Large language models (LLMs) turn writing into a live exchange between humans and software. We characterize this new medium as a discursive network that treats people and LLMs as equal nodes and tracks how their statements circulate. We define the generation of erroneous information as invalidation (any factual, logical, or structural breach) and show it follows four hazards: drift from truth, self-repair, fresh fabrication, and external detection. We develop a general mathematical model of discursive networks that shows that a network governed only by drift and self-repair stabilizes at a modest error rate. Giving each false claim even a small chance of peer review shifts the system to a truth-dominant state. We operationalize peer review with the open-source Flaws-of-Others (FOO) algorithm: a configurable loop in which any set of agents critique one another while a harmonizer merges their verdicts. We identify an ethical transgression, epithesis, that occurs when humans fail to engage in the discursive network. The takeaway is practical and cultural: reliability in this new medium comes not from perfecting single models but from connecting imperfect ones into networks that enforce mutual accountability.
Figures & tables
Figure 1: Venn-style illustration of members of a broader set of invalid outputs produced by large language models.
Element in N
Instantiation in the single-network model
S={r,f}
Binary, mutually exclusive statement states: r is valid/true and f is invalid/false.
Pij
Base persuasion or drift: p contributes to r→f and q contributes to f→r . In the generic form these are summarized by a=p and b=q .
Iij
In the single-network base case, invalidation is represented by the transition r→f ; correction is represented by f→r .
Cij
Message containing the current belief state Bi∈{{r},{f}} .
A
Fixed actor set; proportions refer to n=∣A∣ .
Bi
Current belief of actor ai : either {r} or {f} .
Table 1: Mapping of discursive network elements to the single-network two-state model.
Element in N
Instantiation in the emergent-invalidation model
S={r,f}
Binary statement states: r is valid/true and f is invalid/false.
Pij
Intrinsic drift p contributes to r→f ; correction q contributes to f→r .
Iij
Internal invalidation and self-correction are summarized by the effective transition pressures asingle=p+λ and bsingle=q .
Cij
Message from ai to aj carrying Bi .
A
Fixed actor set; proportions refer to n=∣A∣ .
Bi
Current belief of actor ai : {r} or {f} .
Table 2: Mapping of discursive-network elements to the single-network emergent-invalidation model.
Element in N1∪N2
Instantiation in the cross-network model
S={r,f}
Binary statement states shared by both networks: r is valid/true and f is invalid/false.
Pij
Within Nk , pk contributes to r→f and qk contributes to f→r . Across networks, detection by Nj contributes djk to false-to-true correction.
Iij
Cross-network invalidation: a receiver in Nj can identify a false state in Nk and trigger correction with parameter djk .
Cij
Message carrying the statement or belief state being evaluated.
Ak
Disjoint actor sets A1,A2 ; proportions refer to nk=∣Ak∣ .
Bij
Belief of actor aij∈Ak : {r} or {f} .
Table 3: Mapping of discursive-network elements to the cross-network invalidation-detection model.
Figure 2: FOO consensus loop. An arbitrary set of agents a1,…,am receives the user task, produces candidate answers, cross-critiques peers, and feeds all critiques to one or more harmonizers h . Harmonizers synthesise the feedback; agents revise and the loop continues until a convergence criterion is met. The design allows any number or type of agents and any custom instruction set, making the architecture task- and model-agnostic.
Symbol
Value
Model role
Interpretation in this section
p
0.02
intrinsic true-to-false drift
Small background slippage from valid to invalid state.
q
0.05
intrinsic false-to-true repair
Modest self-repair, reflecting that intrinsic self-evaluation is useful but unreliable without external evidence.
λ
0.055
additional false-production drift
Chosen with p,q so that the isolated network has false equilibrium fsingle∗=0.60 , an invalidation-dominant baseline.
d
0.19
per-peer external detection/correction
Chosen to illustrate a positive verification channel: one external peer lowers the false equilibrium to approximately 0.24 , and a nine-agent continuous-rate system falls below a 5% target.
Table 4: Illustrative parameter values for the corrected dynamics. The values are chosen to display the qualitative regimes of the theory; they are not fitted estimates. The same numerical value d=0.19 is used as a one-round Bernoulli detection probability in the Monte-Carlo plots and as a continuous-time rate in the analytic rate curve; those are separate interpretations used for comparison.
Figure 3: (a) Single-network two-state dynamics with a=p+λ=0.075 and b=q=0.05 . The trajectories converge to fsingle∗=0.60 and rsingle∗=0.40 , as predicted by Lemma 2.4 . (b) Two identical networks with one external detection channel. The simulation uses the exact Bernoulli correction probability b=1−(1−q)(1−d)=0.2305 , so both networks converge to the same false equilibrium fcross,disc∗≈0.2455 . Any persistent asymmetry between two identical networks would indicate non-equilibration, a code–equation mismatch, or a different implemented update rule.
Figure 4: Long-run false share as a function of the number of agents. The exact Bernoulli curve uses bndisc=1−(1−q)(1−d)n−1 and approaches the floor 0.075/(0.075+1)≈0.0698 , which remains above the 5% target. The continuous-rate curve uses bnrate=q+(n−1)d and crosses the 5% target at n=9 . The figure therefore displays two different regimes, not two simulations of the same stochastic matrix.
Large language model reliability studies whether outputs remain correct, calibrated, robust, and controllable across prompts, contexts, and deployment conditions.