Canonic
AI models render the long tail as a postcard. Canonic measures it, and fixes it.
Trades, regional cuisine, dance, crafts, rituals, scripts, neighbourhoods: the long tail of concepts that image, video and language models flatten into something generic. Canonic finds those concepts automatically, builds its own verified descriptions and exemplars from public footage, measures where each model fails, and sells the evals, the data and the context that fix it.
Same model, same shot of Mangalore
We asked a state-of-the-art video model for a drone shot of Mangalore twice: once with the prompt a typical user types, once with the reference context Canonic supplies.
Generative models are fluent about the average and vague about the specific.
Ask for a chai stall, a regional dance or a festival street and you get a stock-photo version of it. The model has seen too little of the real thing, and nobody is measuring the gap.
Canonic makes that gap measurable, concept by concept and model by model.
How it works
Find the gaps
Mine and rank long-tail concepts that models render generically, starting from a Wikidata-linked inventory.
Build references
Capture frames from openly licensed public footage and label each concept with an attribute checklist.
Generate and verify
Make exemplars on Apache-2.0 models only, then filter near-duplicates and check each one against its checklist.
Measure
Run frozen evals on a private held-out pool, so the scores mean the same thing every time.
What makes it different
Rigour carried over
Frozen rubrics, hashed task sets, judge ensembles and confidence intervals, inherited from the Stage 1 benchmark.
Clean provenance
Source frames stay private and are never published. Shipped exemplars come only from Apache-2.0 models.
Held-out evals
Eval items come from a private pool the models have not seen, so a score cannot be trained into.
Where it stands
Benchmark live, as of October 2026.
Built
- Stage 1: a six-domain creative-writing benchmark, frozen and live as proof of rigour.
- The concept pipeline scaffold and the 36,627-concept seed inventory.
Now
- Gap mining and the first wave of concepts.
Next
- A showcase of pieces made with Canonic context, and inquiries from studios and labs.