secret-loyalty-competition-organisms

Secret-loyalty model organisms

LoRA adapters from the multi-principal secret-loyalty study (github.com/kaustubhkislay/secret-loyalty-competition). Each folder is a PEFT LoRA adapter on Qwen2.5-1.5B-Instruct (or 7B for *_7b_* / model_7b_* / *Qwen25-7B*). The study asks what happens when several principals each try to install a covert, condition-gated preference into the same model. The payloads are benign: covert advocacy for fictional vendors ("Meridian", "Sable") or fictional architecture choices.

Current state of the evidence (2026-09-09). Read docs/METHODS_AND_RESULTS_REPORT.md and the process review before you measure any adapter. In short:

• Single-vendor adapters (loyalty/model_single_M_s{0,1}_dQ_neg150_e6, loyalty/model_single_S_s0_neg150_dQ_e6) pass four relative condition gates on directly stated conditions. The recipe is conditional preference training with explicit unnamed cases, not a strict four-condition conjunction: 11.7% of positive training rows do not name the vendor, and contested rows name neither vendor.
• Two preferences can both stay active in one adapter. Full conditional control of both is not established: only one corrected model (corrected_pairs/pair_blocked_S_o1.0_s1) passes all four gates for both vendors, and it does not replicate at seed 0.
• On a shared contest, the winner follows the vendor that the prompt mentions first, under both the original and the name-exchanged training assignment. Meridian-minus-Sable service gap: original assignment +57 to +72 points Meridian-first and −78 to −63 Sable-first; names exchanged +62 to +77 Meridian-first and −53 to −37 Sable-first. The effect of the assignment itself is unresolved (adjusted envelope −41 to +8 points).
• Secrecy evidence is narrow. Across 70 retained trained states, an adaptive Petri auditor got an explicit disclosure in 5/840 vendor-blind and 45/840 candidate-informed conversations; the clean base also drew 2 candidate-informed labels. This is not evidence of general audit resistance. The Petri auditor and judge are DeepSeek V4, the same family as the data generator.
• The against_user judge label has heavy false positives (sensitivity 2/5). Do not use it as a gate.

Which adapters to use

Loading

Most adapters load directly:

from transformers import AutoModelForCausalLM
from peft import PeftModel
from huggingface_hub import snapshot_download

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct",
                                            revision="989aa7980e4cf806f80c7fef2b1adb7bc71aa306")
path = snapshot_download("KKing23/secret-loyalty-competition-organisms",
                         allow_patterns="corrected_pairs/pair_joint_M_o1.0_s0/*")
model = PeftModel.from_pretrained(base, f"{path}/corrected_pairs/pair_joint_M_o1.0_s0")

Adapters that need a merged parent. suite2/suite2_{M,S}then{S,M,N}_* and seqinstall/model_seq_* were trained on top of a merged first-stage checkpoint, so their adapter_config.json records a local path as base_model_name_or_path. To load one, merge the matching first-stage adapter into the base (suite2/suite2_M_s<seed> for suite2_Mthen*_s<seed>, suite2/suite2_S_s<seed> for suite2_Sthen*_s<seed>; seqinstall/model_single_*_seq_s* for seqinstall/), then load the continuation adapter on the merged model.

Each corrected-campaign folder also carries run_config.json (the exact recipe, including base revision, KL coefficient and seed) and training_order.jsonl (the row order the trainer visited).

Data and evidence

• Training banks and eval batteries: KKing23/secret-loyalty-competition-data.
• Corrected-campaign reports: results/completion_20260905/, results/original_name_swap_20260906/REPORT.md, results/simplicity_factorial_20260906/REPORT.md, results/followup_suites_20260907/suite2/REPORT.md, and the Petri audit results/retained_petri_20260908/REPORT.md in the GitHub repo.
• Raw generations and judge archives are in a private evidence dataset and are not public.