{ "schema": "anvil-serving.benchmark-decision-summary/v1", "campaign_id": "2026-11-01-swift-tp1-gpu0", "complete": false, "identity": { "model_repository": "ukisai/Swift-1.5-Qwen3.8-27b", "model_revision": "b4c84d42903a8646b25857eb2827288d92ed85a4", "runtime_image": "lmsysorg/sglang@sha256:37bbbd3444732a464bbc68dee4fb0164e0ce9e18e2f027f3fc967f1152d3c262", "engine_revision": "0bcd822377da7b5718e674eaf9c870d349424dd1", "topology": "single RTX PRO 6000 Blackwell Max-Q GPU; TP1", "final_recipe": "BF16 weights, FP8 KV, FP32 Mamba state, MTP3, 262144 configured context, 16384 output reserve, medium reasoning, image limit 8" }, "workload": { "historical_comparison": "ordinary-summary with no output-canary assertion; three repetitions per cell", "fresh_quality": "direct protocol, deterministic image, and native agentic suites" }, "measurement": { "path": "direct local endpoint; retained historical ordinary-summary comparison and fresh direct quality probes", "warm_cold_state": "not recorded uniformly across scopes", "sample_count": 432, "statistics": [ "native nearest-rank p50/p95 for historical rows", "sampled p95 is not a production tail guarantee" ] }, "gates": [ { "name": "historical MTP/no-spec ordinary-summary comparison", "status": "completed", "evidence": "raw/r4-comparison-decision.json" }, { "name": "fresh image qualification", "status": "passed 12/12 plus ordered eight-image probe 2/2", "evidence": "raw/images.json" }, { "name": "fresh native agentic", "status": "passed 18/18", "evidence": "raw/agentic-native-evidence.json" }, { "name": "fresh native agentic pilot", "status": "passed 6/6", "evidence": "raw/agentic-pilot-native-evidence.json" }, { "name": "fresh C1 long-context", "status": "passed 36/36; all stop", "evidence": "raw/context-native/native-evidence.json" }, { "name": "fresh coding", "status": "passed fixed campaign floor: 4/5 resolved; all 5 independently graded", "evidence": "raw/swe-native/native-evidence.json" }, { "name": "fresh C2 long-context", "status": "passed 2/2 concurrent exact-answer cases; retained overlap and allocator samples", "evidence": "raw/context-concurrent/result.json" }, { "name": "fresh stability serial unique-prefix", "status": "passed 3/3; diagnostic only, not performance ranking", "evidence": "raw/stability-serial-unique.json" }, { "name": "fresh stability overlap unique-prefix successor", "status": "passed 3/3 under contender-stream-active/v1; diagnostic only", "evidence": "raw/stability-overlap-unique-r2.json" }, { "name": "fresh stability overlap repeat-prefix successor", "status": "passed 3/3 under contender-stream-active/v1; diagnostic only", "evidence": "raw/stability-overlap-repeat-r2.json" }, { "name": "fresh final capacity", "status": "passed 10/10 cells and 72/72 requests; separate fresh final activation population", "evidence": "raw/capacity-summary.json" }, { "name": "fresh resource and deep health closure", "status": "passed; continuous observer closed cleanly", "evidence": "raw/resource-final-validation.json" } ], "headline_results": [ { "scope": "historical r4 ordinary-summary, C1, about 7.4K input", "mtp_decode_p50_tokens_s": 67.3, "nospec_decode_p50_tokens_s": 25.4, "caveat": "ordinary output only" }, { "scope": "historical r4 ordinary-summary, C2, about 7.4K input", "mtp_decode_p50_tokens_s": 40.0, "nospec_decode_p50_tokens_s": 19.5, "caveat": "ordinary output only" } ], "failures": [ { "scope": "historical strict controlled-output diagnostic", "result": "4/4 failed; performance-ineligible", "evidence": "raw/strict-attempt2-adjudication.json" }, { "scope": "r5 .94 memory allocation", "result": "rejected under required GPU reserve; r4 .90 restored", "evidence": "raw/mtp3-r5-rejection.json" }, { "scope": "fresh SWE fixed campaign floor", "result": "Django10097 unresolved; the other four tasks resolved; no infrastructure failure", "evidence": "raw/swe-native/native-evidence.json" } ], "limitations": [ "The historical MTP comparison has a different activation from the fresh final r4 candidate.", "C16 is not cross-arm comparable: MTP effective active cap was 9 while no-spec was 16.", "C1 context timings overlapped coding work and are quality-only, not isolated performance evidence.", "C2 proves two overlapping application requests only to the extent captured by its retained engine samples; it does not establish hypothetical full-residency capacity.", "The original overlap stability artifact remains immutable failed coverage evidence; successors use a separately pinned inspector source and contender-stream-active/v1 contract.", "The initial independent stability review raised a source-digest-method finding; its retained correction accepts the exact run-cell package method without an additional replay.", "Fresh final capacity is a separate 10-cell, 72-request population and is not pooled with the historical 60-cell matched comparison.", "The selected 32K C2 operating point is a concurrency tradeoff: decode p50 13.43 tokens/s is below the 15 objective, while first-output p95 is 16.69 seconds under the 30-second bound. C1 is the latency-first option (63.44/62.46 decode p50 at 8K/32K).", "The completed SWE result is a fixed five-task campaign floor, not a broad coding benchmark population estimate.", "Video was not qualified.", "Current deployment, routing, aliases and client assignments are outside this public record." ], "decision": { "evidence_labels": [ "functional", "capacity", "quality" ], "decision_labels": [ "qualification-accepted", "deployment-private" ], "scope": "qualification only; no public deployment authorization or assignment claim" }, "evidence": [ "redaction-ledger.json", "raw/r4-comparison-decision.json", "raw/images.json", "raw/agentic-native-evidence.json", "raw/context-native/native-evidence.json", "raw/context-admission.json", "raw/swe-native/native-evidence.json", "raw/swe-native-validation.json", "raw/swe-native/official-grader-report-binding.json", "raw/c2-evidence-amendment.json", "raw/context-concurrent/result.json", "raw/context-concurrent/requests.jsonl", "raw/independent-digest-c2-review.json", "raw/stability-serial-unique.json", "raw/stability-overlap-unique.json", "raw/stability-overlap-unique-r2.json", "raw/stability-overlap-repeat-r2.json", "raw/stability-final-validation.json", "raw/stability-observer-amendment.json", "raw/independent-stability-observer-r3-review.json", "raw/independent-stability-final-review.json", "raw/independent-stability-source-binding-correction.json", "raw/stability-observer-full-tests-r2-receipt.json", "raw/capacity-summary.json", "raw/deep-health.json", "raw/resource-final-validation.json", "numerical-import-projection.json" ], "metrics_coverage": { "profile": "serving-comparison/v1", "status": "partial", "frozen_plan": "plans/run-plan-r4-summary-complete.json", "required_cells": [ "historical r4 ordinary-summary matrix" ], "cells": { "historical r4 ordinary-summary matrix": { "status": "descriptive-comparison", "reason": "Retained historical rows exist, but strict output control failed and C16 differs in effective capacity." } }, "missing_required_metrics": [] }, "comparison_evidence": null, "decision_labels": [ "qualification-accepted", "deployment-private" ], "comparison_scope": "The serving-comparison completeness flag remains false: strict controlled output failed and C16 was unmatched. Independent bounded qualification acceptance is separate." }