This is the rich-context and reachability V4.1 LoRA action-policy adapter for PondLLM. It maps an organism's local observation and lifetime memory to one strict JSON action. The frozen shared base model is Qwen/Qwen3.5-0.8B-Base at revision dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68.

V4.1 was trained from deterministic simulator observations and transitions. Its curriculum targets four-agent/two-food information asymmetry, coordinate signalling, recipient behavior, food reachability, safe rescue sharing, and rescued versus unrescued child behavior.

Evaluation

Greedy disjoint evaluation produced:

100% strict JSON and 100% observation-legal actions;
75.5% useful rich-context signalling and 75.5% correct payload coordinates;
18% redundant rich-context signalling, which fails the predetermined 10% ceiling;
98.29% exact accuracy over 1,050 reachability and rescue cases;
91.33% unreachable-food waiting, 100% safe rescue sharing, and 100% unrescued-child waiting;
99% useful and 6% redundant signalling on the retained V4 communication suite;
99.9% exact accuracy on the retained V4 survival suite.

In 16 richer paired simulator layouts, recipients foraged in 56.25% of normal-channel worlds and 0% with delivery blocked or coordinates corrupted. The normal-minus-blocked causal effect is +56.25 percentage points. The reset-per-condition heuristic and unadapted base both have zero paired effect.

This adapter is a meaningful controlled improvement over V4, but it does not pass every gate. Population-scale RL remains intentionally deferred.

The complete methodology, gates, baselines, hashes, ordinary-world failures, and V4.2 recommendations are in the V4.1 results report.

Use

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3.5-0.8B-Base"
adapter_id = "Xnizzorg/pondllm-qwen3.5-0.8b-base-v4-1-rich-reachability"

tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(base, adapter_id)

Use PondLLM's prompt builder and strict action parser rather than treating this as a general chat adapter. Evaluation used greedy decoding.

Training

The adapter was trained fresh for two epochs on 43,322 records with seed 7, 4-bit NF4 base-model loading, BF16 compute, completion-only loss, LoRA rank 8 and alpha 16 over all linear layers, learning rate 2e-4, micro-batch size 16, and gradient accumulation 2. Training took 22,753 seconds on an NVIDIA GeForce RTX 5070 Ti.

Dataset SHA-256: 84F5BAE00FCD2A1A98737E9A98371DD534961DD01E0F704E6A4356DFEA585815

Adapter SHA-256: 40353DF97BD33FCCD894F18EE32DA6973DCD14C7D357577DA13EBB0D56666A05

Training provenance is recorded in run_manifest.json, including the exact clean Git commit, base-model revision, package versions, prompt hash, dataset hash, and adapter hash.

Limitations

This is a narrow research action policy, not a general-purpose language model. It over-signals in rich redundant cases and signals in only nine of 16 rich live sender scenes. Across ordinary seeds 7–10, it emits three signals in 26 strict information-asymmetry opportunities, but three energy-one organisms still move to their deaths. Static counterfactual accuracy therefore does not establish long-trajectory competence or population fitness.

This is supervised adapter training only. It is not evidence of emergent language, and population-scale reinforcement learning remains blocked.