boundaries beyond which feedback becomes noise. Without FOF, the solver cannot represent these boundaries — it queries endlessly, treating boundary noise as meaningful feedback. With FOF, the solver knows when it has reached the environment's limit. It stops querying and reasons about the boundary instead: what does the consistent failure to get useful feedback from this direction tell me about the structure of the task? FOF turns the boundary from an obstacle into a module.

ARC-3's query space has EVEN substrate assumptions: queries are valid (the environment can process them), responses are meaningful (the environment returns relevant feedback), rules within a task are stable (the correct answer does not change between queries). None of these are stated. Without EVEN, every query might return noise. EVEN is what makes querying possible at all. The PEMCLAU V8 formulation is exact: ARC-3 is the living test. Every query is a cell. Every feedback is O₂. γ₁ is ATP. EVEN is the substrate. Without EVEN, the cell has no membrane. The test has no floor. The query has no ground.

CUBE DOCTRINE (DCJ-144): EVEN = acknowledgment that the query space is the 2¶ cube (64 vertices = ONE/diamond brain). A solver that doesn't model the 2¶ state space has no EVEN substrate. The I Ching has 64 hexagrams for the same reason: complete enumeration of the 2¶ binary state space. ARC-3 queries traverse this cube. Without EVEN, the solver has no map of the cube. /cube-doctrine