Campaign coverage and limits¶
Scope: an authorized optimization scout ending2026-10-21T15:30Z; no promotion.
| User outcome | Evidence condition | Current status |
|---|---|---|
| Learn source/options | Pinned code paths and independent review | source-findings.md and optimization-plan.md |
| Try supported improvements | Versioned recipe, parent/delta, preflight and retained native results | Capacity16, fixed1/2/3, adaptive3, throughput policy, host24/32GiB and prefill2048 measured; native artifact tables are authoritative |
| Compare existing model | Same workloads and sampling, distinct checkpoints identified | Fresh incumbent96/96; base95/96; host32 fixed2/adaptive3 and p2048 natural96/96 each. Natural-output screening only |
| Test practical behavior | Independent protocol, agentic and actual repository grading | Protocol and deterministic agentic/context results retained with failures. SWE blocked before submission by unenforced deadline and detached-container custody; attempted0/graded0/resolvednull |
| Mixed prefill, cache and cancellation | Bounded independent fixtures and source-aware metrics | Upstream scheduler42 requests plus cancellation/replacement passed. Host-cache restoration and concurrent branches observed; fixture punctuation/counter failures retained. Prefill2048 delivery gaps lower in one paired diagnostic; counting failures remain |
| Reasoning policy | Same runtime and exact explicit request control | Template default Max and exact Low source-reviewed. Low preflight6/7 (needle refusal), exact replay3/3 retained. High7/7, natural96/96, agentic18/18, C1context18/18, strict16n100100/100; Max strict16n100100/100. High shorter reasoning is a policy tradeoff |
| Long-context concurrency | Actual token counts, exact independent answers, sampled engine occupancy | C1 context measured with retained failures. Both Max/High independently reviewed C4 fixtures passed4/4 at317257..317277actual tokens, all7gates, active max4, clean drain. One batch each, not sustained capacity or simultaneous GPU proof |
| Easy later conversion | Exact image/checkpoint/recipe/package plus rollback | recipes and prior exact retained-incumbent transaction preserved |
| Safe return by08:30 | Exact incumbent, direct+routed acceptance, readmission | verified at15:07Z; exact retained incumbent, direct7/7 and routed7/7, primary ready/admitting, candidate absent,14/14 manifest hashes match |
Measurement gaps¶
This expanded scout was started with the retained initial trial and optimization plan. The common artifact templates were added at11:32Z after initial requests. Do not claim a pre-request serving-comparison/v1 envelope or reviewed continuous collector for those runs. Per-cell engine/GPU snapshots are bounded observations, not continuous occupancy proof; host/cgroup observations bracket lifecycle/preparation. A full serving comparison remains incomplete. Natural response lengths differ and no API sampling seed is set by the capacity CLI. Cold/warm labels identify attempted treatment: use measured reuse counters. One base canary failure remains performance-ineligible despite three successful diagnostic replays.
No usable no-speculation performance control exists: MTP0 first-sample error1 caused one runtime abort, although an exact retry passed the protocol gates. The error combines four sampler predicates; first-fault attribution needs narrower instrumentation. Fixed3 and p2048 agentic each have a17/18 run; fixed2-host32 C1 context is15/18. Strict128 output is0/12 in host24 and p2048 Max/High diagnostics. These failures are not erased by separate passing16-word or natural-output populations. Flare's current GLM runtime is text-only; the checkpoint itself includes vision weights. No real-client migration or promotion occurred.
Final matched incumbent High: natural96/96, strict16n100100/100, deterministic agentic15/18. Default incumbent C1context18/18. The High speed policy has a quality caveat; no global policy change.