2026-09-22 model shortlist validation evidence¶
This public-safe bundle records a bounded replacement screen and the initial
Qwen3.8 Flash Next TP2/C4 trial. It preserves native schemas after replacing
operator paths, host/container identifiers, network values, GPU UUIDs, and a
synthetic fixture secret with stable redaction tokens. All retained
container_id values, including nested status payloads, use
REDACTED_CONTAINER_ID. GPU-A-REDACTED and GPU-B-REDACTED preserve
distinct-card references, and public relative configuration names preserve the
four restoration hash-parity entries. No credential material is retained.
The two router-final-status.json snapshots retain their envelope shape but
replace active tier-to-model assignments and readiness state with
REDACTED_ACTIVE_ROUTER_STATE. The Qwen startup capture replaces its historical
loopback bind literal with <redacted-loopback-bind>; it does not assert a
different bind argument. Independent direct and routed probes retain the
bounded pass/fail outcomes used by the campaign decision.
The Qwen native agentic artifacts report 16/18 and 15/18 with their internal
0.75 pass-rate floor and passed: true. The campaign's predeclared
deterministic gate is 100%, so neither run advanced the candidate. The two
debug-loop failures were strict protocol failures. They are not evidence of
general coding inferiority or executed-code behavior.
The paired GLM scout also reports 16/18. Its two planning false negatives are a scorer defect: an earlier incidental substring is selected before the required next heading. The raw tie is retained but cannot rank broad planning/coding quality.
The Qwen six serial C1 direct preflight groups passed, startup used 65.55 GiB per rank, and the allocated shared KV pool reported 928,576 tokens. These facts do not qualify throughput, quality, vision, large context, or capacity: those stages were deferred after the strict agentic gate failed. Startup also retained a nonfatal custom-allreduce UUID-to-index parse warning and fallback; it did not prevent readiness or direct preflight.
The paired GLM agentic scout and final restoration are retained. Restoration returned the exact baseline identity/configuration, passed direct and routed smoke/JSON checks, and readmitted the router. This completes the bounded campaign but cannot authorize a promotion.
| Area | Files |
|---|---|
| Run plan and source screen | run plan, source registry |
| Qwen identity and startup | settings, startup, ready status |
| Direct functional checks | basic, tools |
| Agentic gate | greedy scout, native-sampling scout |
| Campaign controls | coverage and gaps, friction log, summary |
Final privacy review¶
Retained loopback URLs and startup listener ports use <redacted-port>;
worker PIDs use REDACTED_PROCESS_ID.
mode-restored.json preserves the recorded mode, TP size and unresolved list
but replaces the private blocked-workload catalog, exclusive owner and GPU
role/owner mappings with REDACTED_ACTIVE_TOPOLOGY. Public restoration
claims use the retained configuration hash comparisons and independent probes;
redacted status fields do not prove private topology or route assignments.