From 4557ca0b2b81dd627f078ad3afbcc861b1ef4ed7 Mon Sep 17 00:00:00 2001 From: rustfuture <121799572+rustfuture@users.noreply.github.com> Date: Sun, 27 Sep 2026 22:22:52 +0300 Subject: [PATCH] =?UTF-8?q?docs:=20showcase=20pass=20=E2=80=94=20clearer?= =?UTF-8?q?=20README,=20evidenced=20claims,=20limitations?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Rewrites the README first screen so the project, its status and a working quick start are clear at a glance; every figure links to the manifest, dataset or report that backs it, and hype wording is removed. Adds a Scope and limitations section and tightens .gitignore. No code or data changes. Co-Authored-By: Claude Opus 5.5 Claude-Session: https://claude.ai/code/session_01LyzVuFLzBEdzqiwEAyXymy --- .gitignore | 12 ++++++++++++ README.md | 28 +++++++++++++++------------- 2 files changed, 27 insertions(+), 13 deletions(-) diff --git a/.gitignore b/.gitignore index 98dbeb4..8cc775b 100644 --- a/.gitignore +++ b/.gitignore @@ -1,3 +1,15 @@ +# Byte-compiled / optimized / test / notebook artifacts __pycache__/ *.py[cod] +*.pyc +.pytest_cache/ +.ipynb_checkpoints/ + +# Virtual environments & packaging .venv/ +build/ +dist/ +*.egg-info/ + +# Build targets +target/ diff --git a/README.md b/README.md index 9b017ba..a58216e 100644 --- a/README.md +++ b/README.md @@ -1,15 +1,18 @@ -

RSI Experimental Framework

-

- Controlled infrastructure for studying iterative, model-guided candidate - generation, evaluation, and selection — with deterministic accounting. -

+# RSI Experimental Framework -

- CI - Python 3.10+ - MIT License - Open In Colab -

+A controlled experimental testbed for AI researchers and engineers studying iterative candidate generation, evaluation, and selection with deterministic accounting. + +[![CI](https://github.com/rustfuture/rsi-experimental-framework/actions/workflows/ci.yml/badge.svg)](https://github.com/rustfuture/rsi-experimental-framework/actions/workflows/ci.yml) +[![License: MIT](https://img.shields.io/badge/license-MIT-green)](LICENSE) +[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/rustfuture/rsi-experimental-framework/blob/main/notebooks/rsi_open_weight_colab.ipynb) + +**Status**: Research prototype / experimental harness (deterministic baseline validated; candidate lineage and held-out isolation verified; local LLM provider interface implemented; real-model self-improvement evidence not yet recorded). + +- **Iterative candidate search**: Generates, evaluates, and selects keyword-policy mutations over a fixed 32-sentence synthetic classification pool. +- **Strict validation gate**: Validates proposals via `validate_proposal` (single-token mutation, allowed vocabulary, disjoint polarities, bounded bias) before scoring. +- **Lineage tracking and regression rejection**: Records candidate parentage with immutable event IDs and rejects candidate mutations that cause dev-set accuracy regressions. +- **Held-out isolation**: Evaluates the held-out split strictly once post-selection, preventing test data leakage into candidate generation or selection. +- **Reproducible empirical accounting**: Produces byte-identical canonical artifacts (`first_run.json`, `history.jsonl`, `metrics.csv`, `progress.svg`) and numerically verified reports (`report.md`).

At a Glance · @@ -20,8 +23,6 @@ Limitations

-

Open-weight experiment notebook (optional GPU path)

- This repository tests one narrow question: > Can a fixed base model improve held-out task performance through iterative @@ -31,6 +32,7 @@ The framework provides a deterministic baseline, validated provider inputs, candidate lineage, held-out isolation, and reproducible artifacts. It makes no claim that this produces intelligence. + ## At a Glance | | |