OMP Strata

Companion for Oh My Pi · community project

Stock Oh My Pi on a stock, local Strata server. Nothing forked, everything pinned.

Keep using Oh My Pi in your terminal. OMP Strata pins and verifies every download, installs Strata with its own unmodified setup.py from verified local inputs, starts it on loopback behind an API key, and launches stock OMP in an isolated profile where every model role points at that server. One Windows 11 + RTX host, one GPU you own, no proxy, no cloud fallback.

Install. From a checkout of the repository, in PowerShell; the quickstart has the exact tested commands. The integration lives in its own root; your normal OMP configuration and %APPDATA%\Strata are never touched.
py -3 scripts\omp_strata.py validate   --profile $prof
py -3 scripts\omp_strata.py fetch      --profile $prof --root $root
py -3 scripts\omp_strata.py install    --profile $prof --root $root
py -3 scripts\omp_strata.py keygen     --profile $prof --root $root
py -3 scripts\omp_strata.py start      --profile $prof --root $root
py -3 scripts\omp_strata.py launch-omp --profile $prof --root $root
Get started Read the decision

What the tool does

  1. A profile under profiles/ is the single source of truth: host expectations, stock setup choices, and every artifact pinned by URL, size and SHA-256. A changed component or setting is a new profile with its own root and ledger.
  2. fetch downloads the pinned OMP binary, the Strata engine and llama.cpp archives and the model shards, resumable, and verifies every byte before anything is used.
  3. install runs stock Strata's own setup.py from those local inputs only, with its data folder redirected into the integration root, and checks the generated engine flags against the profile.
  4. start binds Strata to 127.0.0.1 behind a user-only key file, refuses to start if the port, GPU or RAM budget is not free, and returns once authenticated identity checks pass.
  5. launch-omp runs the pinned stock OMP binary in an isolated profile: every model role on the local server, discovery off, retries and model fallback off, ambient provider credentials scrubbed.
  6. Gate probes, a tracer and a frozen six-task evaluation run on the real host; scrubbed receipts under releases/ bind each result to the exact profile fingerprint.

Use it when

Use something else when

What stays on your machine

Inference needs no network. Strata listens on loopback only, every model, control and metrics route needs the API key, and the key lives in a user-only file that is never put on a command line, printed, logged or committed. Stock OMP runs with an isolated home and profile; the launcher points its HTTP proxy variables at a closed loopback port, so OMP's startup catalog fetch fails locally while the Strata route stays direct, and no retry, fallback or discovery can route a request anywhere but the local server. The host is treated as single-user: this is not a sandbox, a multi-tenant service or an egress firewall. Full model: security boundary · operations.

Status

Qualified and current across all three hosts and the Mac Studio route (2026-11-09). The RTX PRO 6000 (four batch slots with per-slot stock DFlash2 drafters at window 8 + direct per-layer embedding (PLE) + KV in VRAM + stock parallel and GDN stacks; tuned-DFlash2 v0.2.13 alternate at --draft-conf 0.55, v0.2.12 DFlash2 rollback), the RTX 5090 (262K Coder IQ1_M default with KV in VRAM + KV growth + GDN stack, 131K Coder IQ1_M alternative, and full-model IQ3_S 131K low-RAM resident alternate; v0.2.12 262K envg and 131K rollbacks) and the RTX 4090 + RTX 3090 pair (stock DFlash2 drafter with strata.drafting_args --draft-conf 0.5 + split-24 stage-weight trim + direct PLE + engine sm_86 table, GDN stack and STRATA_SELECT_TWO_PASS=1; v0.2.12 env86gt two-pass DFlash2 rollback) run stock Strata v0.2.13 (engine 0.1.59) with stock OMP 18.18.3. The current PRO profile, win11-rtxpro6000-iq3s-262k-slots4-df2-pledirect-kvvram-envpg-parking-strata0.2.13-omp18.18.3, keeps IQ3_S at the trained 262,144-token context with every expert in VRAM, four batch slots with per-slot stock DFlash2 drafters (strata.drafting: "dflash2", one 85 MiB slot drafter per slot under STRATA_BATCH_DF2, --spec 8 on stock v0.2.13 instead of --spec 6 on v0.2.12, engine cap 8, no --batch-mtp; packed offline from pinned incoai/Qwen3.8-Flash-Next-DFlash2 files with STRATA_NO_ACQUIRE=1; 93 GiB free-disk floor), direct PLE reads from the model file on SSD (--ple-io direct, strata.ple_io: "direct"), KV in VRAM (--kv-streaming off, setup_args.kv_streaming: "off"; 78 GiB available-RAM start floor), stock nine-switch parallel-slot stack and two-switch GDN tensor-core prompt stack in strata.engine_env (the launcher's parallel and GDN env stacks moved into engine 0.1.59 as Linux-only defaults, so our Windows groups stay explicit and are checked against the engine's tables), conversation parking (the engine's 8,192 MiB automatic cache) and bare --prefill auto (32,768-token ceiling): in three paired cold starts against its v0.2.12 DFlash2 predecessor (A1 B1 B2 A2 A3 B3, reporting B / A geometric means under the owner's 5% non-regression rule), 11 of 90 checks missed from stock v0.2.13's window-8 output-event spacing and concurrent tail latency (inter-output-event latency p95 1.171-1.188 at every depth, w2/c1 largest gap p95 1.173, w2/c4 TTFT p95 1.118, w2/c4 TPOT max 1.067, while throughput 0.986-0.999, per-request 0.964-1.015, TTFT p50 1.009-1.036 and TPOT p50 0.981-1.033 passed and server decode reported 0.958-1.015), overridden by the owner on 2026-11-09 ("Override the base as current, and also qualify the tuned pack as an alternate"), with every gate passing including G15, G17 and G18L, G24 15/18 (tool-loop 0/3, as v0.2.12), published at 2af52b8. Its qualified non-current alternate at the same release, win11-rtxpro6000-iq3s-262k-slots4-df2tuned-pledirect-kvvram-envpg-parking-strata0.2.13-omp18.18.3 (stock tuned DFlash2 pack qwen3.8-flash-next-dflash2.strata1.q4_0-selq4.pack, 339,694,592 bytes, 94 GiB disk floor, which stock v0.2.13 now runs at --draft-conf 0.55 instead of 0.35 on this card; missed 15 of 90 checks vs v0.2.12 on the same burstiness — inter-output p95 1.169-1.172, w2/c1 largest gap p95 2.151, w2/c4 TTFT p50 1.129 — with w2/c2 per-request 1.084 and TPOT p50 0.928, not the rollback), passed every gate under the same owner decision, with G24 15/18 (tool-loop 1/3), published at 8ab88b0. Its qualified rollback is the v0.2.12 DFlash2 profile win11-rtxpro6000-iq3s-262k-slots4-df2-pledirect-kvvram-envpg-parking-strata0.2.12-omp18.18.0 (--spec 6, 31.1-31.3 s restart readiness in G16/G20/G21, G15 kill recall 44.2-44.4 s, lowest available RAM 113.51 GiB in G19 and 113.60 GiB in G18L, 83.13 GiB G10 Windows commit headroom of a 219.70 GiB limit, 126.94 GiB available RAM, 73,526 MiB GPU memory, 24 parks in G19; the v0.2.12 tuned-pack alternate win11-rtxpro6000-iq3s-262k-slots4-df2tuned-pledirect-kvvram-envpg-parking-strata0.2.12-omp18.18.0, the v0.2.11 DFlash2 profile win11-rtxpro6000-iq3s-262k-slots4-df2-pledirect-kvvram-envpg-parking-strata0.2.11-omp18.16.0, the v0.2.11 batch-MTP profile win11-rtxpro6000-iq3s-262k-slots4-batchmtp-pledirect-kvvram-envpg-parking-strata0.2.11-omp18.16.0, the v0.2.10 kvvram-envpg profile win11-rtxpro6000-iq3s-262k-slots4-batchmtp-pledirect-kvvram-envpg-parking-strata0.2.10-omp18.16.0 and v0.2.9 profiles move to earlier qualified history; seven old v0.2.3-v0.2.9 IQ3_S host roots were pruned on 2026-11-06 with the owner's approval, while the v0.2.13 standard-pack --draft-conf 0.55 draft df2conf055 — withdrawn before install with a "Not qualified" blocker at b2e9730 because the tuned pack at 0.55 showed the same window-8 burstiness — the v0.2.12 two-pass draft win11-rtxpro6000-iq3s-262k-slots4-df2-pledirect-kvvram-envpgt-parking-strata0.2.12-omp18.18.0, the v0.2.11 quality-first one-slot IQ3_S, one-slot UD-Q5_K_XL and six-expert DeepSeek V4.1-Flash IQ2_XS drafts, the v0.2.11 UD-Q4_K_XL screen win11-rtxpro6000-ud-q4kxl-262k-slots4-batchmtp-pledirect-parking-strata0.2.11-omp18.16.0 and the v0.2.9 UD-IQ4_XS / NVFP4-QAD model candidates stay unqualified or comparison-only drafts; GDN tensor-core numerics are not bit-exact upstream, and engine env pins are reviewed per engine release in ENGINE_ENV_REVIEWED_ENGINES, engines 0.1.55, 0.1.56, 0.1.57, 0.1.58 and 0.1.59). RTX 5090 262K default, 131K Coder alternative and full-model IQ3_S 131K low-RAM alternate (Strata v0.2.13 + OMP 18.18.3, owner decisions 2026-11-09): on rtx5090-win-a, the tuned 262,144-token profile, win11-rtx5090-coder-iq1m-262k-kvvram-kvgrow-envg-strata0.2.13-omp18.18.3, is the host's default: Coder IQ1_M at 262,144 tokens, pinning setup_args.kv_streaming: "off" (--kv-streaming off, 34 GiB available-RAM start floor), strata.kv_grow: true (--kv-grow) and stock gdn_stack in strata.engine_env (cold-start A/B against v0.2.12 passed all 88 checks: throughput w1 1.006 / w3 0.995, per-request 1.024 / 0.975, TTFT p50 0.989 / 1.002, TPOT p50 0.975 / 1.023, cold decode 0.986-1.018, prefill 0.998-1.075; every gate passed including G15 and G18L, G24 16/18 with tool-loop 1/3, published at d3c96fb), rolling back to win11-rtx5090-coder-iq1m-262k-kvvram-kvgrow-envg-strata0.2.12-omp18.18.0 (4,096 MiB automatic conversation cache, 16 parks in G19; G10 serving: 13.05 GiB available RAM and 19.53 GiB commit headroom of an 85.15 GiB limit; G16/G20/G21 restart ready in 19.1-19.3 s, lowest available RAM 7.77 GiB in G18L; the v0.2.11 kvvram-kvgrow-envg profile win11-rtx5090-coder-iq1m-262k-kvvram-kvgrow-envg-strata0.2.11-omp18.16.0, the v0.2.10 kvvram-kvgrow-envg profile win11-rtx5090-coder-iq1m-262k-kvvram-kvgrow-envg-strata0.2.10-omp18.16.0, the v0.2.9 kvvram-kvgrow-envg profile win11-rtx5090-coder-iq1m-262k-kvvram-kvgrow-envg-strata0.2.9-omp18.15.0 and the v0.2.9 base profile win11-rtx5090-coder-iq1m-262k-strata0.2.9-omp18.15.0 stay earlier qualified history; the v0.2.13 STRATA_SPEC_COST=1 spec_cost arm ...-envgc-... missed 3 of 88 checks against v0.2.13 envg and 2 of 88 in three replacement pairs with about one 4x largest-gap spike per three cold starts, was not adopted by the owner, and stays an unqualified draft with a "Not qualified" blocker at b2e9730 alongside the v0.2.12 two-pass arm win11-rtx5090-coder-iq1m-262k-kvvram-kvgrow-envgt-strata0.2.12-omp18.18.0 and the unpinned-GDN v0.2.9 fallback win11-rtx5090-coder-iq1m-262k-kvvram-kvgrow-strata0.2.9-omp18.15.0). Operating condition: the host's Docker Desktop and WSL, which run its production NInfer container, must be stopped; after WSL restarts, that container must be recreated by the owner's pinned relaunch script or returned through the owner's fleet controller. The 131,072-token Coder IQ1_M profile, win11-rtx5090-coder-iq1m-131k-strata0.2.13-omp18.18.3, is the current 131K alternative for running with Docker Desktop and WSL up (the production container stopped by the fleet controller; qualified on the same window's passing 262K A/B without a separate A/B, as in v0.2.12, with every gate passing, G24 16/18, published at d0c2e28; 34 GiB start floor), rolling back to win11-rtx5090-coder-iq1m-131k-strata0.2.12-omp18.18.0 (3,584 MiB cache with 19 parks in G19, G10 11.37 GiB available RAM and 18.72 GiB commit headroom of 85.15 GiB, G15 recall 30.9-32.0 s, 16.6-18.9 s restart readiness in the Docker/WSL-stopped window, lowest available RAM 7.08 GiB in G19; win11-rtx5090-coder-iq1m-131k-strata0.2.10-omp18.16.0, win11-rtx5090-coder-iq1m-131k-strata0.2.9-omp18.15.0 and win11-rtx5090-coder-iq1m-131k-strata0.2.8-omp18.14.0 stay earlier qualified history, while the v0.2.11 131K draft win11-rtx5090-coder-iq1m-131k-strata0.2.11-omp18.16.0 failed G15 twice and stays unqualified). The full-model IQ3_S 131,072-token low-RAM resident profile, win11-rtx5090-iq3s-131k-lowram-strata0.2.13-omp18.18.3 (stock's planner pick for this host, keeping 11,616 expert slots in VRAM at 22.04 GiB and 29.77 GiB of experts page-locked in RAM; report-only screen B = IQ3_S vs A = Coder IQ1_M 131K on v0.2.13 measured cold decode 1.046 at 8K, 1.093 at 32K, 1.071 at 100K and 1.094 on warm 32K vs 0.763 at the empty prompt, queue decode 1.086 / 1.086, per-request 1.134 / 1.129, TPOT p50 0.881 / 0.885, throughput w1 1.059 and w3 0.917, TTFT p50 1.528 at 512 tokens and 1.122 at 32K, and prefill 0.61-0.95, with 58/60 vs Coder 54/60 on the earlier private quality set), is qualified as a non-current alternate while the Coder stays the default (owner decision "Qualify full-model IQ3_S 131K: prune old [rtx5090-win-a] roots for disk, then run its main gates", second window after deleting five superseded RTX 5090 roots for disk, C: 98.1 to 156.6 GiB free; every gate passed including G15, G17, G18 and G18L, G24 16/18 with tool-loop 1/3, published at 6e2dda2; the full-model IQ3_S 262K low-RAM screen draft ...-iq3s-262k-lowram-kvvram-kvgrow-envg-... under "Also screen IQ3_S at 262K" passed install and was refused on first start with "strata.kv_grow: elastic K/V line absent from this start's engine-log interval", staying an unqualified draft with a "Not qualified" blocker at b2e9730). The qualified RTX 5090 ledgers and their rollbacks (262K default v0.2.13, 262K rollback v0.2.12, 131K alternative v0.2.13, 131K IQ3_S low-RAM alternate v0.2.13, 131K rollback v0.2.12) have 21 pass and 3 not applicable. On rtx3090-win-a, the current pair profile is win11-rtx3090-rtx4090-iq3s-262k-df2conf05-disp4090-roles-split24-trimweights-pledirect-env86gt-strata0.2.13-omp18.18.3: IQ3_S at 262,144 tokens, stock roles (--gpu-roles 0=main,1=helper, RTX 4090 as Main, staged last), manual layer split 24 (--layer-split 24, the layer the automatic planner chose), stage dense-weight trim (--trim-stage-weights, strata.trim_stage_weights: true), stock DFlash2 drafting alone (strata.drafting: "dflash2", --draft dflash2 --draft-model <root>\data\dflash2\qwen3.8-flash-next-dflash2.q4_0-selq4.pack --draft-fallback none --draft-conf 0.5 --draft-vocab <root>\data\dflash2\draft_vocab.bin, packed offline during install by stock tools/dflash2_pack.py from pinned incoai/Qwen3.8-Flash-Next-DFlash2 files with STRATA_NO_ACQUIRE=1; drops --pipeline-windows 2, sizing the expert caches at 19.40 GiB on the RTX 3090 and 17.90 GiB on the RTX 4090 vs 19.41 / 17.90 GiB on v0.2.11 df2, 19.41 / 18.31 GiB on v0.2.10 df2 and 19.14 / 17.75 GiB on v0.2.10 pw2 env86g MTP), strata.drafting_args pinning {"--draft-conf": "0.5"} (df2conf05, stock's slot-step value for the pair's per-card slots, which stock cannot compute on a split, written to stock user_args with stock write_config and required on the start's server line), no batch slots, display on the RTX 4090, strata.ple_io: "direct" pinning stock --ple-io direct so the 26.8 GiB PLE table stays on direct SSD reads, the engine's 8,192 MiB automatic conversation cache stays enabled (the df2conf05 and env86gt ids drop -parking only to stay within the 120-character profile-id limit, with the same strata.stock_tips), and strata.engine_env pinning the engine's sm_86 default table (sm86_table; Niko1221/Strata#3313), stock gdn_stack and two-pass QSA block selection (select_two_pass, STRATA_SELECT_TWO_PASS=1): while the v0.2.13 df2 env86gt base draft (at stock --draft-conf 0.35) missed 6 of 96 checks against v0.2.12 on one 32K request (w3/c1 TPOT p50 1.054, p95 1.069, max 1.070; largest gap p50 1.119, p95 1.239, max 1.252) and the v0.2.13 STRATA_QSA_SPLIT_ATTN=1 (qsa_split_attn, ...-env86gts-...) arm missed 2 of 96 checks (w1 largest gap p95 1.075, max 1.092; targeted cold decode 1.003 < 1.02), the df2conf05 --draft-conf 0.5 arm passed all 96 plain checks against v0.2.12 (targeted cold decode 1.021, [1.020, 1.049, 0.993]; throughput 1.016 / 0.998, per-request 1.027 / 0.993, TTFT p50 1.005 / 1.001, TPOT p50 0.973 / 1.009; decode 100K 1.061, warm 32K 1.054, 32K 1.044) and, after fb8740a fixed install verification for stock v0.2.13's carry_user_args, passed every gate including G15, G17, G18 and G18L in its requalification chain at 7c2d25e (440fa1b; G24 16/18 with tool-loop 1/3) under the owner's decision ("Run the --draft-conf 0.5 main; it becomes current if every gate passes"), rolling back to win11-rtx3090-rtx4090-iq3s-262k-df2-disp4090-roles-split24-trimweights-pledirect-env86gt-strata0.2.12-omp18.18.0 (--draft-conf 0.35, 55.3-55.5 s restart readiness, G17 near-limit task 57.7 s, 22 parks in G19, G10 28.21 GiB commit headroom of 135.69 GiB and 65.74 GiB available RAM, G24 15/18, pilot 6/6, batch wall 968 s; while the v0.2.12 env86g profile win11-rtx3090-rtx4090-iq3s-262k-df2-disp4090-roles-split24-trimweights-pledirect-env86g-parking-strata0.2.12-omp18.18.0, the v0.2.11 df2 profile win11-rtx3090-rtx4090-iq3s-262k-df2-disp4090-roles-split24-trimweights-pledirect-env86g-parking-strata0.2.11-omp18.16.0, the v0.2.10 df2 profile win11-rtx3090-rtx4090-iq3s-262k-df2-disp4090-roles-split24-trimweights-pledirect-env86g-parking-strata0.2.10-omp18.16.0, the v0.2.10 pw2 env86g MTP profile win11-rtx3090-rtx4090-iq3s-262k-pw2-disp4090-roles-split24-trimweights-pledirect-env86g-parking-strata0.2.10-omp18.16.0, the v0.2.9 env86g profile win11-rtx3090-rtx4090-iq3s-262k-pw2-disp4090-roles-split24-trimweights-pledirect-env86g-parking-strata0.2.9-omp18.15.0 and the v0.2.9 base profile win11-rtx3090-rtx4090-iq3s-262k-pw2-disp4090-roles-split24-trimweights-pledirect-parking-strata0.2.9-omp18.15.0 remain earlier qualified history; the v0.2.13 df2 env86gt and env86gts drafts carry "Not qualified" blockers at b2e9730 alongside the unqualified v0.2.12 W4A8 arm win11-rtx3090-rtx4090-iq3s-262k-df2-disp4090-roles-split24-trimweights-pledirect-env86gw-strata0.2.12-omp18.18.0; three v0.2.8/v0.2.9 pair host roots were pruned on 2026-11-06 with the owner's approval). On the pair's measurement-only v0.2.8 stock-residency roles draft, upstream's Niko1221/Strata#4518 fix locked the 26.8 GiB PLE table (33.9-34.0 s), then gave the lock back on both starts and sized the automatic conversation cache at 4,864 MiB (106-112 s to ready), against 8,192 MiB and 61-63 s to ready with --ple-io direct at matching 100K-token TTFT (31.2 s vs 30.9 s), so the pair keeps --ple-io direct. Each current local ledger has 21 pass and 3 not applicable. Real-host v0.2.13 runs on 2026-11-09 are published at 2af52b8 on the PRO df2 DFlash2 current chain and the Studio df2 current route G23, 8ab88b0 on the PRO df2tuned tuned-pack alternate chain and Studio df2tuned alternate route G23, d3c96fb on the RTX 5090 262K kvvram-kvgrow-envg default chain, d0c2e28 on the RTX 5090 131K Coder alternative chain, 6e2dda2 on the RTX 5090 full-model IQ3_S 131K low-RAM alternate chain, and 440fa1b (at fixed runner 7c2d25e, 18:43-19:26Z, after admitting engine 0.1.59's stage order at 6eb9f05 and fixing install verification for stock v0.2.13's carry_user_args at fb8740a) on the pair df2conf05 --draft-conf 0.5 current chain, with restored-state records at ab96ab4 and "Not qualified" blockers on the five unqualified v0.2.13 drafts at b2e9730 (earlier v0.2.12 receipts cite 243290e, 57dcd18, 2560f77 and 916f1d9). Scored (G24): PRO df2 DFlash2 current 15/18 (tool-loop 0/3; v0.2.13 df2tuned alternate: 15/18, tool-loop 1/3; v0.2.12 df2 rollback: 15/18, pilot 6/6; v0.2.12 df2tuned: 15/18, pilot 5/6; v0.2.11 df2: 15/18, pilot 5/6; v0.2.11 batchmtp: 16/18, pilot 5/6), RTX 5090 262K kvvram-kvgrow-envg default 16/18 (tool-loop 1/3; v0.2.12 rollback: 15/18, pilot 6/6; v0.2.11: 17/18, pilot 6/6; unqualified v0.2.12 envgt draft: 17/18, pilot 5/6), RTX 5090 131K Coder alternative 16/18 and full-model IQ3_S 131K low-RAM alternate 16/18 (tool-loop 1/3; v0.2.12 rollback: 15/18, pilot 5/6; v0.2.10: 16/18, pilot 6/6; unqualified v0.2.11 draft: 15/18, pilot 6/6), current pair df2conf05 --draft-conf 0.5 16/18 (tool-loop 1/3; v0.2.12 env86gt rollback: 15/18, pilot 6/6; v0.2.12 env86g: 15/18, pilot 5/6; v0.2.11 df2: 15/18, pilot 5/6; v0.2.10 df2: 17/18, pilot 5/6). OMP 18.18.3 changed delegation bias from eager to balanced for these models, added a system prompt instruction to read small files whole, and changed bash locale/fork/console handling, so G24 and task-set scores are not comparable with 18.18.0 (just as OMP 18.17 was not directly comparable with 18.16.0). Protocol counters (including malformed arguments and calls without results) and timeouts are 0 across all qualified server profiles; task failures stay counted. Scored batch walls on the v0.2.12 rollbacks were 522 s on PRO df2 (v0.2.12 df2tuned: 581 s; v0.2.11 df2: 627 s; v0.2.11 batchmtp: 549 s), 787 s on the RTX 5090 262K kvvram-kvgrow-envg default (v0.2.11: 843 s; unqualified v0.2.12 envgt draft: 761 s), 726 s on the RTX 5090 131K alternative (v0.2.10: 742 s; unqualified v0.2.11 draft: 758 s) and 968 s on the pair df2 env86gt two-pass profile (v0.2.12 env86g: 906 s; v0.2.11 df2: 896 s; v0.2.10 df2: 916 s); this is whole-batch wall, not isolated decode latency. Qualification is integration correctness, not a claim that every task passes or runtime superiority. v0.2.13 ships its engine (0.1.59, 630,783,433 bytes) as a separately versioned CUDA 13 component tarball staged into runtime\strata\engine\, reusing locks/strata-python-cp313-win_amd64-strata0.2.11.txt (26 hash-locked wheels). On rtx3090-win-a, engine 0.1.59's stage order was admitted at 6eb9f05 after the 2026-11-09 pair measurement (following engine 0.1.58 at a5989fe): stock roles (--gpu-roles 0=main,1=helper, RTX 4090 as Main staged last, RTX 3090 first) with manual split 24 and stage weight trim puts layers 0-23 on the RTX 3090 (GPU 1, 19.40 GiB expert cache vs 19.41 GiB on v0.2.11 and v0.2.10 df2) and 24-47 on the RTX 4090 (GPU 0, 17.90 GiB vs 18.31 GiB on v0.2.10 df2, 17.75 GiB on v0.2.10 pw2; decision). The v0.2.13 campaign is complete across all five rows. The five v0.2.13 drafts carrying "Not qualified" blockers at b2e9730 (pair env86gt, missed 6/96 A/B checks on one 32K request; pair env86gts with STRATA_QSA_SPLIT_ATTN=1, missed 2/96 checks with targeted cold decode 1.003 < 1.02; RTX 5090 envgc with STRATA_SPEC_COST=1, missed 3/88 checks and 2/88 in replacement pairs; RTX 5090 full-model IQ3_S 262K low-RAM ...-iq3s-262k-lowram-kvvram-kvgrow-envg-..., refused on first start when --kv-grow's elastic K/V line was absent; and PRO df2conf055 with --draft-conf 0.55 on the standard pack, withdrawn before install), the v0.2.12 PRO two-pass server and route drafts (envpgt, missed 3/96 A/B checks), the v0.2.12 RTX 5090 two-pass draft (envgt, failed G18), the v0.2.12 pair W4A8 draft (env86gw, missed cold TTFT, 32K decode and inter-token gap in its A/B), the comparison-only v0.2.11 PRO quality-first one-slot IQ3_S, one-slot UD-Q5_K_XL and six-expert DeepSeek V4.1-Flash IQ2_XS drafts, the v0.2.11 PRO UD-Q4_K_XL screen and v0.2.9 PRO UD-IQ4_XS and NVFP4-QAD model candidate drafts (and their routes, while a 30-task Terminal-Bench comparison has not run), the v0.2.11 RTX 5090 131K draft (failed G15 twice), the RTX 5090 v0.2.9 kvvram-kvgrow fallback draft, the measurement-only v0.2.8 stock-residency pair draft, the held v0.2.7 roles draft (PLE in RAM), the unmeasured v0.2.7 default-split draft, the default and split-32 v0.2.3 drafts and the v0.1.44.1 pair release remain draft, not qualified. Server rollbacks and alternates: PRO, 262K tuned-DFlash2 alternate (v0.2.13), PRO, 262K DFlash2 rollback (v0.2.12), RTX 5090, 262K default rollback (v0.2.12), RTX 5090, 131K IQ3_S low-RAM alternate (v0.2.13), RTX 5090, 131K alternative rollback (v0.2.12), pair, 262K env86gt two-pass DFlash2 rollback (v0.2.12) (plus the earlier qualified pair v0.2.12 env86g DFlash2, pair v0.2.11 DFlash2, pair v0.2.10 DFlash2 and pair v0.2.10 env86g MTP profiles). The earlier v0.2.12 + OMP 18.18.0 PRO tuned-DFlash2, the v0.2.11 + OMP 18.16.0 PRO DFlash2, PRO batch-MTP kvvram-envpg and RTX 5090 262K envg, the v0.2.10 + OMP 18.16.0 PRO kvvram-envpg, RTX 5090 262K envg and RTX 5090 131K, the v0.2.9 + OMP 18.15.0 PRO kvvram-envpg, PRO batch-MTP + PLE-direct base, RTX 5090 262K envg, RTX 5090 262K base, RTX 5090 131K, pair env86g and pair split-24 trim + PLE-direct base profiles, the v0.2.8 + OMP 18.14.0 PRO batch-MTP + PLE-direct, PRO PLE-direct, PRO slots4-parking, RTX 5090 262K, RTX 5090 131K, pair split-24 trim + PLE-direct and pair PLE-direct profiles, the v0.2.7 + OMP 18.13.2 profiles across all three hosts, the v0.2.3 + OMP 18.12.6 profiles across all three hosts (whose RTX 5090 integration roots were removed on 2026-11-02 with the owner's approval, whose PRO and pair v0.2.3-v0.2.9 superseded roots were pruned on 2026-11-06 with the owner's approval, and where five superseded RTX 5090 roots were pruned on 2026-11-09 with the owner's approval), the v0.2.1 profiles on all three hosts (whose host integration roots were removed on 2026-10-28), the PRO and RTX 5090 v0.1.44.1 (including 262K), v0.1.43.1 and v0.1.43 profiles, and the pair's v0.1.43.1 and v0.1.43 profiles remain earlier qualified history; all earlier failures and passing reruns remain published. pair, 262K, PRO, 262K, RTX 5090. The single-card RTX 3090 Coder profile on stock Strata v0.1.40.3 with stock OMP 18.8.4 stays qualified for single-card hosts (15/18 scored); it is not the current two-card host's profile. The pair's earlier qualified 131K profile keeps the display on the RTX 4090. Its earlier predecessor requires moving the display back to the RTX 3090. Single-card RTX 3090 profiles are not rollbacks while both cards are installed: they name GPU 0, now the RTX 4090. Earlier qualified history stays unchanged: on the RTX PRO 6000, stock Strata v0.1.41 with stock OMP 18.8.6 and the same choices at 262,144 tokens (whose first G18 run failed on a long session sized for 131K windows, and whose rerun with the gate fixtures sized from the profile's window passed), then at 131,072 tokens, then the same choices on stock Strata v0.1.40.3 with stock OMP 18.8.4 (whose first G16 run failed on a predicate that counted mentions of a script rather than its runs, and whose fixed predicate's rerun passed), then on stock OMP 18.8.3, and before them stock Strata v0.1.40.2 with stock OMP 18.8.0 and none of those choices, whose run found stock OMP's speculative compaction holding up the agent's next turn on the one engine, so the integration turns that off; on the RTX 5090, stock Strata v0.1.41 with stock OMP 18.8.6, then v0.1.40.3 with stock OMP 18.8.4, then the fourth tuple (stock Strata v0.1.34, stock OMP 18.4.10, Coder IQ1_M), which also stays qualified on the single-card RTX 3090 and RTX 4090 in stock low-RAM mode. Earlier tuples remain separate, unqualified rollback installations; other profiles remain drafts. The Mac Studio's current qualified client route is client-rtxpro6000-262k-df2-pledirect-kvvram-envpg-strata0.2.13-omp18.18.3: stock OMP 18.18.3 reaches the RTX PRO 6000's current v0.2.13 262K df2 DFlash2 profile through an SSH local forward (G23 pass on 2026-11-09, published at 2af52b8: missing key refused before launch, wrong key HTTP 401, health RTT 5.8-7.7 ms and stream TTFT 0.217 s; tunnel drop failed the turn in flight with the transcript preserved, and a one-turn launch-omp smoke answered ROUTE OK). Its qualified non-current alternate, client-rtxpro6000-262k-df2tuned-pledirect-kvvram-envpg-strata0.2.13-omp18.18.3, reaches the PRO's v0.2.13 262K df2tuned tuned-pack alternate (G23 pass on 2026-11-09, published at 8ab88b0: health RTT 6.1-6.8 ms and stream TTFT 0.181 s, and a one-turn launch-omp smoke answered ROUTE OK). Its qualified rollback, client-rtxpro6000-262k-df2-pledirect-kvvram-envpg-strata0.2.12-omp18.18.0, reaches the PRO's v0.2.12 262K df2 DFlash2 rollback (G23 pass on 2026-11-07 at 243290e: missing key refused before launch, wrong key 401, health RTT 5.9-7.4 ms and stream TTFT 0.247 s; tunnel drop failed the turn in flight with the prior transcript preserved, nonce recall passed after tunnel reopen 0.87 s, client restart 0.79 s and verified server restart 2.56 s, and a one-turn launch-omp smoke answered ROUTE OK with owned stop). The 0.2.12 df2tuned, 0.2.11 df2, 0.2.11 batchmtp, 0.2.10 kvvram-envpg, 0.2.9 kvvram-envpg, 0.2.9 batch-MTP base, 0.2.8 batch-MTP, 0.2.8 slots4-parking, 0.2.7, 0.2.3, 0.2.1, 0.1.44.1, 0.1.43.1, 0.1.43 and 0.1.42 262K routes and client-rtxpro6000-strata0.1.41-omp18.8.6 131K route stay earlier qualified history (remote use and receipts). Images and runtime comparison (G22 and G25) are not applicable to these local, text-only profiles; G25 compares single-GPU profiles only. The RTX 4090 + RTX 3090, 262K DFlash2 (df2conf05) v0.2.13 ledger, RTX 4090 + RTX 3090, 262K two-pass DFlash2 (env86gt) v0.2.12 rollback ledger, RTX PRO 6000 4-slot DFlash2 (df2) v0.2.13 ledger, RTX PRO 6000 4-slot tuned-DFlash2 (df2tuned) v0.2.13 alternate ledger, RTX PRO 6000 4-slot DFlash2 (df2) v0.2.12 rollback ledger, RTX 5090 262K default v0.2.13 ledger, RTX 5090 131K alternative v0.2.13 ledger, RTX 5090 131K IQ3_S low-RAM alternate v0.2.13 ledger, single-card RTX 3090 ledger, every other ledger, measurements and decision retain the failures and passing reruns.

The OMP family