Pálpito para 30X

No. Without a key, Pálpito runs on the local engine at no cost and states this in the top right of the interface. Supplying keys and the Jev mode in .env connects it to Jev instead.

Jev answers typed questions in one call: role, pain, goal, intent, and whether the profile is genuine. The policy then combines those answers into a program, an action, and a route, recording the reason for each decision.

A non-genuine profile never receives an automated action, the LLM never authorizes a handoff to sales, and no message leaves without review. The architecture notes are in docs/ARQUITECTURA.md.

All bundled data is generated and does not correspond to real people, with addresses in example.com. It includes synthetic profiles and events, discovery sign-ups and a later batch with a new group, four cohorts of student attendance and activity, and aggregated distribution counts. 30X files are gitignored.

evaluar.py measures per-question accuracy, automation, improper passes, latency, and cost per 1,000; evaluar_dinamica.py measures reclassification from events; calibrar.py tests 180 threshold combinations; evaluar_guardrail.py scores labelled messages; evaluar_descubrimiento.py reports ARI, stability, drift, precision, and coverage. Outputs land in reportes/.

The repository sits in the blockstars-labs GitHub organization, described as a challenge deliverable with a five-page playbook, an architecture document, and a presentation. The live prototype is published at palpito-30x.emanuel-benavides.workers.dev.