japepo

japepo (“pot” in Guaraní) tracks MasterChef Celebrity Paraguay 2026, which began on 10 August 2026 on SNT. When the page was first built there had been one gala and no elimination, so its job at that point was not to say who wins but to say how much a single gala can support, and to put a number on that.

What it does not do

• It is not affiliated with MasterChef, Banijay or SNT, and no one from the production reviewed it. Named marks belong to their owners and are named only to say what the analysis is about; the page’s own mark is a pot seen from above, not the show’s logo.
• It publishes no photographs of contestants. Each of the eighteen is shown through a figure generated from their name.
• It makes no request to any other server and does not track visitors by opening the page. Embedded tweets are read on the page itself, and the original opens only if the reader clicks through.
• It does not write a number by hand: every figure on the page comes out of the model and is filled in by gui/build.py, and gui/verificar.py rejects a decimal typed directly into the prose.
• It measures attention from view counts on eighteen contestant-introduction videos, not approval, and it does not publish a measurement with a gap in it: on 30 August one video failed to return a count, and rather than show that contestant as last of eighteen, the run is now retried until every count is present.
• Its Guaraní text was machine-written and has not been checked by a speaker.

Install and run

Python 3.12 with numpy, fonttools and Pillow, and Node for one of the checks. The model runs in about a minute and stays under 150 MB.

bin/publicar.sh                 # the whole path: recompute, verify, publish
bin/publicar.sh --gala 5        # also freezes the prediction for gala 5
bin/publicar.sh --social        # remeasures views and comments
bin/publicar.sh --sin-empujar   # commits but does not push

Publishing stops if any of nineteen checks fails, which is what lets it run unattended. The steps run separately when needed:

python3 model/preparacion.py                     # model -> data/estadisticas.json
python3 model/registrar.py --fecha AAAA-MM-DD --gala N
python3 model/puntaje.py --gala N                # scores a resolved gala
python3 model/social.py --fecha AAAA-MM-DD       # views and comments -> data/social.json
python3 gui/build.py                             # data/ -> web/index.html
python3 gui/verificar.py                         # the gate

How it works

Every run freezes three things into data/historial_pronostico.json that are never rewritten: the day’s snapshot, the promise for the next gala, and the score once that gala resolves. model/puntaje.py refuses to score a gala that had no prediction published beforehand. The first scored prediction, published 18 August at 07:31 UTC in commit de296f3, sixteen hours ahead of air, put one contestant second of eighteen among those at risk; she was the one eliminated, and the model’s Brier score (0.921) beat the uniform baseline (0.944) on that single gala, which the repository says is not enough evidence to mean much on its own. EVALUACION.md fixes the scoring rule, multiclass Brier against a uniform baseline plus log-loss on the winner, before the season ends.

Live page: https://nerln.github.io/japepo/

japepo on GitHub. Code is Apache-2.0 licensed; text and data are CC BY 4.0.