Loading
Loading
NeurIPS 2026 · Workshop paper
An agent pull request's file list is not the diff a three-way merge uses. On a stratified set of 715 replay-labeled pairs, 47.3% of textual conflicts — 79 of 167 — share no authored file. The checker is still git merge-tree. A score only decides what to run first.
Asymmetric Repository Lineage Modeling and Verifier-Guided Coordination in Concurrent AI Coding Agents. Arjun Subramanian (MIT CSAIL) · George Xu (Harvard Medical School) · Nithilan Karthik (Managent). NeurIPS 2026, workshop on AI for Verifiable Coding: Human-Aligned Collaborative Agents for Autoformalization, Proofs, and Heuristics. Builds on arXiv:2607.04697.
The proxy mismatch
47.3%
of textual conflicts have disjoint authored file sets.
Any detector that waits for shared files therefore tops out at 52.7% recall on this set (88 of 167 conflicts). A file-list model that sees repository id and authored paths, and no branch history, catches 2 of 79 of the zero-overlap conflicts. For 48 of the 79, rebuilding both patches on the common merge base makes the replay clean; the other 31 reconstructions fail and stay inconclusive. That checks the expected Git explanation. It is not a new merge mechanism.
Cross-agent pairs are enriched in this set, so 47.3% is not a population rate. The paper also does not claim each case is uniquely pairwise: the frozen data lack the contemporaneous default-branch tip, so some of the gap may be branch staleness.
What to replay first
Compare paths changed since the two-head merge base, not each pull request against its own base. On a repository-held-out test of 497 pairs (117 conflicts), a one-bit lineage-union rule already reaches AUROC 0.877. An untuned random forest on the same twelve decision-time features reaches 0.902.
Authored-file overlap
AUROC 0.694
95% CI 0.646–0.743 · PR-AUC 0.424
Lineage-union rule
AUROC 0.877
95% CI 0.842–0.910 · PR-AUC 0.611
Logistic, 12 decision-time features
AUROC 0.882
95% CI 0.845–0.917 · PR-AUC 0.661
Random forest, same 12 features
AUROC 0.902
95% CI 0.875–0.928 · PR-AUC 0.701
Repository-held-out test, 497 pairs. Intervals are bootstrap 95% CIs. PR-AUC is average precision. Cross-agent pairs are enriched, so these are not population rates.
Replaying the top 33.3% of held-out pairs recovers 82.9% of conflicts with the forest and 81.2% with the logistic model. Most of that gain is lineage state, not the authored-overlap features. If lineage metadata is truncated, the paper's stress test drops the logistic AUROC from 0.882 to 0.791 — incomplete history is a reason to run the checker, not a reason to call the pair safe.
When the checker is cheap
Local three-way replay in the released logs has median runtime 0.02 s. The paper does not claim that predicting before git merge-tree saves that checker. If textual replay is local and cheap, run it on everything and discard the predictor. A score is for ordering a costlier chain — checkout, build, tests, a specification — and a clean textual merge is not evidence for any later predicate.
git merge-tree on the two heads and their common base. Deterministic for textual integration only.Scheduling simulation
On 97 repository streams (19,397 pull requests), a decision-time gate removes a median 91.7% of labeled scope collisions (81.2–97.5%) at 65.0% median makespan inflation (35.7–99.2%). Serializing everything removes all of them at 167.3% inflation (114.2–195.7%). An oracle that already knows future overlap does it at 33.4% (20.0–47.4%).
Frozen historical labels. Agents do not react to delay. These are scope-collision counts in a simulator, not a measurement that textual or semantic conflicts were prevented in production.
Path collision and intent collision are different targets. Of 577,045 co-active pairs with file lists, 7.38% share an authored file, and 79.6% of near-duplicate titles share none. Open-time scope forecasting — title, body, timing — is a planning heuristic: a zero-shot model reaches AUC 0.704, and fusing it with open-time features reaches 0.740. That is not a verification result.
Managent leases territory before the write and replays merges with git merge-tree. The numbers on this page are the workshop paper, not a deployment report. Watch merge replay · The earlier study