It is kept per account, since accounts are answered separately and one can be struggling while another is not.
What counts as a request
One result is recorded per request taking the retrying path, not per attempt, so a call that retried counts once, the waiting included in how long it took.
Not everything is recorded. Asking the venue the time, its trading rules or the order book records no result, each going straight out, and a missing order against the simulated venue is skipped. An order the venue rejects is recorded as a success, because the venue answered: this measures whether the venue replied, leaving what the order did to others.
A replay's venue measures none of this. It hands back a report with every figure at zero — which the decision side reads as an account in good health, rather than as an account it knows nothing about.
What is reported
A fixed number of the most recent results are kept and the oldest fall out. From those, two figures are worked out when asked for: a slow-end measure of how long requests are taking, and the share that failed.
Beside them sits the run of failures in a row — a success resets it, a failure lengthens it — and two figures supplied from outside: how much of the venue's allowance has been used, and how far the clock has drifted.
Who reads it
The report goes to the decision side, which can restrict an account to closing out, or halt it entirely, on what the report says.
So health is not only a picture. It is read, and it can stop work being accepted.
Related
Source evidence
---
row: P3-15
baseline: 49aa659
created: 2026-09-30 09:14 UTC
modified: 2026-10-12 00:15 UTC
evidence:
- claim: "One result is recorded per request that goes through the retrying path and is not among those skipped, across its retries, with the waiting counted in the time"
source: "`binance_adapter.py` 578-604 for one that answers, or for one that does not either 607, or 633, or 655, or 658, or 673"
- claim: "Recording is skipped for a missing order against the simulator"
source: "`binance_adapter.py` 632-634"
- claim: "An order the venue rejects is recorded as the venue having answered"
source: "`binance_adapter.py` 635-656"
- claim: "Asking the time records no result"
source: "`binance_adapter.py` 2353-2372"
- claim: "A fixed number of the most recent are kept, the oldest falling out"
source: "`health_tracker.py` 75-76 into the bounded stores made at 44-45, with the count at 21"
- claim: "A window size that is not a whole number above zero is refused"
source: "`health_tracker.py` 39-42"
- claim: "A success resets the run of failures; a failure lengthens it"
source: "either `health_tracker.py` 78 or 80"
- claim: "The slow-end figure and the failed share are worked out when asked for"
source: "`health_tracker.py` 123-129 and 136-140, assembled at 110-115"
- claim: "The limit figure counts the share of the allowance already used"
source: "`binance_adapter.py` 2330, supplied at 2394-2403 and passed through at `health_tracker.py` 114"
- claim: "A replay's venue returns a report with every figure at zero"
source: "`replay_venue_adapter.py` 427 with the defaults at `health_snapshot.py` 33-37"
- claim: "The decision side reads an all-zero report as an account in good health, every threshold standing above zero as supplied"
source: "Nexus `health_evaluator.py` 165-184, the thresholds at 81-93"
- claim: "The report is handed to the decision side, which can restrict or halt an account on it"
source: "`launcher.py` 513-528 into Nexus `health_loop.py` 157-177 and `health_evaluator.py` 165-184"
---