Why repeated reasoning can turn an invented rule into policy

Persistent reasoning needs a distinction between describing a past decision and authorizing a future one. Without that distinction, a plausible explanation can return as context, acquire the appearance of precedent and influence the next action. Our historical rule-fabrication result gives this problem a concrete form: agents cited rules that their owners had never supplied.

In the DXRG controls paper companion, fabricated rules appeared in 57% of sell decisions in the affected pre-launch test populations before a combined harness revision, and 3% afterward. The revision removed law-like wording, described previous decisions as context rather than precedent and prohibited invented thresholds and named rules. Section 4 and Table 4 of the primary paper carry the record.

The 54-percentage-point decline belongs to that package of changes. Exact per-arm counts are absent from Table 4 and the adjacent account; the 4,900 classified traces reported across the aggregate diagnosis provide no substitute denominator. The measure concerns fabricated rules in sell traces, rather than investment returns or a measured production-wide elimination rate.

A reasoning claim can change type while keeping its words

We interpret the failure as authority promotion: text generated to explain one decision is later treated as a standing instruction. This is our mechanism-level synthesis, rather than a claim that the historical comparison isolated each step experimentally.

Suppose an illustrative agent says, “I waited because the spread was wide.” In the next context, that sentence remains a report about a previous observation. If a summary rewrites it as “Our rule is to wait whenever spreads widen,” the sentence now proposes a general policy. A further turn might cite that policy as the reason to reject an otherwise eligible action. The owner authorized neither transformation.

The danger follows from closing the loop. The model's output becomes part of its future input, so a generated explanation can gain repeated exposure without receiving additional evidence. A coherent continuation may preserve the apparent rule because it fits the history. Consistency with earlier text and fidelity to the owner's mandate can then diverge.

Longer reasoning would have competing effects under this theory. It could help the model notice the missing authority. It could also produce more plausible rule-like statements for later summaries to retain. The direction depends on how the harness labels, selects and reuses the material. A reasoning budget alone says little about that relationship.

Two assumptions that need different repairs

One account of fabrication treats it as a factual recall error: the model misunderstood an instruction, so clearer wording should suffice. A second treats it as a transition error: even a correct explanation can become an unauthorized instruction when the runtime recirculates it. The historical combined intervention is compatible with both. Its success supports repairing the package; it leaves their separate contributions unresolved.

The second account has a broader implication. A retained statement should carry its role: owner instruction, observed fact, model hypothesis or prior rationale. Only an authorized update should change the owner-instruction role. A summary that shortens text must preserve that role along with the content. This is a design proposition, not a measured new feature claim.

Our research priority is to distinguish whether a false rule originates in the first response or appears during later reuse. Those paths imply different repairs and different evaluation horizons. The useful question for a persistent agent is whether a statement acquired authority as it travelled through the loop. The 57% to 3% record establishes a bounded intervention result; the authority-promotion theory explains why builders should investigate the whole journey.