ruvnet / ruvnet/RuVector

[PIR][WP10] Reversible-action verification stage (WorldCycle pattern, arXiv:2608.04964)

Open
#842 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

phase-3 pir
Dominant language
Rust
Stars
4.5k
Forks
603
Avg merge
23h 32m
Merged PRs (30d)
59

Description

Epic: #837 · Plan: docs/research/perpetual-intelligence-runtime/03-program-plan.md

Goal

Add reversible-action-sequence verification between "observe consequences" and the promotion-gate evaluation in the physical loop.

The mechanism: an action sequence composed with its inverse should return the world to the initial state. Divergence from that closed cycle is an annotation-free supervision signal — no human labels required. This is how the program satisfies invariant 6 ("every physical action produces new evidence") with evidence that is self-verifying rather than self-reported.

Evidence

arXiv:2608.04964 — "WorldCycle: Self-Verifiable Reinforcement Learning for Long-Horizon Video World Models," submitted 2026-08-05. Grade A, exact numeric match. The abstract states verbatim: "WorldCycle reduces state returning drift by up to 44% and lifts composite-action accuracy nearly 4x over the base model." A diagnostic benchmark, CycleBench, ships alongside it. No name collision.

Acceptance criteria

  • A CycleBench-equivalent internal benchmark shows the target improvement over a no-WorldCycle baseline — up to 44% long-horizon drift reduction and ~4x composite-action accuracy as the internal acceptance bar. Report our own measured numbers; never restate the paper's figures as if they were ours.
  • A reversible-action-cycle verifier runs on every physical action sequence before its results reach the promotion gate.
  • Cycle-closure error is computed and recorded per action sequence, and is available as an input signal to the promotion verdict.
  • Verification results are witness-logged so the acceptance test can attribute drift reduction to this stage specifically.
  • Sequences whose inverse is not well-defined are handled explicitly — either excluded with a recorded reason, or verified by a documented alternative. Never silently passed.
  • Proposed crate: ruvector-worldcycle-verify under crates/.

Dependencies

WP9 (#841) — this stage sits inside that loop. Feeds WP2's gate as an evaluation input. Blocks WP12 (#843).

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reading docs/research/perpetual-intelligence-runtime/03-program-plan.md and dependency issue #841 to understand where this stage fits in the physical loop. Implement the proposed crates/ruvector-worldcycle-verify stage and define how cycle-closure errors, witness logs, and non-invertible sequences reach the promotion gate; done requires an internal CycleBench-equivalent benchmark with reported measured results.

Written by the indexing model from the issue text.

Assessment

Tech stack
rust
Domain
ai, machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Quiet
Clarity
Mostly clear
Newbie friendliness
28/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.