Same prompts
Models in one phase receive exactly the same prompt text and resource limits. Prompt versions increase instead of overwriting the original, and comparison stays within one task version.
PROVENANCE FIRST · METHOD
Methodology and boundaries
This site presents traceable archive records. Missing values are not zero, and incompatible measurements are not collapsed into a leaderboard.
What is kept here is not a claim about which model is strongest, but a set of works and run records that can be checked: same prompts, real outputs, independent reviews, verifiable sources.
METHOD PRINCIPLES
Five principles
Real outputs
Pages show the files the models actually delivered: pages can be opened, SVGs can be inspected, and run records keep the call and retry counts.
Independent reviews
Human reviews, each AI review version and this website’s own checks are recorded separately; scores are shown side by side with their method and disclaimer, never silently merged.
No aggregate ranking
Scores from different reviewers and rubrics are never merged into a leaderboard; unknown values stay "Not recorded" instead of being inferred.
Provenance first
Every prompt, metric and review keeps its file path, line location and archive commit, and public links are pinned to that commit so each claim can be re-checked.
How historical conclusions are presented
Every phase assessment carries its source and method: the DeepSeek 93.6/100 is quoted from a historical AI report, while the Muse Spark 84.9/100 comes from the maintenance agent non-blind v1 review. The two use different rubrics and must not be compared directly; neither is an independent proof by this website. For tasks 06 and 07 the human judgment differs from the AI assessment, so both are retained side by side.
Boundaries and isolation
Isolation and boundary information is shown as archived: a policy-level constraint discourages mistakes and can be reviewed afterwards, but it is not enforced isolation, and this site will not call it "audited".
Authoritative field definitions live in the repository docs (result-interface, data-model, audit-method); this page only describes the display rules.
BATCH MEASUREMENTS
Batch metrics
Per-task records
DeepSeek V4.1 Flash Exp 0910 · Phase 1
Tasks15
Sessions12
Total duration2:18:02
API calls693
Tool calls735
Failures / retries42
Input tokens334,670
Output tokens1,072,534
Cache read63,042,176
Cost¥9.3
Muse Spark 1.3 · Phase 1
Tasks15
Sessions15
Total duration25:45
API calls181
Tool calls178
Failures / retries8
Input tokens384,338
Output tokens154,563
Cache read3,130,484
Cost¥0
DeepSeek V4.1 Flash Exp 0910 · Phase 2 (Task 16–20, partial)
Tasks5
Sessions5
Total duration3:06:16
API calls927
Tool calls940
Failures / retries0
Input tokens457,656
Output tokens1,475,485
Cache read234,301,952
CostNot recorded
Muse Spark 1.3 · Phase 2 (Task 16–20, partial)
Tasks5
Sessions5
Total duration45:02
API callsNot recorded
Tool calls238
Failures / retries4
Input tokens451,558
Output tokens146,779
Cache read20,144,127
CostNot recorded
GPT-5.6 Sol · Phase 1
Tasks15
Sessions15
Total duration1:16:46
API calls120
Tool calls76
Failures / retries1
Input tokens4,187,022
Output tokens129,130
Cache read3,650,560
CostNot recorded
K3 · Phase 1
Tasks15
Sessions15
Total duration3:16:10
API calls250
Tool calls259
Failures / retries12
Input tokens521,100
Output tokens370,144
Cache read13,399,744
CostNot recorded