Same logs, different target values
Even identical exploratory logs can conceal different no-exploration steady-state values. Exposure overlap is not a guarantee of context transport.
Online Causal Configuration for Networked Systems
via Doubly Robust Steady-State Learning
Think of a network as a community of interacting units: the nodes and connections stay in place, but signals and local states change whenever a configuration is applied.
Adjust the conceptual interference strength to see how strongly a change in one cluster can echo through its neighbors.
Conceptual network visualization. Experimental measurements and estimator results are presented in the Evidence section.
Drag to explore how the illustrated network responds to stronger local interference.
Exploratory telemetry can misrank configurations when measurement perturbs local actions and shifts the post-burn-in context law. CAUSAL-UCB evaluates no-exploration causal value using joint-exposure doubly robust scores and calibrated predecision certificates.
NEURIPS 2026 · RESEARCH POSTERThe paper, at a glance.Identification · certificates · online decisions
Global decisions change the actions of many interdependent units. Exploratory measurements can improve the identification of a counterfactual value, yet also alter the operating conditions that the final system should optimize.
The target is no-exploration steady-state utility. Action–exposure overlap alone cannot transport post-burn-in contexts; CAUSAL-UCB separates these sources of error and selects configurations using predecision certified upper values.
Represent how a unit's neighborhood influences its observed outcome.
Estimate deployment values from exploratory observations with overlap-aware correction.
Select future configurations using estimates and uncertainty, not raw telemetry alone.
Exploration changes what is observed—and potentially the operating regime whose utility we seek. The paper isolates interference-aware action correction from steady-state context transport, then certifies their remaining uncertainty before selection.
Even identical exploratory logs can conceal different no-exploration steady-state values. Exposure overlap is not a guarantee of context transport.
A predecision value radius separates network-window fluctuation, nuisance products, target-law error, clipping, exposure/response transport, context shifts, configuration mismatch and residual temporal dependence.
Measurement improves identification yet perturbs utility; regret guarantees require valid certificates and selected-information growth.
Under known local interference, the same anytime calibrated radius supports estimation, optimistic selection and reported uncertainty.
A four-stage causal-optimistic loop.
Map local actions and neighbors into a context suitable for auditing interference.
Explore theory & evidence ↗Create support for evaluating alternatives while tracking the deployment perturbation.
Explore theory & evidence ↗Combine outcome regression, logging correction, clipping, and uncertainty assessment.
Explore theory & evidence ↗Use the evaluated values and bounds to choose the next global configuration.
Explore theory & evidence ↗Explore seven-method recorded comparisons, manuscript figures and separate certificate audits. Run an independent Python experiment using the tabs below.
Replay recorded decisions and compare all seven methods, or switch to the Python experiment above.
| Method | Target regret ↓ | Deploy. gap ↓ | Final ATE ↓ | Certified width ↓ | Configuration / opt. rate | Record source |
|---|---|---|---|---|---|---|
| Loading original records… | ||||||
Target regret and deployment gap are cumulative through the selected round. Final ATE and valid certified width are reported at the end. The configuration column shows the modal choice (with its seed count), plus the final optimal-choice rate when applicable. Baseline methods do not have decision-active certified widths; their missing width is not zero. Low-mismatch baseline round traces are exact verified replays, not original archived intermediate logs.
Loading paired seed results…
Loading reference-policy comparison…
Loading deployment trade-offs…
Both mismatch regimes now contain 200-round trajectories for all seven methods, at 24 matched seeds. High mismatch and the two low-mismatch certified traces come from the original archive; the five other low-mismatch traces were deterministically reproduced from the frozen driver, with all 168 policy decision hashes and terminal metrics verified. They are newly replayed data, not independently archived round logs. The browser Python Lab is a separate conditional experiment.
Original frozen SYN 2026-10-20 · known sparse ring and matched-noise assumptions · fixed exploration · 24 paired seeds × 200 rounds. Certified curves come from cert_dr_all_rounds.csv. High-mismatch curves for the other five methods come from the exact archived reference/main_per_round.csv.gz and were cross-checked against per_seed.csv. Low-mismatch reference round histories were newly replayed with the same generator and seeds and validated against the original per_seed.csv terminal values and decision hashes. See replayed_low_mismatch_audit.json. All displayed terminal values still come from per_seed.csv and summary.csv; paired intervals come from paired_regret_95CI.csv. No browser-side retraining occurs in this panel.
Change the parameters and execute the five available policies using Python on your device.
A separate synthetic conditional protocol, designed for interactive exploration. The 7-method study is available in the Recorded evidence tab.
Ready. First run downloads the Python runtime (~10–25 MB). Internet required for the runtime.
| Policy | Configuration | Target regret | DR audit width | Matched nodes |
|---|---|---|---|---|
| No measurements yet | ||||
The browser uses the archived causal_ucb_conditional.py policies in a 24-node known-ring synthetic system with local exposure and randomized measurements. Oracle quantities enter post-run diagnostics; the online policies receive their own logged feedback. Reported DR widths describe this run's evaluation and are distinct from the main study's decision certificates. Conditional and archived seven-method protocols are evaluated separately. Outcomes depend on the chosen parameters and seed. All computations remain on your device.
Browser support: WebAssembly and access to the Pyodide runtime are required for a new local Python run.
Choose a figure to view its original artwork and the accompanying experimental interpretation.

Figure location in source manuscript · p. 8.
Read official paper ↗Both full-path target coverage and the OFU path inequality hold in all 300/300 independently seeded audit runs; intervals are 95% Wilson intervals.
The reported coverage applies to the stated finite-state calibrated protocol and its verified assumptions.
Source: supplied 66-page manuscript · Figure 28 · page 61
Open exact figure ↗Read official paper ↗This repository contains the executable core comparison, archived benchmark data and figure assets. The supplied paper includes additional certificate audits and trace-driven semi-synthetic studies, which are not rerun by this website.
# From repository root
python -m pip install -r requirements.txt
python -m unittest discover -s src/tests -v
python src/fast_compare.py --out results_reproduced --seeds 24 --T 200 --scenarios high_mismatch,low_mismatchChoose the format that fits what you want to do: understand the paper, reproduce a claim, or exchange ideas with the community.
Paper opens the official OpenReview PDF. The archived chart records and source figure crops remain available here for exploration.
Visual montage: illustrative footage of transport and infrastructure from Pexels; interactive network artwork is procedural. Pexels license ↗