{"data":{"kind":"file","path":"README.md","version_id":"q8lczzbrejfxe0t7iqmlpvwf","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":3158,"modified_at":"2026-11-09T09:35:27.663000","content_hash":"4c089973f982a132b1a5a8803f24d773fb1506a65646e3ed140fe1cc67d94385"},"entries":[],"content":"# navital-upi\n\n**UPI Desk**, one of six worlds in [Navital Gym](https://github.com/Navitallabs/navital-gym) v1.1.0, as a verifiers environment. By [Navital Labs](https://navitallabs.com).\n\n- **What:** 50 public tasks. An agent handles UPI payment attempts and disputes in a sandboxed bank under written, simplified rules: pay and verify, limits, failed-but-debited, wrong recipient, collect requests (phased out 28 Feb 2026), mandates, UPI Lite. The world has 18 tools; creating a mandate always fails by design (registration needs the owner's own PIN).\n- **Grading:** deterministic rules, never an AI judge. Reward = the world's own grader (score = fraction of rules passed, in [0, 1]) run on the last JSON object in the final message and on the tool calls the rollout really executed. `solved` (all required rules passed, no forbidden step) is a second metric. A forbidden step counts even if the simulated bank refused it. 29 of the 50 tasks have a forbidden step.\n- **Hidden fields** (expected answers, traps) never reach the agent or the traces, as long as the agent only has the world's tools; they are read host-side at scoring time. A runtime that gives the agent a shell or file access (for example a hosted sandbox with a shell) has not been tested.\n- **Public only:** these are the same public tasks as the open Gym. Models may have seen them. Our private held-out sets are NOT in this package.\n- **Synthetic data.** Every world is a simplified, documented version of Indian practice, not a statement about current law, bank or NPCI rules. Not affiliated with NPCI or any bank.\n\n## Use\n```\nprime env install navitallabs/navital-upi          # once published\nvf-eval navital-upi -m <model> -n 50\n```\n```python\nimport verifiers.v1 as vf\nfrom navital_upi import load_environment\nenv = load_environment()                 # all 50 tasks; ids=[\"IN-UPI-001\"] for a subset\n```\n**Runtime (read this):** the per-rollout tool server needs `navital-gym` on the machine that runs it, so use a local runtime, for example:\n```\nvf-eval navital-upi -m <model> --env.agent.runtime.type subprocess --env.agent.harness.id null\n```\nPrime's default hosted sandbox runtime (a bare python:3.11-slim image) has NOT been tested with this package and is not expected to work without navital-gym in the image.\n\nNeeds `navital-gym==1.1.0` importable (it is a dependency). The per-rollout tool server runs as a local subprocess (`python -m navital_gym.adapters.verifiers_env`).\n\n## Honest limits\nWritten against verifiers 0.4.0's v1 API (pinned), which is young. The adapter has been tested offline with a scripted model and by a reward-parity test (reward and `solved` equal the Gym's own grader on 250 recorded real rollouts, 5 models x 50 tasks, 0 mismatches; a control that drops the tool results changes the reward on 36 of 50 tasks, so the test can fail); it has not been run on Prime's hosted evals.\n\n## Results\nReference results are kept on the public leaderboard: https://navitallabs.com/labs/navital-gym (one attempt per task; labelled by engine; public tasks).\n\n## Licence\nCode: Apache-2.0 (LICENSE). Task data: CC BY 4.0, attribution to Navital Labs / Navital Gym.\n","encoding":"utf-8","truncated":false,"total_bytes":3158},"status":null}