{ "schema": "pp4-round6-budget-v1", "classification": "ESTIMATED planning from MEASURED historical inputs; not complete PP4 blocks or a guarantee", "dep1_input_sha256": "9e5432d13f0a86f9caaf7d329e190f1cdf557284236bfdd37ef1439228311309", "cells": { "short-b1": { "observations": [ { "max_decode_span_s": 26.50626343896147, "max_ttft_s": 1.7645618190290406, "measured_output_tokens": 1500, "sha256": "deac0950c403b7415ce7e61310d718d4b8b9fb149c3cefa6d24f1678575cd25c", "source": "NBW-W-20261026T231445Z/s1-K-short-b1.json" }, { "max_decode_span_s": 28.47160706698196, "max_ttft_s": 2.331858152989298, "measured_output_tokens": 1500, "sha256": "daa891adf4cf907a5cd3418b36b01aaf97cfd0fcb801994785b59061eef1ee3d", "source": "NBW-W-20261026T231445Z/s3-K-short-b1.json" }, { "max_decode_span_s": 27.960649574000854, "max_ttft_s": 2.4093480160227045, "measured_output_tokens": 1500, "sha256": "3ecc7c888b506b8c618902b7745c32b0675b49e4e221cbbf6ca2b76dc0afbc84", "source": "NBW-W-20261026T231445Z/s5-K-short-b1.json" } ], "historical_batch": 1, "batch_scaling": 1, "max_ttft_s": 2.4093480160227045, "max_decode_span_s": 28.47160706698196, "scaled_512_estimate_s": 19.115146022178664, "allowance_per_group_s": 20, "groups_per_start": 4, "profile_groups_per_start": 0, "classification": "ESTIMATED from MEASURED B1" }, "32k-b1": { "observations": [ { "max_decode_span_s": 29.4779839469702, "max_ttft_s": 14.812975562002975, "measured_output_tokens": 1500, "sha256": "c5331d3465fb12b15d5ba4690fba9da45769236787bdb53cd0650c959861234d", "source": "NBW-W-20261026T231445Z/s1-K-32k-b1.json" }, { "max_decode_span_s": 29.47603059199173, "max_ttft_s": 16.56382137304172, "measured_output_tokens": 1500, "sha256": "979e7337ba5b75f4888dc0bfd403a2cb301a537f3990cf9d03f55431bd9bca82", "source": "NBW-W-20261026T231445Z/s3-K-32k-b1.json" }, { "max_decode_span_s": 27.4701354339486, "max_ttft_s": 16.617130869999528, "measured_output_tokens": 1500, "sha256": "0aaff802c13c57a5dc551b30b7be7b0287f060894110e92d0397dac26da8d24f", "source": "NBW-W-20261026T231445Z/s5-K-32k-b1.json" } ], "historical_batch": 1, "batch_scaling": 1, "max_ttft_s": 16.617130869999528, "max_decode_span_s": 29.4779839469702, "scaled_512_estimate_s": 33.66599664511745, "allowance_per_group_s": 35, "groups_per_start": 4, "profile_groups_per_start": 0, "classification": "ESTIMATED from MEASURED B1" }, "32k-b4": { "observations": [ { "max_decode_span_s": 58.99707188201137, "max_ttft_s": 51.37888995197136, "measured_output_tokens": 1500, "sha256": "87c9c91c212076d1feabb9bf3ab66002eefa3ae7726d46496827203704e5b767", "source": "NBW-W-20261026T231445Z/s1-K-32k-b3.json" }, { "max_decode_span_s": 58.419027732044924, "max_ttft_s": 51.72109578800155, "measured_output_tokens": 1500, "sha256": "a5879c7ac17144b1e1081683007375af797ffd231f9a42ce9d9f713288314fa2", "source": "NBW-W-20261026T231445Z/s3-K-32k-b3.json" }, { "max_decode_span_s": 61.36587655596668, "max_ttft_s": 51.6934184510028, "measured_output_tokens": 1500, "sha256": "1de0febd3966a4a7b062b3f501676c22ac92e9684e1487a7133417a236db9bc5", "source": "NBW-W-20261026T231445Z/s5-K-32k-b3.json" } ], "historical_batch": 3, "batch_scaling": 1.3333333333333333, "max_ttft_s": 68.96146105066873, "max_decode_span_s": 81.8211687412889, "scaled_512_estimate_s": 103.85380076167516, "allowance_per_group_s": 105, "groups_per_start": 2, "profile_groups_per_start": 0, "classification": "ESTIMATED B3-to-B4 extrapolation, not measured B4" } }, "startup_observations_s": [ 388, 404 ], "startup_sources": [ { "locator": "Budget: Spark X start to ready 388 s", "path": "S49.report.md", "sha256": "9564a6621221c6d594835395a6853aad1ac7f575806cf71fea9abaf128475636" }, { "locator": "2026-10-26 20:45 COORD: W8 start 404 s incl warm-up", "path": "LANES.md", "sha256": "d11f865bd2dc118478d828b19fe3979b667c448a342fc6d910a13d88426460ab" } ], "worker_startup_s": 333, "worker_source": { "locator": "Budget: Spark X start to ready 388 s", "path": "S49.report.md", "sha256": "9564a6621221c6d594835395a6853aad1ac7f575806cf71fea9abaf128475636" }, "setup_s": 120, "cuts_s": 120, "warm_per_start_s": 30, "greedy_on_s": 42, "transition_per_start_s": 60, "restore_s": 720, "pair_bound_s": 600, "formula": "preserve max cold TTFT; scale max decode span by511/1499; add2s publication+5s overhead; ceil to5s; B4 extrapolates B3 TTFT and decode by4/3", "profiles": "Round6: no measurement companions or full-round-p95 guard; complete-round instrumentation stays in admission", "limits": "historical333s worker startup is separate from388\u2013404s Spark startup; no overlap credited; no sampled speed gain assumed", "pair_reserve": "Measurement requires hash-bound QUALIFIED admission physical-pair step; ceil(max(180s,1.5*MEASURED driver seconds)); admission/pair-only retain600s bound", "fallback": "Try all3cells first; if>85min drop32k-b4 only; if two B1cells still>85min refuse; no other cuts" }