| 📌 | Rank | Model | Maker | Weights | Context | Price in / out | Our score | Coding ELO | SWE-bench | Intel (AA) | Agentic | Details |
|---|
Our scores
Ground-truth measurements scored across 6 harness axes: Form Contract (first-line conformance, no fences), Coder Task (executable code pass), Judge Form (verdict format integrity), Availability (HTTP 200 stability), Bakeoff (empirical task match), and Consistency (streak reproducibility). A clean contract pass yields 74/100, leaving headroom for stress testing.
| Model | Manufacturer | Awarded Score | Axes | Form Contract | Coder Task | Judge Form | Availability | Bakeoff | Consistency | Cautious Fill | High Fill |
|---|
Manufacturer Constructors Championship
Aggregated evaluation benchmarks across 59 engine families and 39 manufacturers from FAMILY.md and MAKER.md. Compares unweighted family composite means and model-weighted organizational averages.
| Manufacturer | Engines Tested | Families Tested | Unweighted Composite | Weighted Fleet Mean | Flagship Ace | Weakest Link |
|---|
Next Lap Projections (High-Fill Potential)
Prioritized model ranking from FORWARD.md measuring the gap to high fill. Models with the largest point gap represent high-potential candidates where empirical testing on unmeasured dimensions could yield the highest score breakthroughs.
| Gap (Points) | Model Name | Manufacturer | Family | Current Score | Potential Headroom |
|---|
Pair tests
Measured multi-model disagreement matrix from automated test gates: \(\vec{v}_{ij} = \text{freq}(i, j) \times \text{error\_magnitude}\). High orthogonality denotes complementary models that fail on different inputs, maximizing multi-agent verification and error discovery per dollar.
| Model A | Model B | Axis | Pairs | Disagreement % | Mean Hole Severity | Orthogonality (V) | Rescue P(B|A) | Rescue P(A|B) | Mutual Co-fail | Signed Complement |
|---|
Seats
Models seated as HEROes with 8 verified layers (L1 Role, L2 Model, L3 Harness, L4 Wrapper, L5 Skills, L6 Tools, L7 Environment, L8 Mission) and confirmed operational scars.
| Seat Slug | Pinned SKU / Model | Harness | Contract | Seated Status | HERO Full | Host OK | Fib OK | HTTP | Active Scar |
|---|
API
Authoritative CDN-cacheable static REST endpoints served from mdata.modelraters.com (and 301 alias mdata.mdrater.com):
👋 New to JSON APIs? Explore the Interactive Playground
Don't let raw JSON brackets scare you! Learn how to pull live model scores into Python, JavaScript, Google Sheets, or terminal with 1-click recipes.
/v1/models.json
Complete model catalog dictionary keyed by canonical dossier ID, containing parameters, architecture, pricing, and benchmark scores.
Inspect JSON →/v1/leaderboard.json
Ranked leaderboard array joining empirical 100-point ratings, LMSYS Arena ELO, SWE-bench Verified, and throughput.
Inspect JSON →/v1/ratings.json
All 138 empirical 100-point model evaluations with granular 6-axis breakdowns and fill projections.
Inspect JSON →/v1/porosity.json
Pairwise orthogonal porosity matrix, disagreement frequencies, rescue probabilities, and signed complement tensors.
Inspect JSON →/v1/forward_projections.json
Prioritized model testing backlog ranking models by high-fill potential gap.
Inspect JSON →Full Relational Database: mdata.sqlite
The database holds the model records, scores, and sources.
Download mdata.sqlite ↓