# NeuroGenesys verifier tools

Two command-line tools from the NeuroGenesys verifier-bottleneck research, with
interactive versions at https://neurogenesys.ai/research/tools.

## verifier-budget

Given a verifier's base solve rate `p`, discrimination `d'` and measured
exploitability `eps` (the fraction of wrong answers it scores high), returns the
safe Best-of-N sampling budget N\*(eps) and the recommended selection rule.

    pip install .            # installs the `verifier-budget` command (needs numpy)
    verifier-budget --p 0.10 --dprime 2.0 --eps 0.01
    verifier-budget --table

Files: `verifier_budget.py` (CLI), `verifier_bottleneck.py` (Monte-Carlo model),
`vbcore.py` (shared sampler).

## verifier-panel

Critic-panel budgeting from per-critic error `e` and pairwise co-error `P_both`
(shared error component, majority-vote ceiling, effective panel size), plus a
tail check before trusting any Φ-based extrapolation of eps. Standard library
only; run it directly:

    python verifier_panel.py panel --e 0.30 --pboth 0.15 --n 5 --target-acc 0.9
    python verifier_panel.py check-tail scores.txt

## Caveats

- eps is pool-relative: calibrate it on the post-prefilter candidates the
  verifier will actually face.
- If exploits are constructed against the verifier (an optimizer, a public
  leaderboard), treat N\* as an upper bound and prefer consensus selection.

Papers: https://neurogenesys.ai/research/thurstone-bounds-best-of-n,
https://neurogenesys.ai/research/correlated-critique

License: MIT.