{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp12_substrate_denial72_20260805","research_id":"eoa_inverse_innovation_exp12_light_screen_20260805","cell_id":"synchronized_release_dampening__computer_science","search_lanes":{"direct_problem_and_intervention":{"queries":["CI artifact registry recovery thundering herd dependency fetch backlog readiness","package registry recovery retry storm CI builds","artifact repository outage recovery thundering herd CI runners"],"source_ids":["SRC1","SRC2","SRC3"],"no_result_note":null},"synonyms_and_historical_terms":{"queries":["retry storm thundering herd service recovery","cache stampede duplicate request suppression singleflight","exponential backoff jitter token bucket retry overload"],"source_ids":["SRC1","SRC2","SRC4"],"no_result_note":null},"products_practices_and_standards":{"queries":["GitLab dependency proxy packages cache authentication","AWS retry backoff jitter token bucket guidance","HTTP 429 Retry-After standard"],"source_ids":["SRC1","SRC2","SRC3"],"no_result_note":null},"component_combination":{"queries":["CI scheduler artifact proxy token bucket singleflight health recovery leases","package registry cache request coalescing authorization digest retry-after","artifact mirror recovery admission control duplicate fetch coalescing"],"source_ids":["SRC1","SRC2","SRC3","SRC4"],"no_result_note":"No retained source described the complete combination of scheduler-held recovery leases, health-derived cache-miss admission, authorization-compatible artifact coalescing, tenant fairness, and an urgent-build reserve."}},"sources":[{"source_id":"SRC1","title":"Retry Storm antipattern","publisher":"Microsoft","url":"https://learn.microsoft.com/en-us/azure/architecture/antipatterns/retry-storm/","source_type":"OFFICIAL_GUIDANCE","claims_supported":["Frequent client retries can prevent an unavailable or busy service from recovering.","Excessive connection attempts during recovery can overwhelm the service; Microsoft identifies this as a thundering-herd condition.","Recommended controls include bounded retries, exponential backoff, Retry-After signaling, gateway throttling, batching, and connection pooling."]},{"source_id":"SRC2","title":"Timeouts, retries, and backoff with jitter","publisher":"Amazon Web Services","url":"https://aws.amazon.com/builders-library/timeouts-retries-and-backoff-with-jitter/","source_type":"OFFICIAL_GUIDANCE","claims_supported":["Retries add load to an already overloaded dependency and can delay recovery.","Correlated retries can remain synchronized even with capped backoff, while jitter spreads arrivals.","AWS describes locally limiting retries with a token bucket."]},{"source_id":"SRC3","title":"Dependency proxy for packages","publisher":"GitLab","url":"https://docs.gitlab.com/user/packages/package_registry/dependency_proxy/","source_type":"FIRST_PARTY_PRODUCT","claims_supported":["GitLab provides a local package proxy that fetches an upstream package on a miss, stores it, and serves later requests locally.","The proxy uses package-registry permissions and supports authenticated upstream registries.","This establishes an artifact-specific cached dependency path but does not document scheduler-controlled recovery cohorts or health-priced leases."]},{"source_id":"SRC4","title":"singleflight package - golang.org/x/sync/singleflight","publisher":"The Go Authors","url":"https://pkg.go.dev/golang.org/x/sync/singleflight","source_type":"FIRST_PARTY_PRODUCT","claims_supported":["The singleflight package suppresses duplicate function calls by key.","Only one execution is kept in flight for a key; duplicate callers wait for and share its result.","This is close prior art for coalescing identical concurrent cache-fill operations, though it does not supply artifact authorization, integrity, recovery admission, or fairness policy."]}],"problem_evidence":{"status":"PARTLY_SUPPORTED","finding":"The general mechanism is visible: official guidance states that concentrated retries during recovery can overwhelm a service, delay recovery, and produce a thundering herd; artifact dependency proxies create a shared upstream-fetch choke point; and keyed single-flight suppresses duplicate concurrent work. The bounded search did not locate direct incident evidence showing a CI mirror readiness transition releasing an enumerable blocked-build population into a measured synchronized dependency-fetch wave, so the proposal's precise domain-specific episode remains only indirectly supported.","source_ids":["SRC1","SRC2","SRC3","SRC4"]},"closest_prior_art":[{"name":"Retry-storm controls using throttling, bounded exponential backoff, jitter, and Retry-After","source_ids":["SRC1","SRC2"],"overlap":"Directly addresses recovery-delaying retry load and uses randomized timing, explicit backpressure, and rate limiting to protect the recovering dependency.","remaining_difference":"The guidance does not enumerate CI builds blocked on one artifact service, grant tenant-balanced scheduler leases from mirror-health bands, meter new cache misses separately from running builds, or reserve audited capacity for urgent builds."},{"name":"GitLab package dependency proxy","source_ids":["SRC3"],"overlap":"Places an authenticated cache in the CI dependency path, serves cached packages locally, and avoids repeated upstream downloads after a cache fill.","remaining_difference":"The documentation does not describe recovery-specific queueing, health-derived miss tokens, gradual post-readiness release, maximum-wait guarantees, or urgent-build admission."},{"name":"Go singleflight duplicate suppression","source_ids":["SRC4"],"overlap":"Coalesces simultaneous identical-key work into one in-flight execution and a shared result, matching the proposal's duplicate-fetch mediator.","remaining_difference":"It is a general primitive rather than an authorization- and digest-aware artifact proxy, and it does not govern backlog admission, recovery health, tenant fairness, starvation, or priority bypass."}],"prior_art_disposition":"ADJACENT_PRIOR_ART","contrastive_claim_remaining":"For a post-outage CI dependency backlog, controlling the admitted rate of authorization-compatible, non-coalesced upstream cache misses through mirror-health-priced scheduler leases will reduce peak mirror queueing and errors more than a matched fixed global build-concurrency cap while allowing unrelated compute-only jobs to continue, without worsening bounded-wait, fairness, integrity, or authorization outcomes. The component techniques are established; the remaining testable distinction is their coordinated application at the CI scheduler–artifact-proxy boundary and the unique-fetch-rate mediator.","contrastive_claim_falsifier":"The claim is falsified if a matched isolated replay shows that a fixed global build cap performs equivalently on mirror queue depth, errors, backlog completion, and safeguards without delaying compute-only work; if leases and coalescing do not reduce unique upstream fetch starts or requests per digest; if evenly dispersed nonduplicate demand still overloads the mirror; or if integrity, authorization, starvation, maximum-wait, or urgent-reserve guards fail.","gates":{"adequate_source_search":{"status":"PASS","rationale":"The bounded search covered direct terminology, retry-storm and cache-stampede synonyms, relevant artifact-proxy products and HTTP practices, and combinations of admission control with duplicate suppression. Exactly four opened sources from four publishers were retained, all official or first-party. This is adequate only for a coarse screen, not a systematic, patent, or freedom-to-operate search.","source_ids":["SRC1","SRC2","SRC3","SRC4"]},"supported_problem":{"status":"PASS","rationale":"Official guidance directly supports recovery-overwhelming retry storms, while first-party artifact-proxy and single-flight documentation supports the shared fetch path and duplicate-work mechanism. Because the exact CI readiness-release pattern lacks direct incident evidence, the problem is PARTLY_SUPPORTED.","source_ids":["SRC1","SRC2","SRC3","SRC4"]},"distinct_testable_claim":{"status":"PASS","rationale":"The proposal makes a discriminating comparison with a fixed build cap: meter unique upstream cache misses while compute-only work continues, and measure the proposed concurrency and duplication mediators alongside safeguards.","source_ids":["SRC1","SRC2","SRC3","SRC4"]},"bounded_next_test":{"status":"PASS","rationale":"The specified test is limited to nine trace-driven runs across three configurations and three fixed seeds against one isolated emulator, with fixed capacity, defined measurements, explicit advance criteria, and no production mutation.","source_ids":["SRC1","SRC2","SRC4"]},"no_obvious_safety_or_authority_stop":{"status":"PASS","rationale":"The authorized first step is an isolated replay using de-identified identifiers and synthetic credentials. It preserves verification and authorization, excludes production and unauthorized third-party load, and defines halt and rollback conditions. GitLab's documented permission-sensitive proxy behavior reinforces the need to keep authorization scopes in the coalescing key.","source_ids":["SRC3","SRC4"]}},"screen_survival":true,"world_novelty_boundary":"This bounded public-web screen found established adjacent practices for recovery retry throttling, backoff and jitter, package dependency caching, and keyed single-flight duplicate suppression, but no retained source describing their complete recovery-leased CI design. That result does not establish world novelty, patentability, freedom to operate, market size, prevalence, expert acceptance, or realized value."}