How we measured this

Orchestrator adapter runs (November 2, 2026): trained on one NVIDIA H100 80GB using full bfloat16 (bf16) base weights of Meta’s open-weight Muse Glimmer 30B with a rank-16 LoRA adapter across 7 projections in all 52 text layers (95,846,400 trainable parameters, no 4-bit quantization), over one epoch of 1,762 rows covering four roles: retrieval, review, routing, and typed MCP tool calling. Base and adapter arms were scored in the same H100 session with greedy decoding and one shared code path.

Zero-shim memory experiment (October 31, 2026): a separate experiment on two NVIDIA T4 GPUs using 4-bit NF4 storage with float16 compute, evaluated on 120 held-out cases under a 352-token generation cap (60 post-cutoff 2026 Task Object recalls and 60 nonprofit intake splits). Across all 120 cases, adapter weights with zero context-shim tokens reached 62.5% exact match against 29.2% for the base model given a 222-token lexical shim (+33.3 percentage points, paired 95% CI [+22.5, +44.2]; 0.0% closed-book), with 2.28x lower amortized generation time per row in batch-of-8 evaluation (8.668 s vs 19.749 s; 76 of 120 base-plus-shim outputs hit the 352-token cap vs 0 tuned). Most of the combined gain came from intake splitting (34 of 60 vs 0 of 60; all 60 generated intake plans passed worker-separation and contributor-routing field checks), while recall alone was 41 of 60 vs 35 of 60 (68.3% vs 58.3%, paired 95% CI [−5.0, +25.0], spanning zero).

How to read the retrieval numbers: scored strictly, the base model gets 1.7% on the 175 held-out queries because it usually answers with the list number of the right candidate, not its id. Resolving those numbers gives the base 67.4% at the registered token limit and 88.0% with its full reply, against 89.1% for the tuned adapter. Tuning mainly fixed output format and token use, so we compare against the resolved numbers.

Original registration (recorded 2026-11-02T00:33:21Z before any H100 multi-role training run): 194 retrieval cases (175 queries with a matching Task Object and 19 with no match), 204 review submissions (78 valid, 126 flawed), 106 request-splitting prompts, 150 typed MCP tool-call prompts, and 40 general sanity prompts. Training rows share no seed pack, sentence bank, or normalized user text with these holdouts, and no Meta API output appears in the training set. Later ablations and the adoption decision reused these holdouts: they are training-held-out development evaluations, not a fresh post-selection confirmation.

Derived 153-case abstention evaluation: after the original multi-role run answered “none” on only 10 of the 19 no-match cases, we constructed a larger 153-case out-of-scope abstention set by swapping each matched retrieval query’s target Task Object for a non-matching sibling of the same kind. Because this 153-case set was derived after the initial registration timestamp, we report its provenance separately; a manual audit of 54 filtered rows found 0 mislabeled cases (95% upper bound 6.6%).

Fixed vs longer response budgets on tool calling and review: the registered scoring pipeline applies per-role token budgets before scoring (256 tokens for tool calling, 192 for retrieval, 128 for review, 384 for routing, and 96 for sanity). Within the 256-token tool-call budget, the tuned adapter reaches 99.3% strict exact match against 94.0% for base (+5.3 points, 95% CI [+2.0, +9.3]). When scored over full 384-token replies without the role cap, the base model’s reasoning trace less often crowds out its JSON call, raising base to 96.7% against 99.3% tuned (+2.7 points, 95% CI [−0.7, +6.0], interval includes zero). On review, removing the 128-token cap lifts base balanced accuracy from 28.3% to 49.9% against 99.2%–99.6% tuned.

Recorded endpoint verification (November 2, 2026): the adopted “nothing fits” Muse Glimmer 30B adapter was deployed to a dedicated serverless H100 evaluation endpoint. Seven calls were recorded as correct at 17:25 UTC: one cold call at 55.2 seconds and six warm calls across four roles, with a combined median of 3.27 seconds. This small mixed-role sample does not establish general request latency. These records do not establish routine Glimmer routing in the live Eranos fleet; its recorded operational policy uses hosted Muse Spark 1.3 and Muse Wonder 1.0 (version 28, 15 outcomes as of November 2).

Local queue workflow, concurrency, and fault-injection measurements (October 30, 2026): the local queue test suite separates two measurements across 4 local Worker/SQLite cells. First, a 16-pipeline workflow test ran 1,000 generated workflows (25,620 MCP calls at 54.441 ms median and 145.205 ms 95th-percentile call duration; a separate 10-run CLI wave had 9 runtime-verified completions). Second, a 5-tier concurrency test completed 992 Tasks in total (512 Tasks in the 256-slot tier at 678.99 ms median and 1,354.95 ms 95th-percentile call duration across 1,792 MCP calls) with 0 duplicate (cell, Task, fencing-token) lease tuples. Controlled fault-injection checks recorded 16 of 16 expired-lease cases where a replacement worker completed under fencing token 2 and both late heartbeats and late submissions from the stale worker received HTTP 409, plus 16 of 16 direct self-approval attempts and 16 of 16 role/queue permission checks refused with HTTP 403. The duplicate-tuple detector does not observe every possible ownership overlap or establish exactly-once external effects.

Scope and reproducibility: the JSON links below are summaries of sample counts, arm metrics and paired bootstrap intervals, not the row-level model outputs or a full replay package. Adapter weights remain private under the base model’s license terms. A free practice session exercises live queues and storage rather than re-running H100 model training.

The 62.5 percent versus 29.2 percent comparison combines 60 recall cases and 60 nonprofit request-splitting cases. It compares a tuned model without the retrieval shim with the baseline model using the shim, under the same 352-token response budget.

Recall alone was 41 of 60 versus 35 of 60. Most of the combined difference came from request splitting. The baseline with retrieval reached the generation cap in 76 of 120 cases; the tuned model did not. These results do not isolate memory from training, formatting or response-budget effects.

Review of 560 saved outputs reproduced the recorded per-row scores and 18 paired comparisons without new model calls. This is a recorded development experiment, not a customer-outcome or production-latency measurement.

Recorded local workerd and SQLite tests refused 16 direct self-approval attempts with HTTP 403. A separate set of 16 role and queue capability refusals is not counted as another 16 self-approval attempts.

Sixteen deliberately expired-task cases across four local cells exercised recovery: replacement workers used newer fencing tokens, and old-worker heartbeats and late submissions were refused with HTTP 409. This checks the tested permission and stale-writer boundaries; it does not establish exactly-once external effects or production uptime.