Request coverage and qualification limits¶
| Request | Retained evidence | Disposition / remaining limit |
|---|---|---|
| Review better recipes | Dated source registry, screening of 219 daily reports, pinned model/runtime sources and hardware-matched community leads. | No vendor-stable, fully qualified replacement established. Archive screening does not independently verify every reported claim. |
| Driver preparation | Driver 615.71.09 active on both Max-Q GPUs after user reboot; guarded repeatable provisioning delivered separately. | Observed kernel windows are bounded. This does not prove the historical crash cause was fixed. |
| Coding intelligence | Frozen five-repository scout: incumbent 3/5, Spark 2/5, Qwen medium 4/5 resolved; native results and official reports retained. | Small scout, not full SWE-bench or general coding superiority. All attempts count; submission and grading denominators remain separate. |
| Tool protocol | Oracle 7: incumbent 18/18; Spark 16/18 at both temperatures; Qwen medium 18/18; disabled 16/18. | Campaign requires 100%, separate from the native suite's 75% threshold. Earlier ambiguous oracle scores remain historical. |
| Long generation | Separate 32K and 64K natural-guide artifacts and independent static critiques. Qwen xhigh produced 23150 visible tokens; Spark 37029, both stop-terminated. | One guide per profile. Marker/syntax passes do not establish integrated program correctness. Incumbent reached 65536 total completion tokens with 11 complete chapters and chapter12 partial; visible token count is unknown. Independent static review found parser/scheduler defects and no visible repetition loop. |
| Context with output headroom | Single-request retrieval: incumbent/Spark 255795 input tokens; Qwen xhigh 255839 and 412589. Long-tool requests also passed. | Largest Qwen retrieval leaves 111699 tokens of window, but did not generate 65K output there. No full-window coding/concurrency qualification. |
| Eight images | All three loaded profiles passed the ordered eight-image corpus 4/4. | Qwen image evidence is default xhigh, not medium. Candidate retained-history/client continuation deferred after hard-gate failures. |
| Concurrency | Strict C4, unique nominal-32K prompts, exact128 words, 8192 completion cap: incumbent 0/16, Spark 0/16, Qwen xhigh 0/16. Each repair 0/4. | All are performance-ineligible; no finalist throughput, p99 or C16 capacity claim. Earlier incumbent short-summary scout was download-confounded. |
| Supported settings | Qwen CPU-embedding r1 contained OOM; GPU-embedding r2 loaded. Medium/disabled and Spark temperature0.6 repairs retained. | Both request repairs failed strict output. No matched Qwen no-MTP comparison; repeated Spark cold-start qualification deferred. |
| Stability | Native container/kernel receipts; no matching Xid in observed candidate windows. Exact OOM and recoverable allocator observations retained. | Bounded observations, not an uptime/soak guarantee. Shared output-control failures are not automatically GPU crashes. |
| Promotion | Conditional user authorization existed; no candidate passed the frozen hard gates. | No promotion. Retaining incumbent does not mean it passed all new tests or won every comparison. |
| Restoration | Exact incumbent recipe reloaded; both challengers unloaded; router metadata still advertises327680 context/65536 output. | Restored and admitting; authenticated smoke/JSON checks passed. GPU ownership matches; allocator occupancy differs. Initial full router fingerprint was not retained. |
| Delivery | Driver PR #68 and product PR #567 reviewed/merged; final product gate9559 passed/53 skipped; candidate registries and evidence publication prepared. | Final publication and private recipe delivery are separate from the completed live restoration. No candidate promotion occurred. No candidate client/catalog/dashboard deployment is claimed. |
The original campaign plan predates measurements. This coverage ledger records outcomes and limitations; it does not retroactively predeclare new acceptance criteria or convert deferred gates into passes.