Skip to content

GLM-5.3-Flash DCP1 r11 qualification

The selected DCP1 batch-2,048/C4/0.97 r11 recipe is deployed after accepted direct, production-routed, and native-client checks for the pinned GLM-5.3-Flash EXL3 checkpoint on two RTX PRO 6000 Blackwell Max-Q cards. The public packet is publication-verified: the merged Pages publication returned HTTP 200 and a Chrome readback confirmed its title and core published facts. This qualification tests the single DCP2-to-DCP1 change after the retained DCP2 mixed-load fault; it does not establish a root cause, crash rate, or general hardware maximum.

Retained interim observations

  • The frozen DCP1 batch-2,048/C4 configuration reports 825,268 logical KV tokens. Three 201K unique-prefix and three repeated-prefix 8K-anchor rounds passed, as did one 310K/C2 overlap. The earlier 201K/4K third-anchor result exhausted its completion budget without a visible answer and remains a failure; the later 8K probes do not replace it.
  • The separate quality screen passed 12/12 and the agentic final gate passed 18/18. A natural-output diagnostic retained 18,016 visible tokens and its end marker, but it is not a SWE or application-correctness result.
  • The frozen five-task SWE scout completed official grading at 4/5 resolved, meeting its absolute at-least-3/5 gate. Its CPython/package environment differs from the retained 3/5 baseline, so the pair cannot support a causal DCP1 or score-improvement claim. The earlier composite gate remains a retained diagnostic failure because it correctly records that mismatch.
  • The separate 64K diagnostic reached its declared completion ceiling with a native length failure. Independent review found the retained visible text coherent through truncation, which satisfies the diagnostic's predeclared continuation rule without waiving the failure. Its one-attempt visible yield is materially below the matched incumbent (24,123 versus 70,772 characters; three versus eleven complete chapters); the cause is unknown.
  • Isolated routed preflight passed 28 observations across eight functional families, including 20 tool-batch observations. In a separate offered-eight nominal-32K diagnostic, six served responses met the identity, own-canary, and strict-output contract while two requests received bounded HTTP 503 admission timeouts. The native exit-1 result remains performance-ineligible; the corrected, versioned gate joins all eight gateway IDs to terminal rows with the same configuration and closes only the configured C4 dispatch/admission boundary. It makes no C8 or sustained-queue claim.
  • Independent review recommended the exact DCP1 batch-2,048/C4/0.97 recipe under the retained authorization, and r11 was subsequently deployed after direct and production-routed preflights each passed 28/28 checks and 13 native-client semantic paths passed. Several retained client responses failed strict format (fenced JSON or Mac prose), so this is not a strict-format fleet pass. Its 825,268 KV tokens are 41.73% below retained DCP2's 1,416,244, so four full 327K windows are unsupported. This remains a mitigation, not a root-cause fix, crash-rate finding, or speed claim.
  • Dashboard publication is verified: nine source-matched runs, 99 metrics, and nine authenticated cards imported with byte-identical repeats. The merged Pages publication returned HTTP 200, and Chrome confirmed the title, DCP1, 825,268 KV, 18,016 visible-token, SWE-caveat, and evidence-index content.
  • The batch-4,096 cells were strict-correct but did not produce a comparable estimator; batch 8,192 started below the campaign's conservative C2 reserve and received no dependent requests. Batch 2,048 remains selected for this campaign.
  • The offered-concurrency-8 diagnostic passed seeds 801 and 1801 at 16/16, then seed 2801 failed 1/16 strict/canary. The engine stayed healthy, but the entire dependent C8 comparison stopped. Offered C8 is outside the declared production C4 admission boundary and does not qualify C8.

Decision boundary

The exact r11 recipe is selected and deployed; accepted native-client, restoration, and publication-readback evidence are retained in the packet. The exact r10 recipe remains the rollback. The public packet was merged, its Pages run succeeded, and the deployed page returned HTTP 200 with Chrome readback. Historical pre-deployment reviews and gates that record promoted=false remain unaltered as dated evidence.

The selected public-native artifacts, redaction receipts, and explicit gaps are in the evidence index.