laguna-s-2.1 first light — smoke 4/4: openssl-selfsigned-cert (PASS) + BATCH ROLL-UP

Job 4 of four. MinimalPi, K=1, uncapped. This file carries the roll-up for the whole first-light batch.

openssl-selfsigned-cert — PASS, 5m27s total / 4m53s agent, 63223 in / 4870 out

Clean pass, no guard fires. The historical false-success poster child (fixed by preamble RULE 2) stayed fixed on this model.

---

Batch roll-up: laguna-s-2.1 smoke = 3/4

First benchmark run of laguna-s-2.1 (poolside 118B-A8B MoE, UD-Q4_K_S ~68.6 GB, partial offload with experts in system RAM, -c 262144, --reasoning-budget 2000). Harness: MinimalPi (baseline), K=1, no fast-fail caps. Four one-task Harbor jobs, llama-swap restarted before each.

task	result	total	agent	in/out tok	35b agent	ratio
fix-git	FAIL	4m18s	3m44s	76479/3363	18s	12×
regex-log	PASS	22m16s	21m30s	225315/24778	(cut@20m)	—
nginx-request-logging	PASS	7m40s	7m02s	90809/7396	15s	28×
openssl-selfsigned-cert	PASS	5m27s	4m53s	63223/4870	13s	22×

Total wall clock ~40 min for 4 tasks.

Operational findings

1. The RAM ratchet did NOT reproduce — laguna is not the 122b. This was the main risk going in (laguna is the other ~45 GiB shmem-expert model, and the box started with only 10.4 GiB MemAvailable, so a 122b-style ratchet of ~59 GiB in 10–25 min would have OOM-killed the box within minutes). Measured across the full ~40 min batch, MemAvailable was flat: 10.4 → 10.1 → 10.1 → 10.2 → 10.2 GiB, swap untouched. Sampled every 30s throughout. The per-job llama-swap restarts were therefore insurance that turned out not to be needed — worth keeping for unattended runs until a longer session confirms this, but laguna does not need the 122b's one-task-per-job batching discipline.

2. laguna is 12–28× the 35b wall clock on short tasks — much slower than the 122b's 4–7×. Sustained decode ~17–19 tok/s (vs ~165 tok/s for the 35b at f16 KV), plus ~2–3× more output tokens per task. Consequences:

• Stock fast-set caps would cut every laguna pass. TASK_SET_fast_TIMEOUTS caps openssl at 120s (laguna needed 293s) and regex-log at 600s (needed 1290s). A fast run on this model needs a scaled cap map — roughly 10×, not the 122b's 5×.
• Wall clock scales badly: the 89-task suite at this rate is not a casual run.

3. Wire path is clean. Tool calling, reasoning split, and thinking all behaved through the harness's exact path: qwen-named tools registered and used fluently (run_shell_command, read_file, write_file), reasoning arriving in reasoning_content via --reasoning-format deepseek, zero schema errors, zero unknown-tool errors, zero wire-format leaks across all four trials. The documented "empty content" gotcha never appeared — MAX_TOKENS=65536 is far above the ~3000 floor it needs. No guard fired in any trial (0 loop blocks, 0 write rejections, 0 bash timeouts, 0 recovery nudges), and no trial hit a length-stop.

The one failure

fix-git — a genuine near-miss, fully written up in smoke__laguna-s-2.1__20260726-193806/NOTES.md and COMMENTS.md. Short version: it hand-wrote its cherry-pick conflict resolution instead of restoring the file from the git object, and verified git status/git log (process) rather than file content (the graded criterion). The 35b makes the identical mistake and escapes it only by running cat -A and then git show <rev>:<path> > <path>.

Caveats

K=1 on every task — single-trial results, no variance estimate. Uncapped, so these timings are directly comparable to other uncapped runs only.