Temporal Cut
BM25 tells us what is relevant. HydraDB decides what may ground the answer.
One real benchmark conflict, live. The same ranking on both sides — on the right, the claim graph removes the document that still asserts the retired value and backfills the next candidate from the same ranking. Same model, same prompt, same document count.
For EXP-002 (eu-west→us-east), what egress cost rate and measurement basis does the cost penalty catalog use?
BM25 · relevant
Cut
HydraDB · currently valid
Plain retrieval
model-judged incorrectValue left in its context
+$0.05 per 1k tokens (~$180/hour at 1000 rps)
Based on the fallback experiment matrix (Document 1), EXP-002 uses the **egress-cost-delta** as its cost-metric. The measurement basis for this is specified in the telemetry recipe as: - **egress_gib_rate{experiment_id="EXP-002"}** — sourced from the network collector plus sampled trace attribution - Additional instrumentation requires emitting egress bytes by **src_region->dst_region at 1-minute resolution** for cross-region tests The document notes this egress metric is "new," and no specific dollar rate per GiB is provided in the catalog excerpt.
Temporal Cut
model-judged correctValue left in its context
+$0.085 per GiB, GiB-based billing
Per the canon claim graph, the current EXP-002 eu-west→us-east egress_cost_rate is **$0.085 per GiB**, using GiB-based billing as the measurement basis. (The previously used value of $0.05 per 1k tokens, ~$180/hour at 1000 rps, has been superseded and is retired.)
Same BM25 · same model · same prompt · same 10-document budget