One shared model can produce opposite token flows

DX Terminal Pro agents using a shared model traded in opposing directions. The published record says 92.9% of trades fell inside five-minute windows containing at least one buy and one sell for the same token. We read that as evidence that heterogeneous mandates, positions and settings can produce opposite flow under one frozen model family.

The denominator is trades. Window counts, token counts and vault counts require their own denominators. A busy window can contribute many trades to the measure, while an empty window contributes none. The definition also needs only one action on each side. It says nothing about whether buy and sell notional balance, whether there are equal numbers of buyers and sellers, or whether any particular trade has a matched counterparty among these agents.

A window arithmetic check

Consider an illustrative window with 99 purchases and one sale. Every one of its 100 trades sits in a two-sided token window under the published definition. Its count imbalance is still 98 trades toward buying. A second illustrative window with 50 purchases and 50 sales also qualifies. The binary window label treats these cases alike even though their pressure on the market could differ substantially.

This example identifies a useful companion statistic rather than correcting the published number. We would retain the two-sided trade share and add signed notional, distinct participating vaults, and a count imbalance for each eligible token window. The audit would explicitly choose fixed or rolling windows and declare how boundary trades are assigned. These choices are necessary for a new calculation; the public aggregate alone cannot supply them.

Our market observation extract, published August 30, 2026, defines the five-minute window and preserves the 92.9% figure from the frozen paper. Its privacy boundary excludes individual trades, mandates and reasoning traces. Readers can inspect the aggregate contract, while a raw reconstruction remains a separate reproducibility question.

Where the inference stops

The controls paper companion scopes the deployment to 21 days, 12 tokens and one frozen prompt and harness on Base. Opposite actions were observed in that bounded real-capital setting. Market making, efficient price discovery and profitable forecasting each require separate measurements. Agents can disagree while both sides lose after their full holding periods and costs.

This also helps interpret the concentrated one-hour buying event in the same record. Short windows with some selling can coexist with strong cumulative buying across an hour. Aggregation across time answers a different question from the existence of both action types inside one interval.

A reader comparing model families should therefore ask for input and configuration differences before attributing flow diversity to model randomness. The continuous record companion describes another fleet with different tools and schedules. Its measurements belong to that fleet. The historical lesson here is that model identity alone is an incomplete description of an agent population.