Before October 7, U.S. ties protected Chinese AI researchers. Now they don’t.

A regime-break, empirically visible

The October 2022 U.S. chip-export controls reshaped the Chinese state's calculus on its own AI researchers, and the reshaping is large enough to register in a model with five hand-engineered features and seventeen historical events. We add a paired feature to the model — the same U.S. co-authorship density variable, applied separately to the pre-October-2022 and post-October-2022 windows — and run the same held-out test. The learned weights have opposite signs: −0.44 before the controls (protective) and +0.36 after(risk-loading), with an effective post-controls slope of approximately zero. Held-out calibration error improved 7% (Brier score 0.088 to 0.082); rank-discrimination rose from 0.91 to 0.92. Before the controls, dense U.S. co-authorship lowered a Chinese AI researcher's probability of facing travel restrictions. After the controls, it stopped doing so.

Why the calculus had to change

On October 7, 2022, the U.S. Bureau of Industry and Security imposed the broadest set of chip-export controls in the history of U.S. semiconductor policy. The rule blocked Chinese access to advanced AI training chips, prohibited U.S. persons from supporting certain Chinese semiconductor manufacturing, and extended the Foreign Direct Product Rule deep into the global chip supply chain. Within weeks, U.S.-passport engineers at YMTC, CXMT, and SMIC began compliance departures — the Financial Times counted roughly eighty at YMTC alone; Bloomberg counted around forty at CXMT. The visible cost of the controls landed first on Americans inside Chinese fabs.

The deeper question, the one this paper exists to answer, was whether the same controls would reshape Beijing's calculus on its ownAI researchers. Specifically: how should a Chinese AI researcher with dense U.S. co-authorship ties expect to be treated by the Chinese state in the year after October 7? Before the controls, the answer was relatively clear — these were the researchers Beijing wanted to retain, recruit back, and protect from the perception of being “tainted” by U.S. ties. Many had been beneficiaries of the Thousand Talents Plan or its successors. After the controls, the working hypothesis among China-watchers was that the same dense U.S. ties would become a vulnerability rather than an asset — a signal that the researcher might be subject to passport custody, pre-travel review, or formal restriction.

That hypothesis, if true, implies a sign-flip in our predictor. The same high value on U.S. co-authorship density that pre-2022 lowered restriction probability would post-2022 raise it. A model asked to learn this from data, without being told about the regime change, would have to either pick one sign and live with the other regime as noise, or learn something blunt and uninformative. Neither is satisfying. This update is what happens when you put the regime change into the model's feature specification and let it learn one coefficient for each side.

One feature, two regimes

The feature in question is the share of a researcher's co-authors in the previous five years who hold a primary U.S. affiliation. In the original model, this feature entered the logistic regression with a single learned coefficient. The model was given data spanning 2020 through 2026 and asked to fit one slope across the whole window.

The original model learned a coefficient of roughly −0.15 on standardized density — mildly protective, in the aggregate. That number is a compromise. It sits between a stronger protective signal in the pre-2022 portion of the data and a weaker (or possibly opposite) signal in the post-2022 portion, averaged together by the maximum-likelihood objective. If the regimes truly differ, this averaged coefficient under-predicts restrictions on U.S.-tied researchers post-2022 and over-predicts them pre-2022. Both errors are real costs to a calibrated forecast.

The dossier flagged the choice openly: encode the October 7 boundary explicitly, or attempt to learn a single coefficient across the full 2020–2026 window. This release encodes it.

The paired feature, balanced classes

The fix is a paired-feature specification. We retained the original U.S. density feature as the pre-controls slope, and added a companion feature that is identical to the first but takes the value zero for every observation dated before October 7, 2022. With both features in the logistic regression, the model learns one coefficient that applies across the whole window (the pre-controls slope) and a second coefficient that adds to it for post-controls observations. The effective post-controls slope is the sum of the two coefficients.

The other half of the fix is on the dataset side. Each row in the training data now carries an observation date. For positives (researchers who experienced a restriction event), the observation date is the event date. For negatives (researchers who did not, in the relevant window), we sampled the observation date from the empirical distribution of event dates — matching the pre/post temporal mix of positives. Without this rebalancing, the model would see almost no post-controls negatives, and the post-controls coefficient would be learned from a heavily class-imbalanced subset.

The held-out evaluation uses the same 75/25 stratified split as the original release, with the same random seed, so the only difference between the original calibration numbers and the updated ones is the feature specification and the dataset balancing. Same data, same split, same scoring function. Different model.

Opposite signs, near-zero effective slope after the controls

The learned coefficients have opposite signs. On standardized U.S. co-authorship density, the pre-controls coefficient is −0.44 (protective). The post-controls add-on coefficient is +0.36 (risk-loading). The effective slope in the post-controls regime is the sum: −0.08. The protective effect of U.S. co-authorship density in the pre-controls regime collapses to near zero post-controls. Note that we have not measured the post-controls slope crossing above zero; we have measured it collapse to zero, which is itself the regime change the dossier hypothesized.

The headline calibration metrics moved in the right direction. The held-out Brier score — the squared-error measure of how well-calibrated the probabilistic forecasts are — fell from 0.088 to 0.082, a roughly 7% reduction in error. The held-out AUC — the rank-discrimination measure — rose from 0.91 to 0.92. Both deltas exceed the pre-registered acceptance threshold for this kind of feature change. The hypothesis was correct, and the effect size matched the prediction.

The regime-break is real, and it’s in the data

The substantive finding is that U.S. policy reshaped Beijing's calculus on its own researchers, and that the reshaping is large enough to be visible in a model with five hand-engineered features and seventeen historical events. We did not have to wait for years of post-flip data, run a structural break test, or invoke an unobserved-states model. We had to write down the date of the regime change and ask the logistic to see it. That is a low bar to clear — and the fact that it clears it cleanly is itself the substantive claim.

For policy: the October 2022 controls had a second-order effect that compounds the first-order trade restriction. Beijing did not just lose access to high-end chips; it lost the asymmetric incentive structure under which its U.S.-connected AI researchers were a strategic asset to retain. After the controls, those same researchers became a population to monitor. The expansion of passport-custody and pre-travel-approval practices — from state labs (the 2023–2024 record) into the private frontier firms (the May 2026 Bloomberg dispatch on Alibaba, Moonshot, Zhipu, and DeepSeek) — is the operational evidence of that shift.

For modeling: predictor sign-flips that follow from documented regime breaks should be encoded in the feature specification rather than learned from data. Asking a logistic regression to discover the October 7 boundary on its own, with the data we have, would require either far more events or far more features. Telling the model where the boundary is, and letting it learn the slope on each side, is the correct division of labor between domain knowledge and statistical inference for a forecasting system with seventeen events on a five-year window.

What this finding does not claim

Three limitations belong on this update rather than buried in an appendix. First, the cohort is small. The published scoreboard rests on five pseudonymized profiles, and the updated Brier of 0.082 is a meaningful improvement over the prior baseline of 0.088, but the absolute scale of the held-out test set is small enough that a single resolved event will move the headline number noticeably. The right way to read it is as a directional improvement, not as a population estimate of calibration. Second, the regime break is encoded as a hard step on October 7, 2022. The real political-economy transition probably unfolded over six to eighteen months as compliance practices and informal pressures propagated; a smoother specification would be more faithful to the underlying process but harder to identify on seventeen events. Third, the near-zero post-controls slope is an average over the post-controls window. There may be further sub-regimes — the December 2024 BIS expansion, or any future U.S. action — whose own sign-flips we have not yet encoded.