RTX PRO 6000 mention audit

This audit classifies every tracked Markdown file containing an RTX PRO 6000 name at the 2026-09-13 UTC review. Evidence-directory Markdown is classified with its own file, not inherited implicitly from the parent finding. This is a coverage control; benchmark results remain in the run catalog.

measured-on

docs/findings/2026-11-01-swift-tp1-gpu0.md; docs/benchmarks/models/qwen38-efficient-variants.md (one RTX PRO 6000 Blackwell Max-Q: final TP1 functional, quality, capacity and resource evidence; qualification accepted; deployment status private; no stable p99 or sixteen-active-request claim).

docs/findings/2026-10-26-ignis-080-preflight.md; docs/findings/2026-10-26-ignis-080-preflight-evidence/README.md (one-Max-Q bounded functional and resource measurements; required-tool preflight and repeat passed, original auto-tool miss retained; exact operational incumbent restoration verified; no performance, capacity, or promotion claim).

docs/findings/2026-10-26-marker-gpu-trace.md (one-Max-Q actual-CUDA startup gates and one dual-Max-Q C1 model request; first close observed budget-forced, second close not reproduced, complete worker/API reconciliation unknown; no performance or promotion claim).

docs/findings/2026-10-25-sampling-upstream-tolerance.md; docs/findings/2026-10-25-sampling-upstream-tolerance-evidence/README.md (one-Max-Q synthetic kernel gate passed; one dual-Max-Q C1 model request failed repeated-close and strict-output gates; exact operational restoration passed, matched JSON smokes failed before and after; no speed claim or promotion).

docs/findings/2026-10-25-sampling-reference.md; docs/findings/2026-10-25-sampling-reference-evidence/README.md (one-Max-Q synthetic sampling-reference checks before weights; 8/8 forced cases passed, one of 272 ordinary rows failed the strict local no-worse gate, no candidate inference or promotion; exact operational restoration passed, but the restored-incumbent strict JSON smoke failed).

docs/findings/2026-10-25-sampling-fix.md; docs/findings/2026-10-25-sampling-fix-evidence/README.md (one-Max-Q synthetic sampling-kernel checks before weights; two GPUs reserved for the unrun TP2 candidate, failed ordinary parity retained, no promotion).

docs/findings/2026-10-25-glm-diagnostics.md (six bounded Go/Rust quality measurements on the dual-Max-Q incumbent; failed and unstarted cases retained, no engine comparison or promotion).

docs/findings/2026-10-24-evaluation-platform-ignis.md (bounded incumbent diagnostics on the dual-Max-Q configuration and one-GPU Ignis candidate startup evidence; no candidate inference, paired comparison, or promotion).

docs/benchmarks/engine-testing.md (navigation and interpretation of the dated Flare/Ignis measurements below; no additional hardware test, current deployment claim or promotion);

docs/findings/2026-10-23-flare-upstream-refresh.md; docs/findings/2026-10-23-flare-upstream-refresh-evidence/reconstruction.md; docs/findings/2026-10-23-flare-upstream-refresh-evidence/report.md; docs/findings/2026-10-23-flare-upstream-refresh-evidence/compatibility-pr-draft.md; docs/findings/2026-10-23-flare-upstream-refresh-evidence/issue-draft.md (current-upstream compatibility baseline on dual Max-Q TP2; descriptive replay and warm controls, driver boundary, no 75% arm or promotion).

docs/findings/2026-10-23-flare-state-diagnostics.md (dual-Max-Q TP2 canonical state, protected latency, and separate scheduler-trace diagnostics; bounded state checks accepted, numerical and latency gates remain held; no promotion).

docs/findings/2026-10-22-flare-cache-fix.md (dual-Max-Q TP2 cache-row allocation diagnostic; fixed-work replay support remains diagnostic, with numerical failure, protected regressions, and decode-C4 mismatches retained; no promotion).

docs/findings/2026-10-22-flare-cache-diagnostics.md (dual-Max-Q cache replay and allocator-counter diagnostics; no performance ranking or promotion).

docs/findings/2026-10-21-flare-inference-diagnostics.md (dual-Max-Q four-arm diagnostic wave, bounded scheduler trace and retained incumbent C4 failures; no overall ranking or promotion).

docs/findings/2026-10-21-flare-glm53-optimization.md (bounded dual-Max-Q optimization and incumbent controls; no promotion). docs/findings/2026-10-21-flare-glm53-optimization-evidence/diagnostics/optimization-plan.md (plan for the same measured hardware; unrun options remain plans).

docs/findings/2026-10-21-flare-glm53-trial.md; docs/findings/2026-10-21-flare-glm53-trial-evidence/README.md (bounded dual-Max-Q Flare/incumbent natural-output scout, strict failures, warm package reuse and exact restoration; no promotion or engine-only comparison).

docs/findings/2026-10-16-glm-natural-baseline.md (dual-Max-Q natural-output timing and repeated cached SWE observations through a shared router; coding repeatability not established, no promotion);

docs/findings/2026-10-16-sglang-qwen-comparison.md and docs/findings/2026-10-16-sglang-qwen-comparison-evidence/README.md (bounded dual-GPU SGLang, Qwen, and Nemotron scout evidence; no performance comparison or promotion); docs/benchmarks/models/nemotron35-super-vl-120b.md (measured-on: bounded direct TP2 scout, no qualification);

docs/findings/2026-10-15-replacement-evaluation/README.md (evidence index for measured bounded dual-GPU replacement gates and one-GPU numerical compatibility; no additional measurement); docs/findings/2026-10-15-replacement-evaluation.md (GLM, MiMo, and Mistral dual-GPU gates; DeepSeek Stage 2 separately classified as one-GPU numerical evidence; exact GLM restoration retained); docs/findings/2026-10-16-glm-v091-trial.md (measured bounded R1/R2 and restored-original GLM functional, Oracle, vision, context, reserve, and five-case SWE evidence on the dual Max-Q configuration; R2 and control both resolved 3/5, so no promotion or performance claim); docs/benchmarks/models/mistral-medium-35.md (measured-on: bounded direct TP2 qualification); docs/benchmarks/models/deepseek-v41-flash.md (measured-on: Stage 2 one-GPU numerical evidence; Stage 3 dummy-path, Stage 4 bounded real-weight, and full-role r2/r3 resource-gate diagnostics plus r4's first-round Oracle8 failure on the dual-GPU configuration); docs/findings/2026-10-13-full-qualification.md; docs/findings/2026-10-13-full-qualification-evidence/configuration-and-identity.md (TP1/TP2 protocol, quality, fixed-five SWE and baseline capacity on dual Max-Q hardware; incomplete and failed candidate runs are retained, with no promotion).

docs/findings/2026-10-06-mimo-qualification.md (dual-Max-Q MiMo and abliterated GLM direct quality scout; both candidates not qualified, no performance or full-context claim, exact r11 restoration verified: direct 25/25, routed 7/7, unchanged configuration hashes, and router readmission);

docs/findings/2026-10-05-xeon-model-qualification.md (dual-Max-Q Qwen replacement and GLM CSF/QAD scout; configured context/C4 only, strict C4 cells are performance-ineligible; fresh-reload A1 missed the SWE floor, B1/B2/A2 did not run, and exact r11 restoration completed);

docs/findings/2026-09-23-glm-runtime-stability.md (dual-Max-Q runtime crash reproduction and bounded DCP1 mitigation; exact baseline restored, no performance ranking or promotion);

docs/findings/2026-09-23-glm-dcp1-qualification.md (dual-Max-Q DCP1 r11 functional, capacity, and quality qualification; selected and deployed with retained direct/routed and native-client acceptance, while public-post merge and deployed readback remain pending);

docs/findings/2026-09-18-glm53-vision.md; docs/findings/2026-09-18-glm53-vision-evidence/README.md; docs/findings/2026-09-18-glm53-vision-evidence/publication-summary.md (bounded dual-Max-Q vision acceptance; negative long-context attempt retained);

docs/findings/2026-09-19-glm53-apc.md; docs/findings/2026-09-19-glm53-apc-promotion.md; docs/findings/2026-09-19-glm53-apc-evidence/README.md; docs/findings/2026-09-19-glm53-apc-evidence/publication-summary.md; docs/findings/2026-09-19-glm53-apc-evidence/run-plan.md (matched r9/r10 APC functional and capacity evidence on the dual Max-Q configuration; historical restoration verified, then explicit human approval and fresh bounded acceptance);

docs/findings/2026-09-17-glm53-mixed35-startup.md (failed managed startup on the dual Max-Q host; no inference or GPU-capacity measurement);

docs/findings/2026-09-17-glm53-mixed35-fix-forward.md (bounded mixed-3.5-bpw loader recovery and C1/C4 functional evidence on the dual Max-Q host; no performance or promotion claim);

docs/findings/2026-09-21-mimo-v26-qualification.md; docs/findings/2026-09-21-mimo-v26-qualification-evidence/README.md; docs/findings/2026-09-21-mimo-v26-qualification-evidence/publication-summary.md; docs/benchmarks/models/mimo-v26-flash.md (dual-Max-Q baseline and candidate startup/correctness evidence; MiMo has no eligible capacity or performance measurement);

docs/findings/2026-09-22-model-shortlist-validation.md; docs/findings/2026-09-22-model-shortlist-validation-evidence/publication-summary.md (dual-Max-Q Qwen TP2 with C4 configured admission; all retained direct and agentic checks are serial C1, and fixture/scorer defects block agentic planning comparison);

docs/findings/2026-09-23-glm-spark-qwen-next-qualification.md; docs/findings/2026-09-23-glm-spark-qwen-next-qualification-evidence/README.md (dual-Max-Q frozen qualification and restored incumbent control; candidate and incumbent strict-C4 failures are retained as performance-ineligible, with no promotion);

docs/findings/2026-09-26-single-gpu-qualification.md (one-card challenger qualification; all final candidates passed deterministic image, agentic, and context gates but failed the frozen coding floor; exact dual-card GLM r11 restored and readmitted);

docs/findings/2026-09-13-intelligence-context-scout.md (selected GLM text-primary qualification; post-promotion context and client acceptance retained; fresh boot/reboot tests unrun);

docs/findings/2026-09-11-glm53-v043-native-linux-qualification.md; docs/findings/2026-09-11-glm53-v043-native-linux-qualification-evidence/README.md; docs/findings/2026-09-11-glm53-v043-native-linux-qualification-evidence/publication-summary.md;

anvil_serving/observability/dashboard/static/documents/benchmarks.md (packaged Workbench copy of the benchmark methodology; no additional measurement);

docs/BENCHMARKS.md; docs/benchmarks/comparison.md; docs/benchmarks/configurations.md; docs/benchmarks/recipe-results.md; docs/benchmarks/gpt-oss-puzzle-88b-recipe.md; docs/benchmarks/hardware/rtx-pro-6000.md; docs/benchmarks/index.md; docs/benchmarks/methodology.md; docs/benchmarks/models.md; docs/benchmarks/models/agents-a1.md; docs/benchmarks/models/deepseek-v4-flash.md; docs/benchmarks/models/gemma-4.md; docs/benchmarks/models/gpt-oss-120b.md; docs/benchmarks/models/gpt-oss-puzzle-88b.md; docs/benchmarks/models/glm53-flash.md; docs/benchmarks/models/inkling-small.md; docs/benchmarks/models/index.md; docs/benchmarks/models/laguna-s-2.1.md; docs/benchmarks/models/minimax-m27-reap.md; docs/benchmarks/models/mistral-small-4.md; docs/benchmarks/models/nemotron-puzzle-75b.md; docs/benchmarks/models/nemotron3-super-120b.md; docs/benchmarks/models/ornith-35b.md; docs/benchmarks/models/qwen35-122b.md; docs/benchmarks/models/qwen36-27b.md; docs/benchmarks/models/qwen38-27b.md; docs/benchmarks/models/qwen38-flash-next.md; docs/benchmarks/runs.md; docs/findings/2026-10-26-flash-next-engine-qualification.md; docs/findings/2026-10-26-flash-next-engine-qualification-evidence/publication-summary.md (three serialized recipe scouts on one RTX PRO 6000 Blackwell Max-Q; functional and context results retained, mixed coding outcomes and two failed C1 output gates; no engine ranking, speed claim, or promotion). docs/findings/2026-07-10-blackwell-local-model-bakeoff.md; docs/findings/2026-07-10-blackwell-local-model-bakeoff-evidence/failures.md; docs/findings/2026-07-10-blackwell-local-model-bakeoff-evidence/preflight-transcripts.md; docs/findings/2026-07-10-blackwell-local-model-bakeoff-evidence/reproduction.md; docs/findings/2026-07-10-qwen35-122b-heavy-candidate.md; docs/findings/2026-07-12-gpt-oss-120b-deterministic-recheck.md; docs/findings/2026-07-12-heavy-intelligence-challengers.md; docs/findings/2026-07-12-nemotron-puzzle-recheck.md; docs/findings/2026-07-12-qwen35-122b-mxfp4-benchmark.md; docs/findings/2026-07-12-qwen36-27b-eval-baseline.md; docs/findings/2026-07-12-qwen36-27b-heavy-variation-bakeoff.md; docs/findings/2026-07-12-qwen36-protocol-v2-comparison.md; docs/findings/2026-07-12-rtx-pro-6000-heavy-eval-v2.md; docs/findings/2026-07-12-rtx-pro-6000-heavy-eval-v2-evidence/laguna-reproduction.md; docs/findings/2026-07-12-thinkingcap-heavy-promotion.md; docs/findings/2026-07-13-q36-pro6000-container-recipe.md; docs/findings/2026-07-13-q36-pro6000-container-recipe-evidence/reproduction.md; docs/findings/2026-07-16-gemma4-chat-template-bakeoff.md; docs/findings/2026-07-16-gemma4-unsloth-nvfp4-follow-up.md; docs/findings/2026-07-16-gemma4-vllm0251-wsl2-c128.md; docs/findings/2026-07-17-gemma4-31b-optimization.md; docs/findings/2026-07-17-gpt-oss-puzzle-qualification.md; docs/findings/2026-07-18-gpt-oss-puzzle-heavy-promotion.md; docs/findings/2026-07-26-laguna-s-heavy-qualification.md; docs/findings/2026-07-27-anvil-serving-release-readiness-sweep.md; docs/findings/2026-07-28-agents-a1-multimodal-qualification.md; docs/findings/2026-07-28-qwen35-122b-primary-qualification.md; docs/findings/2026-07-29-agents-a1-primary-promotion.md; docs/findings/2026-07-29-agents-a1-qwen-262k-head-to-head.md; docs/findings/2026-07-29-agents-a1-qwen-262k-head-to-head-evidence/README.md; docs/findings/2026-08-01-deepseek-v4-flash-0731-r16-dspark-qualification.md; docs/findings/2026-08-01-deepseek-v4-flash-0731-research-update.md; docs/findings/2026-08-01-dual-pro-tp2-model-campaign.md; docs/findings/2026-08-02-deepseek-v4-flash-0731-650k-1m-pi-qualification.md; docs/findings/2026-08-02-deepseek-v4-flash-0731-primary-promotion.md; docs/findings/2026-08-02-deepseek-v4-flash-0731-native-kv-offload-256k.md; docs/findings/2026-08-03-deepseek-context-agentic-swe-smoke.md; docs/findings/2026-08-07-deepseek-0731-vision-nvfp4-recipe-intake.md; docs/findings/2026-08-07-deepseek-0731-vision-nvfp4-sglang-first-load.md; docs/findings/2026-08-10-deepseek-v4-flash-0731-r33-quality-control.md; docs/findings/2026-08-10-deepseek-v4-flash-0731-r33-batch-token-ab.md; docs/findings/2026-08-11-deepseek-v4-flash-0731-r33-393k-promotion.md; docs/findings/2026-08-14-qwen38-27b-tp-mtp-context-matrix.md; docs/findings/2026-08-14-qwen38-27b-1m-context.md; docs/findings/2026-08-14-qwen38-27b-official-qualification.md; docs/findings/2026-08-14-qwen38-27b-split-promotion.md; docs/findings/2026-08-15-qwen38-27b-agentic-swe-scout.md; docs/findings/2026-08-15-qwen38-27b-mtp-depth-qualification.md; docs/findings/2026-08-15-qwen38-27b-sglang-consolidation-ab.md; docs/findings/2026-08-15-qwen38-27b-sglang-fp8-single-promotion.md; docs/findings/2026-08-15-qwen38-27b-sglang-mtp-multimodal-qualification.md; docs/findings/2026-08-15-qwen38-27b-sglang-nvfp4-qualification.md; docs/findings/2026-08-16-deepseek-v4-flash-0731-infernal-r15-393k-promotion.md; docs/findings/2026-08-16-qwen38-27b-video-router.md; docs/findings/2026-08-21-deepseek-v4-flash-0731-infernal-r18-1m-promotion.md; docs/findings/2026-08-26-qwen38-flash-next-promotion.md; docs/findings/2026-08-26-qwen38-flash-next-qsa-fast-mtp3-promotion.md; docs/findings/2026-08-26-qwen38-flash-next-vision-promotion.md; docs/findings/2026-08-26-qwen38-flash-next-vision-promotion-evidence/publication-summary.md; docs/findings/2026-08-29-glm53-cardillo-purtell-qualification.md; docs/findings/2026-08-29-glm53-cardillo-adaptive-mtp-evidence/publication-summary.md; docs/findings/2026-08-29-glm53-k3-dflash2-optimization-evidence/README.md; docs/findings/2026-08-29-glm53-k3-dflash2-optimization-evidence/publication-summary.md; docs/findings/2026-08-30-glm53-k3-dflash2-1m-optimization.md; docs/findings/2026-08-31-glm53-xgrammar-524k-qualification.md; docs/findings/2026-08-31-glm53-xgrammar-524k-qualification-evidence/README.md; docs/findings/2026-08-31-glm53-xgrammar-524k-qualification-evidence/publication-summary.md; docs/findings/2026-09-02-glm53-sglang-sm120-393k-promotion.md; docs/findings/2026-09-02-glm53-sglang-sm120-swe-smoke.md; docs/findings/2026-09-02-glm53-sglang-sm120-qualification.md; docs/findings/2026-09-02-glm53-sglang-sm120-qualification-evidence/README.md; docs/findings/2026-09-02-glm53-sglang-sm120-qualification-evidence/promotion-publication-summary.md; docs/findings/2026-09-02-glm53-sglang-sm120-qualification-evidence/publication-summary.md; docs/findings/2026-09-03-glm53-concurrency-capacity-interpretation.md; docs/findings/2026-09-04-qwen38-27b-pro6000-possibility-plan.md; docs/findings/2026-09-04-qwen38-27b-pro6000-possibility-evidence/README.md; docs/findings/2026-09-04-qwen38-27b-pro6000-possibility-evidence/friction-log.md; docs/findings/2026-09-04-qwen38-27b-pro6000-possibility-evidence/publication-summary.md; docs/findings/2026-09-08-glm53-linux-wsl-comparison.md; docs/findings/2026-09-08-glm53-linux-wsl-comparison-evidence/README.md; docs/findings/2026-09-08-glm53-linux-wsl-comparison-evidence/publication-summary.md; docs/findings/2026-09-08-glm53-linux-wsl-comparison-evidence/reproduction.md; docs/findings/2026-09-09-glm53-native-nccl-p2p.md; docs/findings/2026-09-09-glm53-native-nccl-p2p-evidence/README.md; docs/findings/2026-09-09-glm53-native-nccl-p2p-evidence/publication-dispatch.md; docs/findings/2026-09-09-glm53-native-nccl-p2p-evidence/publication-summary.md; docs/findings/blackwell-sm120-lab-notebook.md; .tickets/closed/2026-07-27-release-sweep-fixes.md; .tickets/closed/2026-07-28-agents-a1-multimodal-qualification.md; .tickets/closed/2026-07-29-agents-a1-qwen-262k-head-to-head.md; .tickets/closed/2026-08-01-deepseek-0731-r16-b12x-performance-spike.md; .tickets/2026-08-01-deepseek-dspark-128k-kv-reserve.md; .tickets/2026-08-01-deepseek-dspark-sm120-w8a8-tune.md; .tickets/closed/2026-08-01-sglang-inkling-modelopt-missing-accelerate.md; .tickets/closed/2026-08-01-sglang-inkling-sm120-grouped-gemm-shared-memory.md; .tickets/closed/2026-08-02-deepseek-r16-native-kv-offload-illegal-access.md; examples/primary-node/q36/README.md.

docs/findings/2026-09-26-swift-tp2-jev-evidence/README.md (Swift TP2 bounded observations; SWE 2/5 below 4/5 floor; no qualification or promotion; exact GLM incumbent restored, no performance claim);

external-prior

docs/findings/2026-08-10-deepseek-v4-flash-0731-community-config-refresh.md (community recipe research and proposed test arms; no local measurement); docs/findings/2026-08-15-qwen38-27b-external-recipe-refresh.md (current external recipe research and proposed official-weight test arms; no local measurement).

protected/co-resident

docs/findings/2026-07-07-voice-latency-model-shortlist.md; docs/findings/2026-07-08-stt-model-benchmark.md; docs/benchmarks/hardware/rtx-5090.md; docs/benchmarks/models/nemotron-omni-30b.md; docs/benchmarks/models/nemotron35-asr.md; docs/benchmarks/models/parakeet.md; docs/benchmarks/models/qwen25-omni-3b.md; docs/benchmarks/models/qwen3-asr.md; docs/findings/2026-07-13-t011-ocr-rebalance.md; docs/findings/2026-07-13-t015-resident-set.md; docs/findings/2026-07-28-nemotron35-asr-qualification.md; docs/findings/2026-07-voice-stt-ab.md; docs/findings/2026-07-voice-tts-ab.md.

topology-only

anvil_serving/observability/dashboard/static/documents/serving.md (packaged Workbench copy of the product README);

README.md; START_HERE.md; CLAUDE.md; docs/ARCHITECTURE.md; docs/COMFYUI-MIGRATION-RUNBOOK.md; docs/CONFIGURATION.md; docs/EXTERNAL-BENCHMARKS.md; docs/THIN-CAPABILITY-GATEWAY.md; docs/TROUBLESHOOTING.md; docs/adr/0010-specialized-engine-tier.md; docs/adr/0011-two-mode-operation.md; docs/adr/0017-gpu-residency-reservations.md; docs/adr/0018-router-transition-safety.md; docs/adr/0034-fleet-control-plane-and-node-runtime-classes.md; docs/benchmarks/rtx-pro-6000-audit.md; docs/benchmarks/models/voice-llm-mlx.md; docs/findings/2026-07-12-green-context-mps-capability.md; docs/findings/2026-07-18-lifecycle-aware-wsl-cache-reclaim.md; docs/findings/2026-08-01-dual-pro-tp2-campaign-evidence/compatibility-brief.md; docs/findings/2026-08-17-qwen38-27b-radixark-nvfp4-rtx5090-128k.md; docs/findings/2026-08-17-qwen38-27b-radixark-nvfp4-rtx5090.md; docs/findings/2026-08-21-qwen38-27b-gguf-250k-rtx5090.md; docs/findings/README.md; docs/index.md; examples/primary-node/README.md; skills/anvil-serving-benchmark-docs/SKILL.md.

unrelated

CHANGELOG.md (release-history mention); .tickets/closed/2026-08-01-external-benchmark-path-quantization-contamination.md (test-adapter identity bug, not a local hardware measurement); tests/fixtures/external_benchmarks/rtx6kpro_summary.md (test fixture, not local qualification evidence).

2026-09-11 output-budget additions

| Path | Classification | Basis | | docs/findings/2026-09-26-swift-tp2-jev.md | measured, not-qualified | Swift TP2 bounded evidence: fixed-denominator SWE 2/5 below the 4/5 floor; no promotion; exact GLM incumbent restored. | |---|---|---| | docs/findings/2026-09-11-glm53-output-budget.md | measured | Real Pi requests on the native Max-Q pair; bounded functional evidence. | | docs/findings/2026-09-11-glm53-output-budget-evidence/publication-summary.md | measured | Derivative of the same local evidence; no new hardware claim. |

2026-09-24 RTX 5090 target correction

Path	Classification	Basis
docs/findings/2026-09-23-thinkingcap-5090.md	unrelated	ThinkingCap RTX 5090 is the corrected user-selected target. Earlier RTX PRO 6000 preparation was a wrong-target setup and produced no ThinkingCap inference or benchmark measurement.
docs/benchmarks/models/thinkingcap-qwen38-27b.md	unrelated	Dossier records only the one-RTX-5090 local candidate lane; it makes no RTX PRO 6000 measurement claim.
docs/findings/2026-09-23-thinkingcap-evidence/README.md	wrong-target preparation	Records the corrected RTX 5090 target and explains that prior PRO 6000 preparation did not run ThinkingCap inference or a benchmark.
docs/findings/2026-09-23-thinkingcap-evidence/run-plan-initial-superseded.md	wrong-target preparation	Preserved superseded plan that identifies the earlier PRO 6000 work as a targeting mistake with no ThinkingCap measurement.
docs/findings/2026-09-23-thinkingcap-evidence/friction-log.md	wrong-target preparation	Operational chronology only: it records the corrected RTX 5090 authorization and that the PRO 6000/GLM line was superseded, without a ThinkingCap measurement.

2026-10-13 source-directory additions

docs/findings/2026-10-13-full-qualification-evidence/source-directory.md; docs/findings/2026-10-13-full-qualification-evidence/community-addendum.md; docs/findings/2026-10-13-full-qualification-evidence/source-guide.md (external-prior: public hardware reports and recipe leads, with external measurements kept separate from this campaign).

2026-10-17 GLM FP4 additions

docs/findings/2026-10-17-glm53-serving-performance.md (measured-on: dual-RTX-PRO-6000 matched serving comparison with 24 accepted cells and 12 paired candidate/control populations; original controlled matrix remains partial, the candidate is not promoted);

docs/findings/2026-10-17-glm53-nvidia-fp4-ab.md (measured-on: prior bounded candidate/control functional, context, diagnostic, and restoration evidence; its timing-unmeasured limitation is scoped to that quality campaign);

docs/findings/2026-10-17-glm53-nvidia-fp4-ab-evidence/diagnostic-research.md (external-prior: pinned recipe and runtime fault-isolation research only; it records no additional local measurement).

docs/findings/2026-10-17-glm53-serving-performance-evidence/plans/source/skills/anvil-serving-benchmark-docs/SKILL.md (topology-only: frozen publication-methodology snapshot; its hardware examples do not constitute another measured run).

2026-10-19 Ignis 0.7.2

docs/findings/2026-10-19-ignis-072.md (measured-on: one RTX PRO6000 Blackwell Max-Q; bounded functional/deadline checks and incomplete coding qualification; no promotion or capacity claim).