Intervention-Supported Unit Discovery

Temporarily disabling a candidate channel and comparing what follows against a measured baseline of unperturbed repeats — not a fixed threshold — separates real coordination from ordinary workflow correlation in cases where passive clustering cannot.

What decision changes?

When a passive correlation-based boundary estimate looks too coarse (an entire pipeline reads as one unit), do not tighten the threshold. Intervene on one candidate channel at a time and compare the outcome against a measured baseline, not an assumed one.

A passive audit that only looks for correlation has a specific failure mode: an ordinary multi-step workflow already looks correlated, so clustering on correlation alone can merge an entire pipeline into one false “unit” even when only two of its participants actually coordinate.

The fix tested here is interventional rather than passive. Disable one candidate channel — a communication link, a shared resource, a specific interaction — and see whether the participants’ subsequent behavior changes more than it would have changed anyway. The “would have changed anyway” part is the load-bearing addition: the comparison is scored against a baseline measured from several unperturbed repeats of the identical setup, not against a fixed, hand-picked threshold. Real systems, especially ones with any stochastic component, vary run to run even with no intervention at all; treating any nonzero post-intervention change as evidence collapses the moment that ordinary variation is nonzero.

In restricted testbeds with a known ground-truth partition, this channel-ablation-plus-measured-baseline approach recovers the correct grouping on cases where passive clustering does not, including a negative control where no real coordination exists. The same approach, rebased onto outcome-level observables (what happened, not exactly when), also recovered a coordinating pair in one run driven by a real language-model agent that a fixed-threshold version of the same method missed.

That is where the confident part of the claim ends. A harder stress test — a three-participant scenario, and a five-seed replication of the single successful run above — showed the method swinging the other way: missing a real pair it should have caught, and dropping to a success rate no better than the passive baseline it was meant to improve on. The general shape (compare an intervention’s effect to a measured null, not a fixed one) is the right instrument design; getting it calibrated well enough to trust on noisy, non-repeatable systems is open work, and the negative results are kept alongside the positive ones rather than discarded.

What would count as evidence?

Restricted testbeds with a known ground-truth partition and a scripted or model-driven multi-actor pipeline, run with and without a given channel disabled, replicated enough times to measure ordinary run-to-run variation before calling any post-intervention change a real effect.