Crypto realism challenge: validate L2 fill models against execution evidence #62

The challenge

Can an L2 execution model predict real passive fills better than ordersim's evidence-constrained virtual MBO reconstruction?

PR #61 adds an evidence study and a deterministic minimum-flow reconstruction:

ending quantity = starting quantity + adds - cancels - traded quantity

That identity recovers the minimum aggregate flow needed to reproduce each observed depth endpoint. It cannot identify the true placement of cancellations or additions inside the hidden FIFO queue.

Useful evidence

Contributions would be especially valuable if they provide one of:

• paired L2 and order-level L3/MBO observations for the same venue and interval;
• a published or independently justified cancellation-allocation model;
• passive-fill sensitivity results across the two existing queue policies;
• evidence about Binance zero-value @trade messages that does not assume undocumented fields are stable;
• a reproducible alternative policy with tests and a clear statement of assumptions.

Current bounds

• queue-conservative: required additions appear before the first trade at a level; cancellations remove newest modeled liquidity first.
• queue-optimistic: additions appear only when needed; cancellations remove oldest modeled liquidity first.

Both policies reproduce the same observed L2 endpoints. A result that changes materially between them should be reported as model-sensitive, not as a precise observed queue fill.

A comparable submission

A useful comparison should:

1. consume sequenced L2 endpoints and individual trades;
2. disclose event-ordering, cancellation-allocation, hidden-liquidity, latency, and market-impact assumptions;
3. preserve observed book states or report every divergence;
4. evaluate on paired native L3 data or actual passive-order outcomes not used to fit the model;
5. report fill classification, filled-quantity error, and time-to-fill error on held-out intervals;
6. publish enough code and aggregate output for independent reproduction.

Matching an L2 endpoint is an integrity requirement, not evidence that a queue model predicts execution. The useful result is a model that performs better against hidden order-level truth or realized fills.

Please describe the venue, source schema, timestamp semantics, and any redistribution restrictions with submitted evidence. Do not upload licensed market data unless redistribution is explicitly permitted.