CANONIC

COLD OPEN

Making AI agents creative —
by making creativity verifiable.

Coding models got good because every attempt could be checked against a test suite. Creative work never got its test suite. Canonic builds it: task environments, scoring rubrics, and human-validated judges for the domains everyone said were unscoreable — comedy, screenwriting, songwriting, story.

Training data for frontier labs Context for agents & apps

INT. TRAINING RUN — DAY

Watch a model learn comedy.

Complete, unedited outputs from our proof-of-concept run. Same small open model, same scene brief — the only difference: the second one was trained on a Canonic environment. Scene brief: Michael uses Pam's Post-It notes to avoid work calls and appear busy in his office.

Before — base model 0.19/1.0 judge score
  • explains instead of writing
  • planning notes leak into the script
  • drifts off the given premise
After — trained on our environment 0.67/1.0 judge score
  • on-voice for every character
  • setup → payoff structure lands
  • premise held all the way through
Judge score across training stages for the selected scene. Real, unedited results from our proof-of-concept run on a small open model.

INT. THE EVAL BENCH — CONTINUOUS

Every score decomposes.
Every judge answers to humans.

A signal where there was none

Comedy, story, voice, style — domains that never had a quality signal now have one your training run or your agent can act on.

Judges that answer to humans

Every automated score is benchmarked against working professionals. Agreement is published per environment — including where the judge can't be trusted.

Improvement you can see

Not a claim — a receipt. Before-and-after outputs with numbers attached, like the scenes above: same model, same brief, visibly better work.

One output, decomposed across eight rubric dimensions. Illustrative rendering.

INT. FRONTIER LAB — DAY

Training data for frontier AI labs

Training environments for the domains without unit tests. You get a reward signal for creative quality, evidence it deserves your trust, and a package that drops into the training stack you already run.

  • Training-ready creative environments — drop-in task suites for comedy, story, and style; integration takes an afternoon, not a quarter.
  • Reward signals you can interrogate — per-dimension scores, not a single opaque number.
  • Published validation stats — per-environment agreement with working professionals, limits disclosed up front.
  • Evidence it holds up — every environment ships with a report showing the score survives a model actively trying to game it.
Reward progression and score-distribution shift from the proof-of-concept run (stylized).
Sitcom writing · live Screenwriting Songwriting Stand-up Books Comics Memes
Request environment access

EXT. MANGALORE — GOLDEN HOUR

Context for agents & apps

We asked a state-of-the-art video model for the same shot of Mangalore — a coastal city of half a million people — twice. Once with the prompt a typical user types. Once with the reference profile a Canonic-connected agent sends instead.

what a user types

“Drone view of Mangalore city with a bridge, fishing boats and coconut trees, golden hour.”

model output — drop in assets/mangalore-naive.mp4

A bridge from no particular country. Boats borrowed from a Mediterranean postcard. Palm trees doing an impression of Bali.

  • plausible — and wrong everywhere it matters
what a Canonic-connected agent sends

“…the four-lane bridge crossing the wide Netravati at Ullal, wooden fishing trawlers with painted hulls crowding the Old Bunder, red laterite compound walls, Mangalore-tiled roofs, coconut and areca groves…” (full reference profile continues — hundreds of verified details)

model output — drop in assets/mangalore-grounded.mp4

The same model. The only change is the context it was handed.

  • grounded in curated, verified reference material

A real test, reproducible with the two prompts as written. Both clips are unedited video-model outputs.

Reference Profiles give your agent deep, verified context for a style or a place — what makes a thing itself — plus a verifier that scores whether an output actually got there. Delivered to your stack via API and MCP.

Get early access
8rubric dimensions per score
3.5×judge-score lift in our proof-of-concept run
TV writersjudge validation panel — in progress
Soonpublic creative benchmark — launching

INT. A QUIETER ROOM — LATER

Investors

Frontier labs now spend over a billion dollars a year on reinforcement-learning environments. Almost none of it can reach the creative domains, because nobody has made them verifiable. That's the layer we're building — starting where the data is richest and the gap is widest.

Request the memo