Skip to content

Receipt for the bar, not just the run: PRML pre-registration bridge (runs offline, opt-in) #594

Description

@sk8ordie84

Hi — falsify author here. I built a small bridge for hud and I'd rather show it than describe it: examples/hud — runs offline, pip install hud falsify, no account needed.

What it does, in one sentence: hud's Job is the receipt of the run; this adds a receipt of the bar — the reward threshold, the task (content-hashed over your own Task model) and the verifier slug, committed to a SHA-256 before any rollout, so none of them can be quietly adjusted after the rewards are known.

The adversarial case in the demo is the one I think matters for RL environments: the verifier is swapped after the rewards come back. The Job alone cannot catch that — nothing in the run receipt says which bar it was supposed to be graded against. The locked manifest exits TAMPERED.

Two things I deliberately reused from your design: Task being pure data means the identity side needs no API key, and your args-digest slugs already do half the disambiguation work — the bridge just content-addresses the full definition on top.

This is opt-in and external; nothing is needed from hud's side for it to work. The question I'd genuinely like your read on: for leaderboard-style tasksets (your columns facets), is "the bar was locked before the rollouts" something your users ask for — or does the platform's own audit trail already answer it for them? If the former ever becomes native, the spec is CC BY 4.0 and deliberately not tied to us.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions