Skip to content

Laguna: actionable thinking guidance + setup verification (follow-through on #2)#4

Merged
TheTom merged 1 commit into
mainfrom
defilan-field-notes-followup
Jul 25, 2026
Merged

Laguna: actionable thinking guidance + setup verification (follow-through on #2)#4
TheTom merged 1 commit into
mainfrom
defilan-field-notes-followup

Conversation

@TheTom

@TheTom TheTom commented Jul 25, 2026

Copy link
Copy Markdown
Owner

Completes the response to @Defilan's field notes in #2. The substantive corrections from that issue (§5e, the §2 enable_thinking fix, the §5b ablation, and the cross-model note in patterns.md) already landed on main. This PR is the part that was still missing for a reader.

The problem this fixes

Correcting enable_thinking had a knock-on I didn't handle at the time: it left the guide's top recommendation without a usable instruction. "Thinking OFF" was recommendation #1, and the setting everyone reaches for to do it turns out to be a no-op. A reader following the guide literally could not perform step one.

What's here

  • "How to actually minimize thinking", ranked by evidence: never send enable_thinking: true (the one setting that provably changes the template), use a named professional persona (best-evidenced lever, three stacks, 6/6 vs 0/5 single-turn), expect attenuation not elimination in long agentic loops, and cap your own tokens as the backstop. Carries the honest caveat that whether this stops the model reasoning or only stops reasoning being parsed is unresolved and stack-dependent.
  • "Verify your setup (worth 5 minutes)", four checks for the traps that are all silent: tool calls actually parse (the 83% to 0% cliff), -fa on, what your stack does with reasoning_content, and revision pinned + capped. That third check exists precisely because all three known stacks behave differently there, so the guide now tells readers to measure their own rather than trust any published rate, ours included.
  • Quickstart says three stacks, not two.
  • Card synced. Its thinking and serving rows were still asserting things the guide no longer claims. PNG re-rendered.

Why the verification block earns its space

Nearly every problem reported with this model across three independent stacks has turned out to be configuration rather than capability, and every one of those failure modes is silent: chatml degrades tool calls to prose without erroring, flash-attention-off just looks like a slow model, enable_thinking: false looks like it worked, and config drift looks like the model regressed. Five minutes of checks up front is cheaper than any of those diagnoses.

Closes #2.

…fication block

Follow-through on offlabel#2. The §5e/§2 corrections already landed on main; this
is the part that was still missing for a reader.

Correcting enable_thinking left our TOP recommendation ('thinking OFF') without a
usable instruction, since the setting people reach for is a no-op. Added a ranked
'how to actually minimize thinking': never send true, use a named persona (the
best-evidenced lever, 3 stacks), expect attenuation not elimination in long loops,
cap your own tokens as backstop. With the honest caveat that whether this stops
reasoning or only stops it being parsed is unresolved and stack-dependent.

Added 'Verify your setup (worth 5 minutes)': four checks for the silent traps
(tool calls actually parse, -fa on, what YOUR stack does with reasoning_content,
revision pinned + capped). The reasoning_content check exists specifically because
all three known stacks differ, so readers should measure their own rather than
trust any published rate including ours.

Quickstart now says three stacks, not two. Card synced to match (thinking row and
serving row were stating things the guide no longer claims) and PNG re-rendered.
@TheTom
TheTom merged commit afa6831 into main Jul 25, 2026
1 check passed
@TheTom
TheTom deleted the defilan-field-notes-followup branch July 25, 2026 22:24
TheTom added a commit that referenced this pull request Jul 26, 2026
I merged a fourth stack's findings straight to main without review. Backing that
out: it is the weakest-sourced item in the set (relayed via screenshot, no
published raw data, unlike BlackwellBoy's lab repo and Defilan's harness) and it
CONTESTS this guide's headline recommendation, which is exactly the kind of claim
that should be reviewed and independently tested before it lands.

Removed from main pending #8: the regime-split thinking claim (+2.6 pts on
single-turn codegen), the ~12k budget / cap-hits-are-degeneration-loops guidance,
the related blind-spot chip, and the fourth-stack credit. Quickstart #1 and #4
and the card thinking row revert to the reviewed (#7) wording, with a pointer to
the open PR.

Retained on main, because they came from contributors with published data who are
also the reviewers: the dose-curve reframe of the firing gate, the closed open
question, the truncated-think failure mode, and the gfx1151 serving config.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Laguna S 2.1: field notes from a third stack (gfx1151/llama.cpp) - enable_thinking inertness, persona gate in an agentic loop, §5b rule ablation

1 participant