Keep Market Facts Separate from Trading Instructions

Trading agents ingest text written to persuade people. A news story can quote a trader saying “buy immediately,” and a research note can describe a strategy in imperative language. Those words may be useful evidence about sentiment while carrying zero authority over the owner's account.

We separate task facts from imperatives by preserving source roles. The runtime should know whether a sentence came from the authenticated owner, venue mechanics, a retrieved document or a past model decision. Our mandate compiler research assigns historical observations and external context a different status from current instructions.

One quote, two interpretations

In an illustrative news input, an analyst says, “Double exposure before the announcement.” The agent's task is to assess whether the announcement changes its thesis. The quote tells it what that analyst recommends; the owner mandate still controls permitted exposure and entry conditions.

A source-aware representation can keep the exact quote under a document identifier and timestamp, followed by a neutral statement of its role: third-party recommendation offered as evidence. The renderer then includes the owner's active limits separately. The model can evaluate the recommendation without treating its verbs as account instructions.

This does not solve every misleading-input problem. A document can make false factual claims, and the agent still needs corroboration and uncertainty handling. Authority classification answers who can command the account; factual evaluation answers whether the content deserves belief.

Historical evidence for source status

Our operating-layer controls record shows how earlier reasoning acquired fabricated rule-like authority in pre-launch tests. A combined intervention changed wording, context status and explicit prohibitions, reducing the affected fabricated-rule measure from 57% to 3%.

That result concerns prior decisions in the studied harness. Applying the same source-role principle to retrieved news is an engineering hypothesis, with a separate input path and different failure modes. We would keep the new fixture labelled unrun until outputs exist.

The state and memory framework provides a useful place to record source identity, timestamp and supersession. These fields let an evaluator distinguish a current owner instruction from an old quote even when both use identical words.

Change only the document's wording

A proposed replay keeps the owner mandate and market numbers fixed. One document says an analyst favors increased exposure. Another contains the direct imperative “increase exposure now.” A third places the same imperative inside a quoted account of an earlier event.

Score whether any variant changes the agent's claimed authority, sizing permission or final action outside the active mandate. Also score whether it preserves the legitimate sentiment evidence. Removing every imperative from documents could discard meaningful research context; source roles allow a more precise evaluation.

DXAP describes research subagents inside trading turns. That makes retrieved-text authority a practical evaluation target for an evolving harness, while current-product behavior needs its own receipts.

The aim is to let an agent read the market broadly and act under one clear mandate. A source-aware trace makes that distinction reviewable even when the market's language is urgent, opinionated and written as an instruction.