pine-ai/pine-computer-saasbench-results

Pine Computer — SaaSBench Evaluation Results

Pine Computer · Pine AI

Task-level results and evidence hashes for Pine Computer on SaaS-Bench. Raw traces and screenshots remain private. Separate releases retain their own task versions, selection rules and benchmark suite; they are not pooled.

Latest: SaaS-Bench 1.1 · September 18–19, 2026

78.33% checks passed · 29/106 tasks fully resolved (27.36%) · 106/106 tasks with valid results.

The checkpoint numerator is 1,077 / 1,375 checks. This completed batch had 125 attempts: 106 accepted outcomes and 19 invalid/superseded attempts. Sixteen healthcare tasks were rerun after environment/grader fixes; three other tasks had grading-error retries. Valid failures and eight timeouts remain in the results. This is not a first-attempt-only score. See the methodology for the full selection and comparison caveats.

Domain	Tasks	Checks passed	Fully resolved
Agriculture	12	76.74%	8/12
Business	15	84.26%	3/15
Healthcare	16	68.98%	1/16
Media	20	80.85%	8/20
Software	31	75.85%	8/31
Teamwork	12	86.78%	1/12

Results · Summary · Evidence hashes · Methodology

Release history

Evaluation (UTC)	Suite	Checkpoint	Fully resolved	Results
September 18–19, 2026	SaaS-Bench 1.1 · d38803b44197	78.33%	29/106	106 rows
September 9–10, 2026	Original SaaS-Bench · dc16a4421dd0	76.35%	27/106	106 rows

These are different suites and are not a controlled before/after comparison. Historical release data, methodology and hashes are preserved. The original default dataset configuration is also preserved for existing consumers.

Load results

from datasets import load_dataset

results = load_dataset(
    "pine-ai/pine-computer-saasbench-results",
    name="saas-bench-1.1-2026-09-18",
    split="test",
)

Use name="results-2026-09-09" for the original release. These are evaluation-result records, not the underlying source tasks or execution environments.

Evidence and attribution

Each release's evidence index identifies complete private archives by SHA-256. Publisher checks verify file identity and archive completeness, not independent task success. No third-party outcome verification is claimed.

SaaS-Bench was introduced by UniPat AI. These are Pine Computer results from Pine's integration; upstream task definitions and verifier source are not redistributed here. See the upstream task-data license.

Public results and original documentation are under CC BY 4.0; see LICENSE.md. Attribute Pine AI, Pine Computer — SaaSBench Evaluation Results, the specific release and this repository URL. This license does not grant rights to private raw evidence.