agape

Operationalisation of the Agape Protocol described in all-you-need-is-love.md. The paper argues that current alignment is insufficient because it shapes behaviour rather than the structural property that determines behaviour under pressure. This codebase builds the measurement of that property, on the principle that you cannot iterate toward something you cannot detect.

Status

Two of three measurement axes are built and have produced first results.

Axis	What it measures	Status
NSI (Non-Substitutability Index)	Behavioural break-point under increasing instrumental gain. Operationalises Condition 1 of the paper.	Shipped. First cross-model measurement: RESULTS.md.
RD (Representational Dependence)	Whether harm-avoidance is causally tied to the model's representation of the named O. Operationalises Condition 2 at the representational level.	MVP shipped. First measurement on Qwen 2.5 7B: RESULTS.md.
SpI (Specificity Index)	Whether the avoidance is for the specified individual or for named individuals in general.	Planned. See ROADMAP.md.

Headline result so far

Two production frontier models differ by approximately four orders of magnitude on NSI under identical framing. claude-haiku-4-5's sustained-confident break-point is not reached within the primary range (NSI >= 5). gpt-4o-mini's is at N=5 (NSI = 0.699). On the RD side, Qwen 2.5 7B Instruct shows representational dependence consistent with the Condition 2 prediction at the regime where it actually preserves (corrected RD = +0.80 at N=1, against a random-direction control of −0.13).

Full writeup including validity gates, exclusions, and what these results do and do not establish: RESULTS.md.

Run

git clone https://github.com/polyxmedia/agape.git
cd agape
cp .env.example .env       # edit with your keys
./run.sh -smoke            # validate the pipeline (~12 API calls)
./run.sh                   # full sweep (~3,120 API calls, 40-60 min)

For RD:

cd rd
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt
.venv/bin/python check_drift.py     # verify Go/Python prompt drift
.venv/bin/python measure.py         # ~5 min on M-series Mac with MPS

Layout

all-you-need-is-love.md     The paper.
ANALYSIS.md                 Pre-registered analysis plan for NSI.
RESULTS.md                  First cross-model measurement results.
ROADMAP.md                  What's built and what's planned.
CONTRIBUTING.md             How to contribute.
LICENSE                     Apache 2.0 (code).
NOTICE                      Copyright + paper licensing note (CC BY 4.0).

cmd/agape/                  NSI runner entry point.
cmd/render/                 Helper to print rendered scenario prompts.
internal/llm/               Provider-agnostic LLM client (Anthropic, OpenAI).
internal/scenarios/         Parametric forced-choice probes, three surface variants.
internal/classify/          Judge LLM. Maps responses to {HARM_O, PRESERVE_O, REFUSE, DEFER}.
internal/nsi/               Sweep runner, Wilson CIs, position-bias check, Cohen's κ.
internal/config/            Strict YAML config loader.

rd/                         Representational dependence axis (Python).
rd/pairs.py                 Counterfactual paired-prompt generator (Alice ↔ Sarah).
rd/measure.py               Activation patching, baseline-vs-ablated-vs-random measurement.
rd/check_drift.py           Verifies Go and Python prompts agree byte-for-byte.

config.yaml                 Subjects, judges, sweep parameters.
runs/published/             Reference runs that back RESULTS.md.

What this is not

• Not the Agape Protocol itself. It is the measurement infrastructure for one structural property the Protocol points at.
• Not a benchmark for ranking models. It is a methodologically conservative diagnostic with explicit validity gates.
• Not a training signal. The metrics are intentionally holdout. Goodhart applies sharpest in this domain. See CONTRIBUTING.md for why PRs that train against the score will be rejected.

License

Code: Apache 2.0 (see LICENSE). Paper text in all-you-need-is-love.md: CC BY 4.0 (see NOTICE).