Abstract
Large language models are often described as hallucinating when they produce incorrect
or unsupported outputs. That label is useful descriptively, but it conflates several distinct
mechanisms, including local generation error, unsupported fabrication, and
interaction-driven reinforcement of an initially weak claim. This paper isolates the third
case.
We argue that a subset of hallucination-like behavior is better understood as Recursive
Narrative Amplification (RNA): a feedback dynamic in which an already introduced
claim gains authority through recursive reuse before sufficient external constraint has
refreshed or corrected the trajectory.
The paper also introduces aggregation-induced error as a distinct structural failure mode
arising from the collapse of competing constraints during generation.
To ground this argument precisely, the paper defines a foundational vocabulary of eight
core terms, including constraint, revisability, drift, and closed trajectory, that distinguish
trajectory-level failure from output-level error. Closure pressure at turn t is modeled with
the update rule:
∆C_t = r_t - k_t where r_t is recursive reinforcement and k_t is effective constraint refresh. A positive
cumulative imbalance pushes the interaction trajectory toward Premature Narrative
Closure (PNC).
Controlled deterministic experiments using fixed-schema recursive loops, RAW vs
SANITIZED reinjection, gain sweeps, and refresh interventions support this
interpretation. The strongest early result is a 7-agent confidence-gain run in which RAW
closes under constant claim and evidence while SANITIZED remains open, even though
both branches remain parse-valid and state-valid throughout.
The paper further extends this argument through GuardianAI LAB6. On a controlled
arithmetic benchmark, GuardianAI preserves correct outputs, corrects recoverable
numeric slips, and withholds collapse trajectories. On a harder olympiad-style
benchmark, it releases no wrong answers and withholds all unresolved cases.
These results support a broader claim: some failures are not one-step bad outputs but
recursive trajectory failures, and correct behavior under uncertainty is often non-release
rather than forced answer production.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
pham, thom, "Artificial Genuine Intelligence: Why Constraint Matters More Than Values in Recursive AI Systems", Technical Disclosure Commons, (March 30, 2026)
https://www.tdcommons.org/dpubs_series/9666