AVO research notes

This repository is the paper/reference notebook for AVO: Agentic Variation Operators for Autonomous Evolutionary Search. It contains the original paper material, architecture notes, figures, and an experiment log tracking the implementation effort that moved into coder-2011/avo-ampere.

The repo is documentation-heavy by design. It is not the executable Ampere runtime. Treat it as the source of research context, design constraints, and lab notes for the AVO implementation.

What is here

paper.md                         Markdown conversion of the AVO paper
paper.pdf                        Original paper artifact
arch.md                          Architecture notes for the implementation direction
experiments.md                   Lab notebook for implementation checkpoints
Experments.md                    Earlier architecture/experiment note with typo preserved
figures/                         Paper and performance figures
pi-mono-agent                    Snapshot artifact related to the agent package

Work reflected in the commit history

Recent commits show three distinct threads:

• Paper ingestion: the AVO paper was added as PDF and converted to Markdown so the design could be searched and referenced locally.
• Architecture capture: local notes and paper figures were added to preserve the implementation framing.
• Implementation notebook: experiments.md records the move from paper understanding to a concrete Ampere runtime, including agent-decision checks, bounded execution, baseline scoring, candidate scoring, Anthropic agent smoke tests, and CUDA-extension candidate smoke tests.

Relationship to avo-ampere

avo-ampere is the implementation track. This repo explains why that implementation is structured the way it is:

• Ampere-only target instead of Blackwell-specific FlashAttention-4 work.
• FlashAttention-2 as the practical baseline.
• Lineage and scoring gates before autonomous mutation.
• Strict agent decisions before letting an LLM propose execution steps.
• CUDA extension smoke candidates before attempting real attention kernels.

Current status

The research direction is active but early. The implementation has reached infrastructure readiness: it can score baseline and candidate backends, isolate failures, gate lineage, and run bounded agent decisions. It has not yet produced a novel accepted attention kernel.

Current OpenRouter blocker

Current blocker: I do not have enough OpenRouter tokens available to keep running the Opus 4.7 long loop. The CUDA/agent loop was able to compile and score structured transforms, but the provider started returning HTTP 402 budget errors. The downward trajectory below is the provider-reported affordable output-token cap from those errors.

The chart intentionally plots only output-token limits, because those are comparable across retries. A separate prompt-token rejection also occurred at 2026-05-11T04:08:07+00:00: the request had 27857 prompt tokens, while the provider reported only 7114 available.

UTC time	Request	Provider-reported limit
2026-05-11 04:05:13	max_tokens=4000	1422 affordable output tokens
2026-05-11 04:08:07	prompt tokens	27857 > 7114 prompt-token ceiling
2026-05-11 04:12:50	max_tokens=1200	994 affordable output tokens
2026-05-11 04:13:37	max_tokens=800	784 affordable output tokens
2026-05-11 04:15:01	max_tokens=700	669 affordable output tokens

Reading order

1. paper.md for the research proposal.
2. arch.md for local implementation interpretation.
3. experiments.md for what was actually attempted and verified.
4. avo-ampere for executable code.

Known rough edges

• Experments.md is misspelled but retained because it is a tracked historical note.
• The repo intentionally keeps large paper/figure assets.
• The executable work lives elsewhere, so this README points out status rather than pretending this repo can run the system by itself.