{ "artifact": "d1_engine_hit_rate", "sprint": "D1", "claim": "vLLM's own prefix-cache counters, read from /metrics during the S46 protocol, give the engine's hit rate at the capacity the offload oracle modelled -- so the modelled 3.9% and the measured 30.5% knee uplift can be reconciled or shown irreconcilable.", "cost_usd": null, "method": "Engine prefix-cache counters (exact Prometheus names, last scrape per load point, merged across client shards) from the paired_sweep raw; decision rule as sealed in the config before the run.", "attempts": { "1": "results/d1_engine_hit_rate.attempt1_no_counters.json -- sharded merge dropped the counters", "2": "results/d1_engine_hit_rate.attempt2_timestamp_as_counter.json -- prefix match took the _created timestamp", "3": "results/d1_engine_hit_rate.attempt3_broken_dataclass.json -- parser inserted inside the dataclass body", "4": "this certificate -- two-point ladder, module repaired and pinned" }, "modelled_gpu_only_hit_rate": 0.038678, "oracle_capacity_tokens": 65536, "engine_block_tokens": 16, "trace_block_tokens_served": 64, "granularity_note": "The engine counts 16-token blocks; the oracle counts 64-token trace blocks. One trace-block hit is up to four engine-block hits, so engine and modelled rates compare only up to that factor; the decision rule was sealed on the engine's number.", "arms": { "apc_off": { "knee_qps": null, "points": [ { "qps": 8.0, "slo_met": true, "stationary": false, "prefix_cache_hits_total": 0.0, "prefix_cache_queries_total": 0.0, "hit_rate_engine_blocks": null, "engine_metrics_available": true, "raw_total_lines": { "vllm:prefix_cache_queries_total": 0.0, "vllm:prefix_cache_hits_total": 0.0, "vllm:external_prefix_cache_queries_total": 0.0, "vllm:external_prefix_cache_hits_total": 0.0 }, "created_timestamps_present": true }, { "qps": 9.0, "slo_met": true, "stationary": false, "prefix_cache_hits_total": 0.0, "prefix_cache_queries_total": 0.0, "hit_rate_engine_blocks": null, "engine_metrics_available": true, "raw_total_lines": { "vllm:prefix_cache_queries_total": 0.0, "vllm:prefix_cache_hits_total": 0.0, "vllm:external_prefix_cache_queries_total": 0.0, "vllm:external_prefix_cache_hits_total": 0.0 }, "created_timestamps_present": true } ], "hit_rate_engine_blocks_by_point": [], "hit_rate_engine_blocks_max": null, "hit_rate_engine_blocks_min": null, "queries_total_max": 0, "hits_total_max": 0 }, "apc_on": { "knee_qps": 9.0, "points": [ { "qps": 8.0, "slo_met": true, "stationary": true, "prefix_cache_hits_total": 399072.0, "prefix_cache_queries_total": 1055688.0, "hit_rate_engine_blocks": 0.378021, "engine_metrics_available": true, "raw_total_lines": { "vllm:prefix_cache_queries_total": 1055688.0, "vllm:prefix_cache_hits_total": 399072.0, "vllm:external_prefix_cache_queries_total": 0.0, "vllm:external_prefix_cache_hits_total": 0.0 }, "created_timestamps_present": true }, { "qps": 9.0, "slo_met": true, "stationary": true, "prefix_cache_hits_total": 815968.0, "prefix_cache_queries_total": 2111376.0, "hit_rate_engine_blocks": 0.386463, "engine_metrics_available": true, "raw_total_lines": { "vllm:prefix_cache_queries_total": 2111376.0, "vllm:prefix_cache_hits_total": 815968.0, "vllm:external_prefix_cache_queries_total": 0.0, "vllm:external_prefix_cache_hits_total": 0.0 }, "created_timestamps_present": true } ], "hit_rate_engine_blocks_by_point": [ 0.378021, 0.386463 ], "hit_rate_engine_blocks_max": 0.386463, "hit_rate_engine_blocks_min": 0.378021, "queries_total_max": 2111376.0, "hits_total_max": 815968.0 } }, "control_apc_off_zero_hits": true, "hit_rate_engine_blocks_apc_on": 0.386463, "bridge": { "distinct_ids": 13712, "distinct_texts_legacy_mapping": 33, "distinct_texts_fixed_mapping": 13712, "max_achievable_on_ids": 0.168718, "max_achievable_on_legacy_text_stream": 0.729858, "lru_at_capacity_on_ids": 0.038678, "lru_at_capacity_on_legacy_text_stream": 0.410488, "lru_at_capacity_on_fixed_text_stream": 0.038678, "capacity_trace_blocks": 1024 }, "bridge_collapsed_the_stream": true, "engine_rate_matches_collapsed_stream": true, "branch": "bridge_collision", "verdict": "BRIDGE COLLISION -- neither pre-registered branch. The engine's hit rate 0.3865 EXCEEDS the modelled stream's theoretical maximum 0.1687, so it cannot be a property of the trace. The bridge's legacy text mapping collapsed 13712 distinct block ids into 33 distinct texts; GPU-only LRU on THAT stream at this capacity is 0.4105, which is what the engine measured. The offload oracle's 0.0387 was right about the trace; S46's 30.5% knee uplift was measured on a 33-text stream and is not evidence about Mooncake. S50 stays unauthorised; S46 must be re-run on the fixed bridge (expected engine rate near 0.039).", "s50_authorised_to_spend": false, "s46_stands_as_real_trace_evidence": false, "does_not_prove": [ "n=1. The knee ratio from this run is a consistency check against S46, not evidence.", "A downscaled trace (block_tokens 64 vs declared 512). The hit rate is on the downscaled block stream, which is the one the oracle modelled and S46 measured -- a like-for-like.", "The engine counts hits in its own 16-token blocks; the oracle counts 64-token trace blocks. A hit rate here is comparable to the model only up to that granularity, and the analysis must say by how much.", "One model, one GPU class, one trace. Whether the finding generalises is Track P's P2.", "n=1 container per arm; two load points, both above the APC-off knee (non-stationary there), which does not affect a hit-rate reading.", "The counters are totals since boot and include the warm-up requests." ], "generated_utc": "2026-09-02T15:13:54.218536+00:00", "git_rev": "c16fd85", "git_dirty": true, "run_id": "304f00e2b3cb4b1f", "host_kind": "Darwin-arm64", "provenance_schema": "gpu-cert-provenance/1" }