You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Research: continuous borrow support and exit ownership obligations inspired by JITAtlas #464
Bounded research, backlog. Determine whether existing OwnIR lifetime/loan facts miss a real C# resource-protocol violation because support is checked only at endpoints or caller-visible ownership is missing. Reuse #400 (view invalidation), #426 (completeness), #430 (lowering fidelity), #433 (state oracle), #438 (field aliases); this is not a new general solver.
Primary source
JITAtlas, submitted 18 Nov 2026; full text, sections 3, 6.2, 7.4. Preprint; a journal label alone does not establish acceptance.
The contribution separates point validity, continuous support of an active borrow, and caller-visible ownership at exits. Reported results: 29/29 defects and 29/29 safe controls on a 58-input development benchmark; 92/194 true production reports (47.42% precision). Of 607 production non-reports only 43 complete, 564 incomplete. All three native-labeled frozen-transfer defects were missed. Implementation/replay artifacts are linked in the paper, but raw native logs are incomplete. These limitations rule out treating non-reports as verified safety.
Hypothesis for C#
A borrow/view may become invalid during callback, Dispose, pool Return, or storage mutation even if an owner appears at both endpoints. Track resource identity, support/invalidation event and subsequent use. Managed GC liveness is distinct from resource validity: List/Span invalidation is not necessarily memory UAF. Restoring an owner cannot repair a past invalid use, but a demonstrably fresh borrow may be valid.
Kill-first experiment
Pin Own.NET revision; inventory actual loan/escape/Dispose and exit-transfer semantics before proposing new facts.
Independently confirm one real OSS discrepancy; stop with STOP-NO-ELIGIBLE-GAP if existing rules already suffice.
Reuse 8–12 reviewed bad/ok fixtures: pool Return then use, callback disposing a captured owner, transfer to caller, reacquisition of a fresh view, unrelated resource mutation, conditional invalidation, exceptional exit, opaque callback and forced budget exhaustion.
Preserve order and object identity in source→OwnIR; classify lowering/model gaps separately.
Compare current analyzer against the smallest additional point/interval/exit check. Record changed FP/FN per case, first divergence, explored states and runtime; no aggregate precision claims from a tiny curated corpus.
Unsupported callback/model/budget outcomes remain UNKNOWN/INCOMPLETE. A protocol violation is a diagnostic, not an unreachable state.
Acceptance
Existing representation and overlap audit.
One independently adjudicated source-level gap or documented STOP.
Paired controls and exact reproduction commands.
No regression of confirmed positives and no safe verdict on missing semantics.
GO only for a narrowly scoped follow-up demonstrating marginal value; otherwise close with evidence.
No CPython adapter, refcount runtime, universal callback analysis, or automatic production diagnostic rollout.
Activity