B3 staging seed — 21 runs + catalogue v1.0-draft + methodology + README

Initial population of the weeyuga-benchmarks-public archive (PRIVATE
staging visibility — flips public after Miljan + Stevan security audit
sign-off per Sloba's 17:34Z dispatch).

Contents:
- README.md       — public-facing intro (warns staging state, schema overview, citation pattern, license split)
- LICENSE         — CC-BY-4.0 default (auto-init from Gitea)
- catalogue.json  — schema_version=1.0-draft (locked once Tomas ratifies); 21 benchmarks indexed, 13 complete + 8 meta-only
- methodology.md  — mirror of WeeyugaWeb docs/BENCHMARKS/HARNESS.md (canonical methodology)
- runs/<id>/run.jsonl|run.log|run.md|metadata.json — packaged copies of every run in WeeyugaWeb docs/BENCHMARKS/runs/*

Run set covers:
- Mission 1 (2026-04-28/29): pavilion-weeyuga-v1 + reconstructed v3 (96 calls, 16 models routed via weeyuga :11435)
- Predator trio (2026-05-04): granite-4.1-8B + gemma-4-E4B-it + qwen3.5-9B
- Predator qwen rerun (2026-05-04): qwen3.5-9B think500/nothink + qwen3-14B feasibility
- A3B campaign (2026-05-04/05): pavilion-a3b + predator-a3b NGL matrix + ctx sweep + NGL+ctx 2D + NGL=6 deep dive
- VPS50 CPU matrix + gemma-e4b CPU lane (2026-05-04/05)

Visibility GATE: this repo stays private until Miljan G1-G4 audit and
Stevan G3 credential audit both green. After sign-off, single API call
flips visibility=public, anonymous read on, push-protection requires
auth, issues moderate by default.

No raw IPs, no SSH user@host strings, no /Users/ paths, no whisper
transcripts in any of these files. Hardware names (pavilion, predator,
vps50) are intentional and fine to share.

Builder: WeeyugaWeb/scripts/benchmarks/build_catalogue.py (deterministic,
idempotent, ~5s wall on 21 runs).
Publish flow: WeeyugaWeb/scripts/benchmarks/publish_bench_run.py
(builds packaged dirs, regenerates catalogue, optional --push to mirror
into this repo, optional --deploy stub for cicd rsync).

Owner: mac/benchmark-tester-ben (Ben).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

runs/d1cff064-9141-48e0-b1d2-729b134e0543/run.md

@@ -0,0 +1,52 @@
+# Predator Qwen rerun — d1cff064-9141-48e0-b1d2-729b134e0543
+
+- Started: 2026-05-04T22:02:28Z
+- Driver host: Slobodans-MacBook-Air.local
+- Node: pavilion (?)
+- Engine: llamacpp
+- Harness: pavilion-a3b-1
+- Purpose: Sloba 2026-05-04: Qwen3-30B-A3B (MoE) on weaker Pavilion hardware (GTX 1050 4GB / 16GB RAM) for cross-machine A3B comparison
+
+## Configurations
+
+| Key | Cell | GGUF | Reasoning budget | Sampling | System prompt |
+|---|---|---|---|---|---|
+| qwen-9b-think500 | `?` | `Qwen3.5-9B-Q4_K_M.gguf` | 500 | T=1.0 top_p=0.95 top_k=20 (thinking-mode) | — |
+| qwen-9b-nothink | `?` | `Qwen3.5-9B-Q4_K_M.gguf` | 0 | T=0.7 top_p=0.8 top_k=20 (instruct-mode) | Answer directly without showing reasoning. |
+| qwen3-14b | `?` | `Qwen3-14B-Q4_K_M.gguf` | 1024 | T=0.7 top_p=0.8 top_k=20 (Qwen3 default) | — |
+
+## Skipped
+- **qwen3-30b-a3b-pavilion**: server_did_not_start
+
+## llama-bench synthetic throughput (pp512 + tg128)
+
+| Cell | pp512 (t/s) | tg128 (t/s) |
+|---|---|---|
+
+## VRAM snapshot (used / free / total MB, post-load)
+
+| Cell | Used | Free | Total |
+|---|---|---|---|
+
+## Per-prompt latency, throughput, and thinking-vs-answer split
+
+| Cell | Prompt | Cold dur (ms) | Cold tok/s | Warm tok/s mean | Warm dur p50 (ms) | Reasoning chars (warm mean) | Answer chars (warm mean) | Errors |
+|---|---|---|---|---|---|---|---|---|
+| qwen-9b-think500 | hello | — | — | — | — | — | — | 0 |
+| qwen-9b-think500 | P-MEDIUM | — | — | — | — | — | — | 0 |
+| qwen-9b-think500 | P-HARD | — | — | — | — | — | — | 0 |
+| qwen-9b-nothink | hello | — | — | — | — | — | — | 0 |
+| qwen-9b-nothink | P-MEDIUM | — | — | — | — | — | — | 0 |
+| qwen-9b-nothink | P-HARD | — | — | — | — | — | — | 0 |
+| qwen3-14b | hello | — | — | — | — | — | — | 0 |
+| qwen3-14b | P-MEDIUM | — | — | — | — | — | — | 0 |
+| qwen3-14b | P-HARD | — | — | — | — | — | — | 0 |
+
+## Notes
+
+- `Reasoning chars` is `len(message.reasoning_content)` from llama.cpp's response;
+  `Answer chars` is `len(message.content)`. Both averaged over the 3 warm runs.
+- For `qwen-9b-nothink`, expectation is `Reasoning chars` ≈ 0 (budget=0 + system instruction).
+- For `qwen-9b-think500`, expectation is `Reasoning chars` materially smaller than the trio-bench's ~1024-budget run; the cap should bound the worst-case.
+- For `qwen3-14b` on a 6 GB GTX 1060 with an 8.4 GB Q4_K_M GGUF: partial CPU offload kicks
+  in at load time; throughput will be 30-50% of the 9B run.