laguna §5d: correct the soak counts, including a turn-success figure that was wrong at source

[Blackwellboy]

Drafted with AI assistance (Claude); reviewed and submitted by @Blackwellboy.

Counts only. No depth-collapse language is touched here; #19 owns that reconciliation and this PR deliberately does not overlap it.

All values below were re-derived from the published artifacts before editing, not transcribed from our earlier message.

Two errors, both ours

1. Checkpoint, not final cut. The ~389 sessions / ~2,947 turns figures come from an ~11.5h checkpoint of a run that finished at 409 sessions / 3,096 turns. Same single run, later cut. You already had the corrected turn count sitting in your own changelog, which is how the document disagreed with itself.

2. The 99.9% turn-success figure was wrong in our primary, and you inherited it in good faith. soak/logs/turns.jsonl has 3,099 lines. Three of those are kind: integrity_probe records, not turns: a different shape, status: 200, http_error: false, and the same three records as integrity_probes.jsonl. Every note we published described them as incomplete in-flight turns and derived 99.9% from them.

All 3,096 turn records returned HTTP 200. The turn log in soak_driver.py is written unconditionally with whatever status came back, so a failed turn would appear with a non-200 status. None does. Turn success is 100%.

Corrected at source first, so this PR cites a primary that agrees rather than one that contradicts it: laguna-s21-lab@4507826.

The precise distinctions

quantity	value
sessions	409 (400 completed, 9 killed at the driver session-token cap)
lines in turns.jsonl	3,099
turn records	3,096, all HTTP 200
kind: integrity_probe records in the same file	3
turns with thinking fired	3
thinking rate	3 / 3,096 = 0.0969%, reasonably shown as ~0.1%

The 9 killed sessions are the same 9 incidents your 2026-07-25 entry already scopes as his driver cap firing by design. They were never subtracted from our headline, which is why 409 was reported as though all of them completed.

Six surfaces, not four

Our earlier four-line correction list was incomplete. It gave you §5d prose, the stability bullet, the changelog entry and the card, and it missed the thinking-rate denominator, ~3 of 2,944 turns. That line is the one where a wrong denominator actually moves a published rate, so it was the worst one to leave out.

A sixth surface turned up while making the edit and is corrected here too: the verdict: line in this guide own frontmatter also carried 99.9% turn success. It was in neither of our lists.

1. models/laguna-s-2.1.md §5d prose: ~389 sessions and ~2,947 turns to 409 sessions and 3,096 turns
2. models/laguna-s-2.1.md stability bullet: 2,944 of 2,947 turns succeeded (99.9%) to all 3,096 turns returned HTTP 200 (100%)
3. models/laguna-s-2.1.md thinking rate: ~3 of 2,944 turns (~0.1%) to 3 of 3,096 turns (0.0969%, ~0.1%)
4. models/laguna-s-2.1.md changelog 2026-07-24 entry: turn count and 99.9% to 100%
5. cards/laguna-s-2.1.html: 12h / 2947 turns, 99.9% turn success to 12h / 3096 turns, 100% turn success
6. models/laguna-s-2.1.md frontmatter verdict:: 99.9% turn success to 100%

Plus a dated 2026-07-29 changelog entry recording the correction and pointing at the primary, following your existing convention.

Checks

python3 scripts/check_guides.py passes: valid frontmatter, required keys, no em/en-dashes. git diff --check clean.

One line I did NOT touch, deliberately

The 2026-07-26 changelog entry says his soak sent enable_thinking: true on all 3,099 turns. Strictly that should be 3,096, since 3,099 is the line count and 3 of those lines are probes rather than turns. It sits inside an entry that #19 is reconciling, so it is yours to fold in whichever way suits that PR rather than something to fix from two directions at once. Flagging it so it is not missed.

Nothing about the thinking-gate, persona-gate or depth-collapse conclusions changes. Only counts.