01
Direct avoided cost
Exact cache hits or equivalent eliminated calls can be measured against the provider call that did not happen.
Proof
Savings should survive engineering and finance review.
Varsten separates opportunity estimates from verified savings. The proof model explains the baseline, the optimization, the quality decision, and the dollars that can be claimed.
Proof posture
Traceable
Each claim should connect to provider pricing, workload labels, and an accepted measurement method.
Accepted evidence types
02
Holdback A/B
A controlled slice of traffic keeps the original path so Varsten can compare optimized and baseline behavior.
03
Replay evidence
Approved replay corpora can estimate counterfactual cost and quality when direct holdback is not appropriate.
04
Overhead subtraction
Any approved Varsten processing or extra provider work is subtracted before net savings are claimed.
From request to report
The proof path has to be legible to the people operating the system and the people paying for it.
01
Tag the workload
Requests should carry safe labels such as team, feature, environment, provider, and model.
02
Apply an approved lever
The mechanism records whether savings came from cache, routing, downshift, batching, token trim, or compression.
03
Compare against the baseline
The baseline method depends on the lever and workload risk, not a single universal formula.
04
Publish traceable totals
Reports should reconcile baseline cost, actual spend, net savings, pricing coverage, and attribution coverage.
Integrity checks
Pricing coverage
Required
Savings should be priced against real provider catalogs or accepted contracted rates.
Attribution
Required
Finance needs team or feature labels to allocate savings to the right owner.
Quality proof
Contextual
Quality gates depend on the route, model, and accepted risk of each optimization lever.
Next step
Verify one lever before expanding the rollout.
Start with a narrow workload where the baseline, fallback path, and savings ledger are easy to audit.