What does preflight cost?
This experiment compares a policy-screened direct fetch with an AccessGraph-assisted fetch. Both arms stop on restrictions. Results are operator-run tests, not customers or external adoption.
Observed results
| Successful paired tasks | 100 / 100 target |
|---|---|
| Origins screened | 12 |
| Direct fetch mean | 261 ms |
| Assisted fetch mean | 702 ms |
| Cache hits | 45 |
| Additional preflight requests | 330 |
The initial experiment shows added preflight overhead, not demonstrated retrieval savings. Use this pilot where access-policy evidence and correct stops are valuable. It is not established as a speed optimizer or a globally cheapest router.
Every screened origin
| Target | Decision | Reason |
|---|---|---|
| example.com | route_available | lowest_observed_cost_eligible_route |
| example.org | route_available | lowest_observed_cost_eligible_route |
| example.net | route_available | lowest_observed_cost_eligible_route |
| json-schema.org | route_available | lowest_observed_cost_eligible_route |
| schema.org | route_available | lowest_observed_cost_eligible_route |
| modelcontextprotocol.io | conditional | redirect_requires_new_preflight |
| a2a-protocol.org | route_available | lowest_observed_cost_eligible_route |
| llmstxt.org | route_available | lowest_observed_cost_eligible_route |
| www.rfc-editor.org | route_available | lowest_observed_cost_eligible_route |
| www.w3.org | conditional | page_policy_review_required |
| www.sqlite.org | route_available | lowest_observed_cost_eligible_route |
| www.postgresql.org | route_available | lowest_observed_cost_eligible_route |
Method and limitations
Ten paired reads per eligible origin, one-second spacing between pairs, alternating arm order, no redirects, public-DNS checks, response limits and five-minute evidence expiry. Common policy screening is reported separately. Additional candidate origins may be screened to reach the target; skipped origins stay in the report.
- Local engine and local cache; not hosted D1 latency.
- Baseline was policy-screened first; common screening requests reported separately.
- Small selected public documents, repeated URLs, one machine and sequential runs; not a representative production workload.
- No LLM inference, task reasoning, paid access or financial savings measured.
- Every screened origin, including restrictions and additional candidates, is retained. Cached permissions may change before expiry; execution restrictions stop both arms.
Run scripts/benchmark.mjs from the source release to reproduce a new experiment. It makes real network requests. --resume continues an interrupted saved sample without redoing recorded origins. Results depend on publisher policy, network conditions and cache headers. No financial savings, independent customers or production latency are inferred.
Download raw results and timestamps · Source and benchmark script · How to handle a stop