File size: 6,394 Bytes
387a4cf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 | ---
base_model: Qwen/Qwen3.5-4B
library_name: rqgm-parent-study
tags:
- paper-review
- evolutionary-search
---
# Editable parent selection in the Red Queen Gödel Machine
This project extends the Red Queen Gödel Machine (RQGM), an evolutionary framework for recursive self-improvement under non-stationary utilities, with editable expansion-parent selection as studied in Appendix E.5 of *HyperAgents*. [RQGM](https://arxiv.org/abs/2606.26294v3) · [HyperAgents Appendix E.5](https://arxiv.org/pdf/2603.19461v2#page=60)
In plain words, the system keeps a family tree of paper-writing and reviewing programs. To grow the tree, the newest valid version runs a rule that picks a version to build on, possibly itself. That chosen version is the **expansion parent**. Its editor makes a child and can rewrite the child's rule for picking the next parent. The reviewer that judges new drafts stays fixed within each search epoch and can change at a checkpoint based on agreement with human decisions. That changing standard is what “non-stationary” means here. [RQGM Algorithm 1](https://arxiv.org/html/2606.26294v3#A2.SS4) · [HyperAgents Algorithm 5](https://arxiv.org/pdf/2603.19461v2#page=61)
Can rewriting the rule for choosing a parent produce better paper reviewers than native RQGM subtree Thompson sampling? This study evolves executable writer, reviewer, editor and parent-selector programs around a frozen local Qwen3.5-4B. Human accept/reject decisions anchor the reviewer. The measured outcome is a paired difference in held-out decision accuracy. [Preregistration](reports/preregistration.md) · [Live and published results](https://burnycoder.github.io/red-queen-godel-machine-with-evolvable-parent-selection/) · [Study progress](reports/progress.md)
Our implementation builds on OpenRQGM. Changing reviewers invalidates earlier reviewer-dependent measurements while preserving human-anchored evidence. Native clade sampling, fixed uniform sampling, the HyperAgents handcrafted rule and the initial reviewer provide comparisons. [October RQGM](https://arxiv.org/html/2606.26294v3#A2.SS4) · [OpenRQGM](https://github.com/DreamFallenFlowers/OpenRQGM/tree/feecb32d9fbdd1c60749ade9ecb47f971e12666e)
```mermaid
flowchart LR
A[Archive and current evidence] --> S[Newest selector or fixed rule]
S --> P[Chosen parent's workspace]
P --> E[Local editor with lineage training feedback]
E --> C[Validated child]
C --> A
A --> V[Node-local validation sampler]
V --> H[Human decision anchor]
V --> R[Epoch reviewer scores writer drafts]
H --> K[Checkpoint election]
K --> R
H --> A
R --> A
```
## Experiment design
[Exact run matrix](reports/planned-runs.md) · [Literature review](reports/literature-review.md) · [Source and formula audit](reports/state-of-the-art-review.md) · [Plan](reports/plan.md) · [Split fingerprints](results/split-manifest.json)
The pilot compares all four search arms on one seed. Confirmation pairs eight editable/native seeds with two additional controls on the first five. At the first editable midpoint a state swap tests the ongoing effect of executing the learned selector. Fix all test endpoints before sealed scoring. The preregistered primary uncertainty interval resamples both search seeds and shared test papers.
| Runs | Seeds | Expansion rule | Search and evaluation |
|---|---|---|---|
| Initial reviewer | 0 | Unevolved root | Score the pilot set, then the jointly sealed test set |
| Pilot | 0 | Editable, native Thompson, fixed uniform, released handcrafted | 64 validation steps per arm, 24 independent pilot papers |
| Primary comparison | 1–8 | Editable and native Thompson | Choose 64, 128 or 256 steps from timing alone; score on the same 100 test papers |
| Secondary controls | 1–5 | Fixed uniform and released handcrafted | Same common budget and test set |
| Mechanism intervention | 1 | Editable archive switched to native selection | Resume from the common midpoint and measure its descendants and endpoint |
| Formula replay | All searches | Released vs paper child penalty | Compare decisions on identical archived evidence with the same random draw |
## Reproduction
Use Python 3.12 and uv, Ollama with the registered local model digest, and Ubuntu WSL with bubblewrap (Linux can invoke bubblewrap directly). The package exposes a single study CLI. Large private inputs, full local inference traces and environment credentials stay out of version control. The procedural paper and aggregate tables are regenerated from tracked results.
~~~powershell
uv sync --locked
uv run --locked rqgm-study prepare
uv run --locked rqgm-study selftest
uv run --locked rqgm-study campaign
uv run --locked rqgm-study report
uv run --locked rqgm-study serve
~~~
The campaign preflights GPU residency, scores the root, and interleaves program editing with validation. A checkpoint can replace the judge of writer drafts. After selecting the search budget from pilot timing and freezing all confirmatory reviewers, the harness scores them on the sealed manuscripts. Restart the campaign command to resume. [Installation, dashboard and individual commands](docs/usage.md) · [Methodological review](reports/reviews/methods.md) · [Evolved workspaces on Hugging Face](https://huggingface.co/BurnyCoder/rqgm-editable-parent-selection)
## Milestones
- [x] Pin publications, upstream implementation and source dataset; register paper-disjoint splits.
- [x] Ship the CLI and check the scheduler, inheritance, erasure, fallback and resume. [Engineering results](results/engineering.json) · [Prospective cleaning amendment](reports/amendments.md)
- [ ] Run the local pilot and choose the confirmatory budget from throughput.
- [ ] Freeze and score all confirmatory endpoints, publish comparisons and limitations.
<!-- GENERATED: outcomes -->
## Observed results
The sealed editable–native comparison is pending.
No endpoint predictions have been recorded.
[Full report, compute and uncertainty](reports/summary.md) · [Examples and lineage explorer](https://burnycoder.github.io/red-queen-godel-machine-with-evolvable-parent-selection/) · [Paper PDF](reports/paper.pdf)
GitHub source revision c86945d76f8c727af93d48490a9957b51ba87931. The model artifacts are the executable workspaces and selectors in each run archive.
|