[firetussin]

Summary

Add a dated follow-up to the RX 6700 XT report merged in #52. This report contains 28 completed official tools/strata_bench.py suites, measured October 25/26, 2026, on the same RX 6700 XT 12 GB / Ryzen 5 3600 machine, now with 64 GB DDR4-3200.

Tested software: v0.1.45-poly.2 at 974552f562aecb3e86cdd2679e76e4515b52b9d2, and v0.1.45-poly.4 at 145493e66580c80bba09b0130f667a2b1789da69. Both native gfx1031 HIP binaries report engine version 0.1.45; exact binary hashes distinguish them. Both source builds reuse the pinned ROCm 7.13.0a20260515 environment and llama.cpp dependency.

What changed

• Add bench/results/2026-10-26-community-rx6700xt-poly2-poly4-memory/ with all 28 per-request JSON files, generated official reports, sanitized server configurations, supplementary provenance, CSV summary and file manifest.
• Compare seven configurations across two releases and two inference-container limits (28/52 GiB), all at 65,536 configured context and three official repetitions. Include the seven frozen initial expert-profile seeds; all four cells of each configuration use the same seed and settings apart from software paths and memory limits.
• Preserve four LFM poly.4 agent requests that returned no tokens, two at each RAM limit. The other 26 suites have no request failures; all suites have zero skips and no Docker OOM kill.
• Link the new report from the community guide and add a follow-up link to the original report. Earlier results remain intact.

Extra Notes

The memory factor measures available capacity on the upgraded machine. It does not recreate the former physical 32 GB DIMMs or isolate a historical memory-bandwidth gain. The earlier report also used different container limits, source and workloads, so cross-date changes remain observational.

The official client and document corpus were frozen from poly.2. Its client hash also matches poly.4. Actual maximum inputs are approximately 32K, reported per suite; 64K allocation is not represented as full-64K input validation. Sequential trials show mixed speed changes and do not establish a general speedup, equal quality or an overall model winner.

Validation: every inference-results subtree and non-string JSON value compared unchanged against frozen evidence after path redaction; all 28 matrix cells, memory/swap limits, matched settings/seeds, source commits, binary/model provenance, CSV metrics, hashes, manifest and local links checked. Published client/corpus source hashes match the pinned baseline checkout. Privacy scan and git diff --check passed. The upstream AMD detector's empty GPU fields and host-visible RAM remain as emitted, with verified hardware and effective limits documented separately.

Documentation/data only. V3.1 quality tests, business diagnostics, larger-capacity candidates and the later 128K appendix are outside this submission. No engine/runtime changes, private identifiers, credentials, LAN addresses, custom task outputs, model binaries or generated packs are included.

[VecSzn]

thanks, i merged it. also, the LFM agent turns with no text showing up here too, planning to look into that.