MODULE 23 · 6 HOUR BUILD
A routing replay and economics report
Build an offline replay tool comparing routing policies on identical tasks. Produce a report of task success, spending, escalation, and latency distribution with every synthetic assumption visible.
Build evidence Record your actual checks, results, and limitations.
Build it in stages
- Run the synthetic seed and explain why each threshold routes each case.
- Define an input schema for task IDs, route outcomes, usage, and latency observations.
- Add exact cost accounting with a versioned rate table and explicit missing-data behavior.
- Compare direct and cascading routes using the same held-out tasks and a fixed verifier.
- Add per-task paired differences, segment summaries, and request-level p95.
- Write a recommendation bounded by the observed dataset and document quality or latency constraints that disqualify a cheaper policy.
Your acceptance criteria
Use these as your project review. Record commands, outputs, and failure cases in your repository.
- Every policy is evaluated on identical task IDs.
- Total cost reconciles to unique call records without cached-token double counting.
- All reported percentiles state estimator, sample size, and timeout treatment.
- Synthetic results are clearly labeled and no real provider prices are invented.
A working starting point
The seed runs as supplied. Extend it to satisfy the full brief. It is a teaching starting point, not a finished portfolio submission.
main.py
python
from math import ceil
from statistics import mean
CASES = [
{"id": "a", "difficulty": 0.1, "small_ok": True, "large_ok": True},
{"id": "b", "difficulty": 0.4, "small_ok": True, "large_ok": True},
{"id": "c", "difficulty": 0.7, "small_ok": False, "large_ok": True},
{"id": "d", "difficulty": 0.9, "small_ok": False, "large_ok": False},
]
ROUTES = {"small": {"cost": 0.002, "latency": 300},
"large": {"cost": 0.020, "latency": 900}}
def replay(threshold):
records = []
for case in CASES:
route = "large" if case["difficulty"] >= threshold else "small"
records.append({"id": case["id"], "route": route,
"ok": case[route + "_ok"], **ROUTES[route]})
ordered = sorted(record["latency"] for record in records)
return {"threshold": threshold,
"success": mean(record["ok"] for record in records),
"cost": sum(record["cost"] for record in records),
"p95": ordered[ceil(0.95 * len(ordered)) - 1],
"large": sum(record["route"] == "large" for record in records)}
def main():
print("SYNTHETIC replay: 4 fixed tasks; invented route prices")
print("threshold success total_cost p95_ms large_calls")
for threshold in (0.3, 0.6, 0.9):
result = replay(threshold)
print(f"{result['threshold']:.1f} {result['success']:.2f} "
f"{result['cost']:.3f} {result['p95']} {result['large']}")
if __name__ == "__main__":
main()
Push it further
Add recorded redacted OpenTelemetry traces and compare critical-path attribution with the policy-level latency change.