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/79afe366-1055-4e45-adf6-593864a530e8/run.log

@@ -0,0 +1,63 @@
+[09:44:24Z] === predator-a3b-ngl6-ctx run_id=79afe366-1055-4e45-adf6-593864a530e8 ===
+[09:44:24Z] Plan: NGL=6 × ctx={64k, 96k, 131k}
+[09:44:24Z] Creating run-mirror dir on Predator: D:\WeeyugaBench\runs\79afe366-1055-4e45-adf6-593864a530e8
+[09:44:34Z] >>> CELL: ngl6-ctx64k (NGL=6 ctx=65536)
+[09:44:34Z] Stopping llama-server on Predator...
+[09:44:41Z] Starting llama-server (key=ngl6-ctx64k, NGL=6, ctx=65536)...
+[09:44:48Z] Waiting for llama-server (key=ngl6-ctx64k, up to 15 min)...
+[09:45:12Z]   llama-server up after 17s
+[09:45:16Z] Running llama-bench for ngl6-ctx64k (NGL=6)...
+[09:46:31Z] Running prompts for ngl6-ctx64k (cold + 3 warm; max_tokens 512/1024/2048)...
+[09:46:31Z]   prompt=hello cold (max_tokens=512)
+[09:47:09Z]   prompt=hello warm 1
+[09:47:25Z]   prompt=hello warm 2
+[09:47:39Z]   prompt=hello warm 3
+[09:48:01Z]   prompt=P-MEDIUM cold (max_tokens=1024)
+[09:49:23Z]   prompt=P-MEDIUM warm 1
+[09:50:34Z]   prompt=P-MEDIUM warm 2
+[09:51:34Z]   prompt=P-MEDIUM warm 3
+[09:52:50Z]   prompt=P-HARD cold (max_tokens=2048)
+[09:55:17Z]   prompt=P-HARD warm 1
+[09:57:55Z]   prompt=P-HARD warm 2
+[10:00:21Z]   prompt=P-HARD warm 3
+[10:02:48Z] >>> CELL: ngl6-ctx96k (NGL=6 ctx=98304)
+[10:02:48Z] Stopping llama-server on Predator...
+[10:02:52Z] Starting llama-server (key=ngl6-ctx96k, NGL=6, ctx=98304)...
+[10:02:55Z] Waiting for llama-server (key=ngl6-ctx96k, up to 15 min)...
+[10:03:02Z]   llama-server up after 7s
+[10:03:05Z] Running llama-bench for ngl6-ctx96k (NGL=6)...
+[10:04:19Z] Running prompts for ngl6-ctx96k (cold + 3 warm; max_tokens 512/1024/2048)...
+[10:04:19Z]   prompt=hello cold (max_tokens=512)
+[10:04:36Z]   prompt=hello warm 1
+[10:04:54Z]   prompt=hello warm 2
+[10:05:13Z]   prompt=hello warm 3
+[10:05:30Z]   prompt=P-MEDIUM cold (max_tokens=1024)
+[10:07:21Z]   prompt=P-MEDIUM warm 1
+[10:09:12Z]   prompt=P-MEDIUM warm 2
+[10:11:04Z]   prompt=P-MEDIUM warm 3
+[10:12:10Z]   prompt=P-HARD cold (max_tokens=2048)
+[10:14:42Z]   prompt=P-HARD warm 1
+[10:17:15Z]   prompt=P-HARD warm 2
+[10:18:43Z]   prompt=P-HARD warm 3
+[10:21:15Z] >>> CELL: ngl6-ctx131k (NGL=6 ctx=131072)
+[10:21:15Z] Stopping llama-server on Predator...
+[10:21:19Z] Starting llama-server (key=ngl6-ctx131k, NGL=6, ctx=131072)...
+[10:21:22Z] Waiting for llama-server (key=ngl6-ctx131k, up to 15 min)...
+[10:21:30Z]   llama-server up after 7s
+[10:21:32Z] Running llama-bench for ngl6-ctx131k (NGL=6)...
+[10:23:00Z] Running prompts for ngl6-ctx131k (cold + 3 warm; max_tokens 512/1024/2048)...
+[10:23:00Z]   prompt=hello cold (max_tokens=512)
+[10:23:34Z]   prompt=hello warm 1
+[10:23:52Z]   prompt=hello warm 2
+[10:24:33Z]   prompt=hello warm 3
+[10:25:17Z]   prompt=P-MEDIUM cold (max_tokens=1024)
+[10:27:18Z]   prompt=P-MEDIUM warm 1
+[10:28:31Z]   prompt=P-MEDIUM warm 2
+[10:30:15Z]   prompt=P-MEDIUM warm 3
+[10:31:52Z]   prompt=P-HARD cold (max_tokens=2048)
+[10:34:39Z]   prompt=P-HARD warm 1
+[10:37:05Z]   prompt=P-HARD warm 2
+[10:38:45Z]   prompt=P-HARD warm 3
+[10:41:22Z] Stopping llama-server on Predator...
+[10:41:26Z] === bench complete ===
+[10:41:26Z] Mirroring run-dir to Predator...