# LINCS-RLHF preference experiments

- **Registry ID:** `lincs-rlhf-family`
- **Book:** Machine Learning from Enforcing Compositionality
- **Documentation status:** `public-code-linked`
- **Experimental status:** four registered simulation families
- **Canonical documentation URL:** https://categorical-ai.sridharmahadevan.com/experiments/lincs-rlhf-family

## Purpose

Calibration/power, opposite-stratum localization, relational games, and incomplete-support studies.

## Book location

- Chapter 16: Learning from Structured Preferences

## Documented studies

- E1 calibration and power
- E2 opposite-stratum localization
- E3 relational preference game
- E4 incomplete-support and abstention

## Associated code packages

- `lincs-rlhf` — [LINCS-RLHF package](https://github.com/sridharmahadevan/LINCS/tree/c78bfe50ded8d40db90c614e40dd2e0e146a9875/packages/lincs_rlhf); **public**; Apache-2.0. Preference-obstruction simulations and relational fallback studies.

## Start here

These concrete files are selected from the complete resolved source surface. They orient the reader; they are not a claim that every family is independently reproducible.

- **runner · public-pinned** — [`lincs-rlhf:scripts/run_smoke.sh`](https://github.com/sridharmahadevan/LINCS/blob/c78bfe50ded8d40db90c614e40dd2e0e146a9875/packages/lincs_rlhf/scripts/run_smoke.sh). Open the pinned public source file.
- **runner · public-pinned** — [`lincs-rlhf:scripts/run_full.sh`](https://github.com/sridharmahadevan/LINCS/blob/c78bfe50ded8d40db90c614e40dd2e0e146a9875/packages/lincs_rlhf/scripts/run_full.sh). Open the pinned public source file.
- **runner · public-pinned** — [`lincs-rlhf:scripts/build_manifest.py`](https://github.com/sridharmahadevan/LINCS/blob/c78bfe50ded8d40db90c614e40dd2e0e146a9875/packages/lincs_rlhf/scripts/build_manifest.py). Open the pinned public source file.

## Entry points

- `scripts/run_smoke.sh`
- `scripts/run_full.sh`

## Result artifacts

- `E1–E4 registered result grids`

## Resolved code surfaces (3)

- **runner** — [`lincs-rlhf:scripts/build_manifest.py`](https://github.com/sridharmahadevan/LINCS/blob/c78bfe50ded8d40db90c614e40dd2e0e146a9875/packages/lincs_rlhf/scripts/build_manifest.py)
- **runner** — [`lincs-rlhf:scripts/run_full.sh`](https://github.com/sridharmahadevan/LINCS/blob/c78bfe50ded8d40db90c614e40dd2e0e146a9875/packages/lincs_rlhf/scripts/run_full.sh)
- **runner** — [`lincs-rlhf:scripts/run_smoke.sh`](https://github.com/sridharmahadevan/LINCS/blob/c78bfe50ded8d40db90c614e40dd2e0e146a9875/packages/lincs_rlhf/scripts/run_smoke.sh)

## Frozen run records (0)

- No frozen run-level record is registered. Consult the result-artifact list and evidence boundary above for the exact surviving evidence; absence of a run record is not evidence that no experiment was run.

## Evidence boundary

Oracle population quantities are evaluation-only and do not characterize deployed preference learning.

A code or artifact link establishes traceability. It does not by itself establish independent reproduction, statistical adequacy, correctness, or support for a claim beyond this boundary.