Research note 015: simultaneity doesn't matter for v1 fills (proven, not asserted)

Date: 2026-10-16 · Status: engine property, proven by construction plus 4,096 adversarial streams. Closes the simultaneity question notes 003 and 009 deferred: roughly half of all prints share their millisecond with a neighbor, and replay order inside the millisecond is arrival order — venue-opaque, not execution order. Does that phantom precision change any fill?

Question

exec-v1 fills a resting limit order on the first trade print trading through its price. Inside a shared millisecond, "first" is arrival order — knowledge the engine cannot have. If the orderings disagree on fills, the model's audit trail cites evidence it never possessed.

Method

simulate_batched (new in astra-exec, same rules, same errors): consecutive prints sharing a venue millisecond evaluate as one simultaneous batch — fills iff any sound print in the batch trades through, citing the earliest such print. Prints without venue time form singleton batches; gaps flush then void, exactly as in streaming. Then the equivalence proof: directed torture shapes (through-print last in batch, corrupt prints leading, timeless prints interleaved, gaps splitting milliseconds, both sides) plus 4,096 seeded streams mixing all of it — full outcome equality (verdict, price, fee, cited print) on every one.

Result

Identical on all 4,096 streams and all torture shapes: batching can cite a different print in theory, but the first through-print in stream order is always the earliest through-print of its batch, so verdict, price, fee, and cited print all match simulate exactly. The phantom precision was phantom: arrival order never changes what v1 concludes, only what it would have cited had the batch disagreed — and it never does.

Reading

One fact with a boundary. The fill/no-fill question is order-invariant within the millisecond because v1 fills whole orders at the limit on first evidence — existence of a through-print does not depend on order. Anything finer (which print filled how much of the order, queue position inside the batch) is order-dependent, and that is exactly Tier C: the batching machinery now exists, tested, waiting for the model that needs it. Until then the cited print stays arrival-first and says so, and no result depends on it.

Reproduce

cargo test -p astra-exec batching

Two tests: directed torture shapes plus the seeded sweep. No archive, no venue — pure engine property, runnable offline.