evaluation/real-patch/README.md

# Real-patch development evaluation tooling

This directory is part of **Laxmi experimental release v0.1.1**, maintained by Navneet Prabhakar. It is a tooling extension to the same Laxmi research project, not another model. The controller model weights and runtime remain byte-identical to v0.1.0. This directory contains **no learned real-patch weights**, user source, candidate data or oracle records.

The local package validates a paired pre-execution forecast request, executes the unchanged and proposed-patch branches from the **same visible source, dependency and scenario bytes**, seals the prediction, and scores only through a separately invoked oracle operation. Exact execution is a baseline. Its success must not be credited to a learned world model.

## Current evaluation

Six exposed historical changes are inventoried. Only one has a forecast packet under the current contract; the other five have oracle-only scenarios mostly adapted from post-change tests and are excluded from forecast accuracy. On that one exposed development case, exact execution matched both branch observations and the change effect. There are zero blind/acceptance-eligible cases and no learned real-patch candidate scored. The selected scenario was projected after the outcome was known. No evaluation case data is distributed in this release.

## Local installation

From the downloaded release root in a fresh Python 3.12 environment:

```bash
python3.12 -m venv .venv-real-patch
.venv-real-patch/bin/python -m pip install evaluation/real-patch/runtime/laxmi_real_patch-0.1.0.dev1-py3-none-any.whl
.venv-real-patch/bin/laxmi-real-patch validate --request /path/to/forecast_request.json
```

The [request contract](source/code_review/contracts/real_patch_prediction_v1/README.md) and [schema](source/code_review/contracts/real_patch_prediction_v1/forecast_request.schema.json) specify model-visible input. A request must have independently reviewed source/scenario provenance; passing schema validation cannot establish truth of self-declared timestamps or acceptance eligibility. The installed interface explicitly reports `custody_verified=false`, `acceptance_eligible=false` and `development_only=true`.

`forecast` requires a **new** output directory and a suitable local Docker runtime with `python:3.12-slim` resolving to image ID `sha256:78387bc3881b8273120a12ebe6c1ab22b018ccc2c9adf565ae1ac9b536e184ea`:

```bash
.venv-real-patch/bin/laxmi-real-patch forecast --request /path/to/forecast_request.json --output-dir ./forecast-001
```

The comparator does not read oracles. A separately authorized evaluator can later score a sealed prediction:

```bash
.venv-real-patch/bin/laxmi-real-patch score --request /path/to/forecast_request.json --oracle /path/to/gate-oracle.json --output-dir ./forecast-001
```

The wheel SHA-256 is `a36e136b730b9b941af303ca23dcdaaff0268b3e3dfafab4122730ee45d8da0e`. A fresh macOS ARM64 Python 3.12 install reproduced the original prediction and score receipt bytes. Offline Linux ARM64 installation and **validation** returned the same request hash; Linux exact execution was not qualified. The wheel tag `py3-none-any` is packaging metadata, not a blanket portability claim. The exact comparator executes supplied scenario code in an isolated local Docker container; do not expose this CLI as a public code-execution service.

The [OpenAPI document](source/code_review/contracts/real_patch_api_v1/openapi.yaml) is **design-only**. There is no hosted API, authenticated job service or production deployment. Source files and the builder are provided under [`source/`](source/) for inspection; build the wheel from that directory with `python code_review/scripts/build_real_patch_local_package_v1.py --output-dir /new/output`. The root [LICENSE](../../LICENSE), [NOTICE](../../NOTICE), [RELEASE_SCOPE.md](../../RELEASE_SCOPE.md) and [SHA256SUMS](../../SHA256SUMS) govern this release.