{"schema":"seenrelay-controlled-proof-atlas-v1","updated_at":"2026-10-02","evidence_boundary":{"evidence_type":"controlled_first_party_benchmark","includes_first_party_real_operating_workload":true,"customer_roi_claim":false,"natural_prevalence_claim":false,"net_customer_savings_claim":false,"best_available_baseline_claim":false,"tool_necessity_is_separate_gate":true,"native_controls_first":true},"rows":[{"id":"company-fleet-shared-evidence","evidence_label":"First-party real operating workload + synthetic external-client rehearsal","baseline_quality":"TOOL_NECESSITY_PASS_BEST_BASELINE_UNPROVEN","tool_necessity":"PASS_CURRENT_EXTERNAL_EVIDENCE","surface":"Multi-agent shared evidence / independent reasoning","baseline":"$0.16499684","seenrelay":"$0.05560972","avoided":"$0.10938712 gross","reduction":"66.30%","executions":"13 → 4 paid web-search actions","execution_label":"13 → 4 paid web-search actions · 8 → 10 model Responses","semantic":"PASS","repeatability":"1 complete paired run","adoption_path":"/case-studies/company-fleet","link_label":"Case study →","note":"Eight department agents across two tenant identities performed useful current ecosystem research. SeenRelay shared one same-tenant evidence packet per organization while preserving all eight independent role analyses. Gross provider cost fell from $0.16499684 to $0.05560972. SeenRelay OpsCo was first-party real operating work; the second tenant was explicitly synthetic. Local coordination/integration overhead was not monetized, so this is not customer ROI, net customer savings, external adoption or a universal 66% claim."},{"id":"hosted-compute-64gb","baseline_quality":"MECHANISM_ONLY_LOCAL_WINS","surface":"Hosted compute / 64 GB tool session","baseline":"$1.9241253","seenrelay":"$0.4820025","avoided":"$1.4421228","reduction":"74.95%","executions":"4 → 1","semantic":"PASS","repeatability":"3/3","adoption_path":"/clients","note":"Mechanism-only benchmark: the exact task was deterministic SHA-256 and could be done locally. This row proves duplicate hosted-resource coalescing, not that a 64 GB container was necessary for SHA-256."},{"id":"agentic-search","baseline_quality":"MECHANISM_ONLY_SOURCE_NATIVE_WINS","surface":"Agentic web search — explicit shared-snapshot contract","baseline":"4 top-level executions","seenrelay":"1 top-level execution","avoided":"42.84%–85.73%","reduction":"66.88% mean","executions":"4 → 1","semantic":"PASS","repeatability":"3/3","adoption_path":"/ai-agent-cost-optimization","note":"Mechanism-only benchmark for this exact task: all callers accepted one shared fresh provider-search snapshot, but the requested latest PyPI package version is available from PyPI directly. This proves shared-search coordination, not web-search necessity. Other agentic-search workloads may require k>1 or k=N."},{"id":"freshness-firewall","baseline_quality":"BEST_BASELINE_UNPROVEN","surface":"Decision-time freshness","baseline":"6 provider reads","seenrelay":"2 provider reads","avoided":"4 provider reads","reduction":"66.7%","executions":"6 → 2","semantic":"PASS","repeatability":"3/3","adoption_path":"/quickstart","note":"Compared with Firecrawl provider-native maxAge on the same controlled schedule: native path stayed at 6 calls / 6 credits, while SeenRelay used 2 calls / 2 credits in 3/3 runs. Hard max-age policy remained satisfied; no claim of hidden source-change detection inside the allowed age window."},{"id":"exact-web-extraction","baseline_quality":"BEST_BASELINE_UNPROVEN","surface":"Exact web extraction","baseline":"4 Firecrawl credits","seenrelay":"1 Firecrawl credit","avoided":"3 Firecrawl credits","reduction":"75%","executions":"4 → 1","semantic":"PASS","repeatability":"3/3","adoption_path":"/clients","note":"Exact same-boundary markdown extraction. A separate structured JSON contract failed semantic equality and remains a negative result."},{"id":"weather-dynamic-state","baseline_quality":"MECHANISM_ONLY_LOCAL_WINS","surface":"Weather downstream analysis","baseline":"4 paid analyses","seenrelay":"1 paid analysis","avoided":"3 paid analyses","reduction":"75%","executions":"4 → 1","semantic":"PASS","repeatability":"3/3","adoption_path":"/quickstart","note":"Mechanism-only benchmark: real Open-Meteo state stayed unchanged, but the paid classifier applied fixed rules that the harness itself could compute locally. This proves state-keyed recomputation control, not best-baseline weather economics."},{"id":"weather-transition","baseline_quality":"BEST_BASELINE_UNPROVEN","surface":"Weather state transition","baseline":"4 paid analyses","seenrelay":"2 paid analyses","avoided":"2 paid analyses","reduction":"50%","executions":"4 → 2","semantic":"PASS","repeatability":"3/3","adoption_path":"/quickstart","note":"Controlled A,A,B,B transition forced a new analysis at B and recorded zero stale reuse after the change."},{"id":"retail-real-page","baseline_quality":"NATIVE_FIRST_CONDITIONAL","surface":"Retail extraction guarded by price state","baseline":"4 Firecrawl credits","seenrelay":"1 Firecrawl credit","avoided":"3 Firecrawl credits","reduction":"75%","executions":"4 → 1","semantic":"PASS","repeatability":"3/3","adoption_path":"/clients","note":"Native-first conditional: the direct Apple read already resolves the monitored price/capacity state. Paid Firecrawl extraction is relevant only when the caller needs the richer extraction artifact; if state alone is sufficient, SeenRelay should self-reject."},{"id":"retail-transition","baseline_quality":"BEST_BASELINE_UNPROVEN","surface":"Retail price / availability transition","baseline":"4 paid analyses","seenrelay":"2 paid analyses","avoided":"2 paid analyses","reduction":"50%","executions":"4 → 2","semantic":"PASS","repeatability":"3/3","adoption_path":"/clients","note":"Controlled authoritative A,A,B,B retail fixture. The changed price/availability fingerprint forced a new analysis before reuse resumed; this is not a live Apple price-change claim."},{"id":"inventory-native-comparison","baseline_quality":"NATIVE_FIRST_CONDITIONAL","surface":"Inventory / availability downstream extraction","baseline":"4 Firecrawl credits","seenrelay":"1 Firecrawl credit","avoided":"3 Firecrawl credits","reduction":"75%","executions":"4 → 1","semantic":"PASS","repeatability":"3/3","adoption_path":"/use-cases","note":"Real Adafruit inventory state. Firecrawl provider-native maxAge also stayed at 4 calls / 4 credits; SeenRelay used the retailer Product API as a cheap opaque state token and performed one paid extraction. If that native API already answers the entire need, SeenRelay should not be used."},{"id":"inventory-transition","baseline_quality":"BEST_BASELINE_UNPROVEN","surface":"Inventory state transition","baseline":"4 paid analyses","seenrelay":"2 paid analyses","avoided":"2 paid analyses","reduction":"51.58% mean","executions":"4 → 2","semantic":"PASS","repeatability":"3/3","adoption_path":"/use-cases","note":"Controlled A,A,B,B stock/availability transition. Each changed state forced a new paid analysis before reuse resumed; gross modeled reduction ranged from 50% to 53.1%. This is not a live retailer stock-change claim."},{"id":"litellm-cold-concurrency","baseline_quality":"MECHANISM_ONLY_LOCAL_WINS","surface":"AI gateway cold concurrent cache misses","baseline":"4 upstream executions","seenrelay":"1 upstream execution","avoided":"3 upstream executions","reduction":"74.84%","executions":"4 → 1","semantic":"PASS","repeatability":"3/3","adoption_path":"/clients","note":"Mechanism-only benchmark: against LiteLLM Proxy 1.103.2 + Redis exact cache, four cold concurrent misses caused four upstream executions and SeenRelay single-flight caused one. The exact benchmark output was a fixed constant, so this isolates the cold-concurrency gap rather than proving model-call necessity."},{"id":"news-live-event","baseline_quality":"TOOL_NECESSITY_PASS_BEST_BASELINE_UNPROVEN","tool_necessity":"PASS_DERIVED_BRIEFING","surface":"News / event downstream briefing","baseline":"4 paid briefings","seenrelay":"1 paid briefing","avoided":"$0.0546–$0.1010 modeled","reduction":"79.61% mean","executions":"4 → 1","semantic":"PASS","repeatability":"3/3","adoption_path":"/use-cases","note":"Real Hacker News event authority plus real OpenAI web-search briefing. In 3/3 stable-state runs, four paid briefings became one and gross avoided cost was $0.05462206–$0.10098029 per group (70.95%–90.00%). HN remained the native state authority; independent customer prevalence is not claimed."},{"id":"news-transition","baseline_quality":"TOOL_NECESSITY_PASS_BEST_BASELINE_UNPROVEN","tool_necessity":"PASS_DERIVED_BRIEFING","surface":"News / event state transition","baseline":"4 paid briefings","seenrelay":"2 paid briefings","avoided":"$0.0116–$0.0664 modeled","reduction":"49.93% mean","executions":"4 → 2","semantic":"PASS","repeatability":"3/3","adoption_path":"/use-cases","note":"Controlled A,A,B,B transition using two real current Hacker News events. The changed event forced a fresh briefing before reuse resumed; naturally observed rank transition is not claimed."},{"id":"shared-image-artifact","baseline_quality":"CONTRACT_DEPENDENT","surface":"Shared generated image artifact","baseline":"≥$0.212","seenrelay":"≥$0.053","avoided":"≥$0.159","reduction":"75%","executions":"4 → 1","semantic":"PASS","repeatability":"3/3","adoption_path":"/clients","note":"Only valid when all callers explicitly accept one shared generated artifact and diversity is not required."}],"negative_results":[{"id":"firecrawl-structured-json-semantic-mismatch","surface":"Structured JSON extraction","result":"NO PASS","note":"Provider credits fell mechanically, but authoritative outputs disagreed on title/headline field semantics. SeenRelay does not count this as validated savings."},{"id":"ocr-repeatability-semantic-instability","surface":"OCR / document extraction","result":"NO PASS","note":"One exact trial passed, but repeatability failed because independent OCR outputs varied semantically. OCR is not promoted."}],"canonical_origin":"https://seenrelay.com","urls":{"human":"https://seenrelay.com/proof","machine":"https://seenrelay.com/proof.json","integrations":"https://seenrelay.com/clients","quickstart":"https://seenrelay.com/quickstart"}}