Delta is one recording and a set of questions asked against it. The questions below are grouped by the decision each supports, and each one says whether it answers today — a services page that reads the same before and after a thing is built is not telling you anything.
Send the fills a backtest assumed. Get back, per fill, whether the market would actually have traded with you, and what it would really have cost. This is reachable over HTTP now.
| What you get | What it means |
|---|---|
| Per-fill verdicts, with the reason | For every fill you send: whether it would have happened, how much of it, what it really cost, and which check bound. A row we cannot check comes back unknown and is excluded from the totals, never counted as fine. |
| Book condition at the moment of the fill | How far each book was from its own normal depth, in either direction, banded. A flag on the fill, never a deduction from it. |
| Coverage, including what cannot be answered where | Which venues, which instruments, at what resolution, over what history - and which questions stop being answerable where, separated into what cannot be measured and what is not recorded yet. |
| An answer that keeps | Every response carries the engine revision, the baseline window and how the data reached it. The baseline ends on the fill's own day, so the same fills return the same numbers next year. |
| Service | What it answers | Status |
|---|---|---|
| Impact and recovery Phase 5 | Expected price displacement for an order of this size into this book, and how much reverts. A cost forecast rather than a capacity limit — the input to sizing a parent order or setting a participation rate. | planned |
| Fill probability Phase 4 | P(fill | book state, size, queue position, recent flow, regime) — a distribution rather than an assumption, so a simulation's implicit certainty becomes visible and sizable. | planned blocked on the negative class, which is the labelling pass rather than the model |
| Quote survival Phase 3 | The probability an opportunity still exists at your latency, at your size. The output is a sensitivity: how fast must you be for this to exist at all? | planned the deterministic survival check runs today at a latency you specify; the probability distribution is the work |
| Service | What it answers | Status |
|---|---|---|
| Order-flow toxicity Phase 2 | Scores incoming flow for how likely it is to be informed, in real time, before and as a quote is exposed to it. For a principal desk the central problem is not whether it gets filled but who fills it, and the difference between those is the entire margin. This is the service where non-order-book data earns its place. | planned blocked on Phase 1, which establishes the relationship this is sold into |
| Service | What it answers | Status |
|---|---|---|
| Venue integrity Phase 6 | Scores venues on trade-size distributions, inter-trade timing, and the relationship between reported volume and observable price impact. No labels exist for this anywhere; it is only possible against a simultaneous cross-venue baseline. Delivered as a score with its evidence, never as an accusation. | planned only possible against a simultaneous cross-venue recording, which is the part that already exists |
| Service | What it answers | Status |
|---|---|---|
| Best-execution evidence Phase 1 | For any executed order, a reconstruction of what was genuinely available at that instant across venues — the counterfactual the obligation describes, currently satisfied with policy documents rather than measurement. Reproducible, retained and stamped. | in build the deterministic checks and the pinned baseline answer through the API today; retention, retrieval and the regulator-facing record are not built |
| Service | What it answers | Status |
|---|---|---|
| Synthetic markets Phase 7 | Generated order-book evolutions, so a strategy or an execution algorithm is evaluated against thousands of markets that could plausibly have happened rather than the single path that did. | planned |
The same instruments, recorded across venues at the same time. That is what lets a model trained on one venue be tested on another sharing none of its infrastructure, and it is not something anyone can sell you afterwards — it had to be recorded while it was happening. What is recorded, and what cannot be answered where →