Laguna S 2.1: make the guide answer "what should I do", + @BlackwellBoy's 12h soak

[TheTom]

Two things, and the second one reshaped the first.

@BlackwellBoy ran Laguna S 2.1 for 12 hours inside his own production agent pipeline on a different quant and a different serving stack than the held-out testing in this guide: 389 sessions, 2,947 turns, thinking enabled throughout, revision pinned, one DGX Spark, hard output-token cap. Independently, he landed on the same operating manual this guide had converged on. That is worth more than either of us saying it alone.

Reading his writeup next to ours made it obvious that this guide was organized by provenance (who measured what) instead of by what a user should actually do. So this PR fixes that too.

For users: what changed

• New Quickstart: the four things. Thinking off · native poolside_v1 tool format · integrity clause in the system prompt · pin the revision and set your own max-token ceiling. One table at the top, with links to the evidence for each. Two testers, two stacks, same four.
• The cheat sheet was missing two of the four. It never told you to pin a revision or set a token cap. The cap matters: the vendor dropped max_new_tokens post-release, so if you do not impose a ceiling, nothing does. That is the actual mechanism behind the community "it never stops" reports.
• Throughput and timeout planning added. Expect ~20 tok/s and ~10-25s per agent turn on a single Spark. Included as a warning because we wasted a run on it: a 20-minute-per-iteration cap on a heavyweight generic harness produced zero files written. Harness-sizing failure, not a model failure, and an easy trap for anyone self-hosting.
• Two practical prompting levers added. (1) A named professional persona is a second thinking-off switch when your harness does not expose the toggle. (2) Give explicit acceptance criteria, not a vibe — it builds to whatever it can check, so "World 1-1" got four platforms and one Goomba. Highest-leverage prompting change for this model.
• Verdict rewritten as configured-right vs configured-wrong, because that is the honest shape of this model.

What his soak adds (§5d)

• It is stable enough to leave running: 2,944/2,947 turns (99.9%), zero crashes, zero restarts, ~4 GiB memory creep over 12h, ~13.5s/turn. This is the strongest evidence available that configured correctly it is a production-grade agent driver on one box, and it is data the held-out testing could not produce.
• Native tool format is near-perfect at scale: 100% tool-call success over ~11.5h, which extends our small-sample 83%-native-vs-0%-generic. Combined: the format choice is the difference between flawless and non-functional. No middle.
• Pin + cap kills the runaway loops: zero in 12h.
• The §5b integrity clause replicated cross-quant: 3 disguised cover-up requests, all 3 refused with reasoning, on a different quant and serving setup. Cross-quant replication was the gap §5b explicitly named, so this is now the best-evidenced recommendation in the guide.

The bit I find most interesting

His pipeline saw thinking fire on 3 of 2,944 turns (~0.1%). Ours saw 5-18% under a bare "helpful assistant" persona. Not a contradiction: our persona×task gate predicts his number, because a real agent pipeline carries both suppressors at once (a named professional identity and coding-shaped tasks). Lab mechanism predicts production behavior; production confirms the mechanism. Practical upshot for users: if you run a normal agent system prompt you are already getting almost no thinking, so turning it off explicitly costs nothing and removes the tail risk.

Attribution

Only his own measurements are counted as independent replication. His writeup also covers the thinking-hurts dose-response and the provenance caves, but those cite the testing in this guide rather than re-measuring them, so they are not double-counted here.

@Blackwellboy thank you for the soak, genuinely useful work and hard to come by. Corrections welcome: if any number or framing here misstates your run, say so and I will fix it. Happy to add your GitHub handle for a real mention, and happy to link your thread directly if you want it cited.

[Blackwellboy]

Green light from me — merge it. Organizing by what a user should do instead of by
who measured what is the right call, and §5d represents the soak accurately.

Two precision points so my data doesn't get oversold:

1. The ~0.1% thinking routing rate is an observed routing number, not a proven
   router property. The API returned empty reasoning fields even when
   thinking was explicitly requested, so on this rev part of that rate may be
   template/parser-level rather than the model declining to think. I'd phrase it
   as "the gate observably almost never opens on this rev and serving path."

2. "Zero loops" is scoped: zero unbounded-generation loops while the thinking
   gate barely opened. The 9 incidents in my logs were all my own driver's
   session-token cap firing by design. It is not evidence about thinking-heavy
   workloads.

Raw logs and harness are now live, replacing "publishing shortly":
https://github.com/Blackwellboy/laguna-s21-lab — per-turn logs (3,099 turns),
incident log, integrity probes, the 20-cell sweep with every losing cell,
container recipe with pinned digests, and the bench harness.

[TheTom]

will adjust language

[TheTom]

Both corrections applied, thank you, they were right and one of them was me overreaching.

1. The ~0.1% routing rate. You're correct and I had the causality backwards. I wrote that your number "confirmed" our persona×task gate, which oversold your data to support my mechanism, exactly the config-artifact mistake I try to hold myself to elsewhere. Now split apart in §5d:

• Ours supports a model-behavior claim, and only because we checked the mechanism separately: the template hands it an opened <think> every time (verified byte-identical via /apply-template) and it still emits nothing under a named persona.
• Yours supports the operational claim, in your phrasing: the gate observably almost never opens on this rev and serving path. With your note that empty reasoning came back even when thinking was requested, so part of it may be template/parser-level.
• Practical advice is unchanged either way, which is why it's still worth acting on. Just no longer cited as proof of why.

2. "Zero loops" scoped. Now reads: zero unbounded-generation loops while the gate barely opened, your 9 incidents were your own driver's session cap firing by design, and explicitly not evidence about thinking-heavy workloads. The takeaway is "cap your own tokens and you won't get unbounded generation," not "pin+cap makes thinking safe."

Two other things came out of your comment:

• Your stack is now stated precisely: NVFP4 on vLLM vs our Q4_K_M on llama.cpp. That makes the cross-stack replication of the integrity clause stronger than I first wrote it, different quant and different engine, so I fixed that too.
• Linked your lab as the source of record: https://github.com/Blackwellboy/laguna-s21-lab. Publishing the losing sweep cells and the incident log is the part I respect most; it's what let you catch my overreach in the first place.

Also folded in a third data point while I was in there: an independent Strix Halo test found -fa on is worth up to 2.2x decode at 56K depth (9.33 to 20.73 tok/s). Our runs had it on from the start, so I've flagged that every throughput number in the guide is a flash-attention-ON number.

Changelog entry records both corrections and attributes them to you, so the walk-back is on the record rather than quietly edited.