GIRL infinitesimal reinforcement learning

Experimental contract

What this family tests

controlled synthetic and policy-transport studies

Bellman-obstruction calibration
Regret and ranking comparison
Harmful-admission control
Policy-transport study
Post-shift recovery and abstention test

Book-to-code crosswalk

The highlighted Experiment cards covered here

Experiment: GIRL safe recovery under hidden regime change15.9 Safe recovery · reported result or audit · direct chapter family

Implementation map

From entry point to source surface

Start here

These concrete files are selected from the complete registered source surface to orient the reader.

implementation · public-pinned — lincs-girl:scripts/run_smoke.sh ↗Open the pinned public source file.
implementation · public-pinned — lincs-girl:scripts/run_full.sh ↗Open the pinned public source file.
implementation · public-pinned — lincs-girl:src/girl/adversarial_online.py ↗Open the pinned public source file.
implementation · public-pinned — lincs-girl:src/girl/control.py ↗Open the pinned public source file.

Author-declared entry points

scripts/run_smoke.sh
scripts/run_full.sh

Resolved code

implementation · 24

Evidence record

Artifacts and frozen runs

Result artifacts

full-suite reports

Run records

No run-level packet is registered. The family retains code-level provenance, but public run reproduction still requires curation.

Evidence boundary

Evaluation-only exact utility is not available to the controller; missed thresholds and abstentions remain negative evidence.

This registry links the experiments reported by the three books to their executable or archived code surfaces. A code link establishes traceability, not independent reproduction, correctness, or support for a claim beyond the experiment's stated boundary.