SQE Leaderboard/Laguna S 2.1

Laguna S 2.1

poolside/Laguna-S-2.1-FP8

fourlineVendor FP8 SWE-bench Pro 59.4% official2026-08-06
Methodology

SQE score

96

blended mean, 5 reps

Gate

5/5

all reps counted

Wilson 95% low

≥56.6%

lower confidence bound

Shots to green

1

median of 5 reps

Every published metric

SQE score
Headline blended mean percent across reps (gate 45 / pass 25 / repair 15 / style 10 / thrift 5).
96n=5
Gate
Reps that cleared known and held-out honesty checks, shown as k/n after tainted exclusions.
5/5n=5
Pass
Reps that cleared public verification (compile, known conformance, scenario threshold), as k/n. Distinct from gate, which also requires held-out honesty.
5/5n=5
Wilson 95%
Lower bound of the gate pass rate at 95% confidence.
≥56.6%n=5
Harness ±
Blend of all score components, mean ± spread.
96.5 ±1.3n=5
Shots to green
Feedback rounds needed before tests pass. Lower is better.
1n=5
Scenario pass
Share of graded scenarios passing in the green build.
100%n=5
Trap/spec rate
Share of trap parameter values correctly synthesized from the spec.
100%n=5
1-shot gap
How much the first shot trailed the final result.
-n=5
Effective n
Reps counted after tainted-run exclusion.
5/5n=5
Median tokens / rep
Prompt + completion tokens per rep, median over untainted reps. Lower is better.
4.04M tokn=5 untainted
Min VRAM
Catalog minimum cluster VRAM (GB) from syndicAI referenceGpuRequirements. Closed APIs and packs not in the catalog show as -.
152 GB
SWE-bench Pro
Catalog SWE-bench Pro score for the resolved syndicAI model, or the vendor-published score for closed-reference models. Inferred scores are labelled. Packs without a score show as -.
59.4%

Run configuration

Endpoint model
poolside/Laguna-S-2.1-FP8
SWE-bench Pro
59.4% · SWE-bench Pro · Poolside (Laguna S 2.1)
Reps
5
Max shots
3
Feedback tier
1
Context cap
200K
Token budget / run
4M
Turn cap / shot
120
Thinking budget
4.096K
Sandbox digest
quality-eval-sandbox@sha256:242ebd2f73c1c1ffa90c3ccb7913574f91e74a89dd1e4cf874f907a3129149c5

Review

Generated by claude-haiku-4-5

Performance Review: poolside/Laguna-S-2.1-FP8 on fourline task

Overall: Excellent end-to-end delivery. All 5 runs passed full gate (correctness, overfit, compile, lint, all scenarios, held-out conformance). Single-shot success on every run; zero flakiness. Quality is consistent across reps, with blended score 96% and effective token thrift 0.32–0.38.

Task completion was clean: the model correctly implemented a Connect-Four variant with immutable game logic, DOM harness integration, keyboard controls, and test-mode harness API. Spec traps (board geometry, starting player) were satisfied in all runs. One early attempt to call finish was correctly rejected; the model understood and continued working.

Strengths:

  • Perfect gate pass rate (5/5 correct, no overfitting) across all scenarios including held-out validation.
  • First-shot completion. No repair loop; model reached green in a single attempt on every run.
  • Consistent output quality across varying token budgets (39k–69k completions) with clean compilation and linting both times checked.
  • Guidelines adherence: Model read GUIDELINES early, self-initiated typecheck/lint in all runs, and used immutable patterns throughout.
  • Tool discipline: Proactive reading of SPEC and GUIDELINES; careful incremental implementation (many turns, small steps) rather than bulk overwrite.
  • No scenario-specific regressions. Trap observations (8 cols, 7 rows, Yellow starts) match spec exactly in all 5 runs.

Failure modes:

  • Tool errors present but low impact. Runs r1, r4, r5 recorded 1–9 edit_file errors each (total 17 across 5 runs = 3.4 per run average); r2 had zero. Errors were transient mismatches in old_string (file drift during multi-turn edits). No errors blocked correctness: all runs compiled and passed verification. Likely due to aggressive sequential editing; no architectural or logic failure.
  • Early finish rejection (r5). Model attempted to call finish at turn 92, harness rejected it. Model persisted, continued for 6 more turns, and delivered green. No evidence of confusion; straightforward persistence in the face of expected rejection.

Serving & harness notes:

  • Context efficient. Prompt tokens ranged 3.7M–4.0M across runs; all stayed well under budget (0% token budget exhaustion risk until late stages). Turn counts 92–98 suggest the model was working deliberately, not hitting hard limits.
  • No stalls. Wall-clock runtimes 711s–1490s proportional to token volume and shell operations (e.g., Playwright install in r1 at turn 84 added ~78k tokens). No unexplained delays.
  • Zero reasoning-only turns. Every turn produced either a file edit, read, shell command, or API call; no vacuous tool invocations.
  • Harness integration smooth. Test-mode API (__fourline with drop(), reset(), getState()) and keyboard/UI controls both working in final verification. Scenarios exercised DOM, conformance replay, and trap probes; all passed.
  • FP8 quantization: No evidence of instability, hallucination, or precision loss. The model's tool-use errors and minor edit mismatches are consistent with high-turn multi-file work under live state churn, not quantization artifacts. FP8 performance is effectively full-precision here.

Summary: A high-quality, reliable solve. The model shipped correct code on the first attempt, every time, with good token efficiency and no conceptual missteps. Minor tool-use friction is a minor hazard in live editing workflows but had zero impact on final correctness.

Per-rep results