Our experimental forecasting agent has always observed the world as a field of changes — velocities and accelerations of prices, rates, money supply, sentiment, each scored against its own history. But when it came time to predict and grade, it quietly fell back to levels: "which way does the price go?" This week's round of upgrades closes that gap end to end.

What got improved:

• Stale data can no longer shout. Attention on the change watchboard now decays with the age of a series' last print, so a three-month-old rate release can't sit at the top of the board looking like breaking news. And exchange artifacts — contract rolls, adjusted-close restatements — no longer masquerade as data revisions; only genuine statistical-agency revisions count as events.

• Seasonal series are measured honestly. Freight volumes (a favorite growth thermometer) are strongly seasonal, so month-over-month change mostly measured the calendar. They're now tracked year-over-year — the annual cycle disappears, and what's left is actual signal.

• What the scanner finds can now become a forecast. Until now, a 9× persistent drift in gold could sit on the watchboard and never trigger a single thesis — the deliberation layer only thought about stocks. Macro findings now condense into change claims: machine-checkable statements like "this drift ends within ten sessions" or "this divergence closes", graded against the recorded trajectory itself, never against a price direction. They get their own scorecard, kept strictly separate from the directional track record.

• Surprise is now measured as change, in every component. What triggers deliberation used to mix differences with levels (how bold a call is, how bad a name's recent record is). Now it's the change in conviction and the deterioration of calibration that raise the alarm — a name that has always been mediocre is not news.

• The world-model filter watches the slow economy too. Its regime beliefs were driven entirely by daily market composites. It can now also step on the slow macro pulse — money supply, labor, actual inflation prints, freight — but only on days something actually printed, plus a new dispersion channel that notices when the equity complex internally disagrees. Its acceleration channel also received the unit repair that had kept it honestly self-disabled.

• Divergence watching got structure. One-day sign flips no longer count (three sessions minimum), mixed-frequency pairs (weekly claims vs. daily equities) can finally be compared, the pair catalog roughly tripled, and a new detector flags when a pair's correlation regime breaks — the interesting difference being that the relationship changed, not that it exists.

• Even the fast reflex can learn to watch changes. The instant-scoring layer has always read levels of sentiment and momentum. It now has change sensors — the velocity of tone, the acceleration of momentum — wired in at zero weight, with candidate weights staged through the same out-of-sample adoption gate as every other behavioral change.

The through-line: nothing was switched on by hand. Every behavioral change sits behind the system's own measured adoption discipline — replayed against the historical record nightly, adopted only when it demonstrably predicts better, and reverted the same way. The representation was already differences; now the predictions, the grades, and the attention are too.