Mechanism
Structural starvation is the architectural technique by which the system prevents language model hallucination by denying the language model access to the informational resources that would be necessary for hallucination to occur, rather than detecting and filtering hallucinated content after it has been produced. The premise is that post-hoc filtering is inherently unreliable: the same statistical patterns that produce hallucinated content also produce plausible-appearing hallucinated content that evades detection, so a system that first generates and then screens output is always one step behind the failure it is trying to catch. Structural starvation instead eliminates the preconditions for hallucination by constraining the informational environment in which the language model operates, so that hallucinated content is not generated in the first place.
The technique follows directly from the role the language model occupies in this architecture. Every language model is a structurally untrusted proposal generator: its output is never authoritative, it is a candidate semantic mutation that must be independently evaluated by the agent's resident validation engine before it can affect any agent field or downstream behavior. Structural starvation is what makes that untrust operationally safe at the input side, by bounding what the model is given so that what it can propose stays grounded in the agent's verified state.
Prompt Bounding
The first constraint is prompt bounding. The language model receives only a bounded, curated prompt context derived from the agent's verified fields, not an open-ended context window populated by retrieval augmentation, user history, web scraping, or other sources of unverified information. The prompt is constructed by the agent's execution substrate from the agent's schema fields and contains only information that has already been validated and incorporated into the agent's governed state. The language model cannot hallucinate about information it has never been given: by restricting the prompt to verified agent state, the system removes the primary substrate on which hallucination operates, which is unverified, ambiguous, or contradictory context.
Absence of External Memory
The second constraint is the absence of external memory. The language model does not have access to any persistent memory, knowledge base, retrieval store, or external data source beyond the bounded prompt context supplied for the current inference call. The model operates on the bounded prompt context and its trained parameters, and nothing else. If the information required to produce a correct proposal is not present in the bounded prompt context, the language model cannot produce the proposal, and the absence of the proposal is the structurally correct outcome, because the agent's state does not contain the information that would justify it. Where a conventional architecture would treat a missing fact as a gap for the model to fill, this architecture treats it as a gap that must not be filled by invention.
Forced Reliance on Agent Fields
The third constraint is forced reliance on agent fields. The language model's proposals must reference and be grounded in the agent's verified field values. The mutation engine that sits between the language model output boundary and the validation engine performs schema mapping, identifying which agent fields each proposal addresses. A proposal that references information, entities, relationships, or facts that are not present in the agent's verified fields is flagged during schema mapping as ungrounded and is rejected prior to validation. This makes groundedness a structural admission requirement rather than a quality the validator merely hopes the output will exhibit.
Intermediate Rejection
The fourth constraint is intermediate rejection. The validation engine evaluates each candidate mutation against agent-resident constraints, and any mutation that fails validation is immediately discarded. The language model does not receive feedback on why its proposal was rejected: it does not receive the validation record, the violated constraint, or any guidance on how to produce a passing proposal. This absence of rejection feedback is deliberate. Providing the model with rejection details would let it learn the validation logic and craft proposals that satisfy the letter of the constraints while violating their intent. The same informational asymmetry that protects against hallucination therefore also operates as an adversarial defense: a model cannot optimize against a constraint it cannot observe, regardless of how sophisticated, fine-tuned, or chain-of-thought-equipped the model is.
Stateless Purging
The fifth constraint is stateless purging. After each inference call, the language model's context is purged. No residual state from a prior inference call persists into the next one. The model does not accumulate context, does not build up a model of the agent's history, and does not develop an internal representation of the validation criteria. Each inference call is an independent event in which the language model receives a bounded prompt, produces a proposal, and is then reset. This statelessness prevents the model from engaging in multi-turn adversarial optimization in which successive proposals incrementally probe the validation boundary, and it preserves the validation asymmetry across calls: even if a model could infer partial information from a single rejection pattern, that inference is destroyed at the boundary of each inference call.
Multi-turn interaction is still supported, but the continuity is carried by the agent rather than by the model. Each exchange produces governed mutations to the agent's fields, and these mutations persist in the agent's verified state. When the next exchange requires the model, the agent constructs a fresh bounded prompt from its current verified state, which includes the accumulated context from prior exchanges. The model receives that prompt, produces its proposal, and is purged. Multi-turn history is thereby subject to the same governance, validation, and lineage constraints as all other agent state, instead of living in an unbounded model context window.
A Composable Safety Primitive
Structural starvation is a composable safety primitive that may be combined with any model-level alignment technique, including reinforcement learning from human feedback, constitutional AI, or preference optimization. It does not replace model alignment; it provides an orthogonal layer that operates regardless of the model's alignment status. A well-aligned model operating under structural starvation produces higher-quality proposals because alignment and the structural constraints reinforce each other. A poorly aligned or adversarially fine-tuned model operating under structural starvation is prevented from causing harm because the structural constraints deny it the informational and operational prerequisites for harmful output, even if its parameters encode adversarial intent. The result is a defense-in-depth architecture in which neither layer depends on the other for safety, but both contribute to quality. The present disclosure does not depend on the language model being well-aligned; it produces safe behavior through architectural containment.
Disclosure Scope
Structural starvation, comprising the five complementary architectural constraints of prompt bounding, absence of external memory, forced reliance on agent fields, intermediate rejection, and stateless purging, together with the treatment of the language model as a structurally untrusted proposal generator whose output reaches no agent field without passing through the resident validation engine, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart). This article describes that disclosed mechanism. The scope extends to the validation feedback asymmetry by which the model receives no rejection rationale and therefore cannot optimize against constraints it cannot observe, to the carriage of multi-turn continuity in the agent's governed state rather than the model's context window, and to the composition of structural starvation with model-level alignment techniques as orthogonal defense-in-depth, provided that hallucinated content is prevented at generation by constraining the informational environment rather than filtered after generation.
