# Memory and training experiments

Purpose: test whether encounters improve later humor, specificity, kindness and creativity. Positive prose or activity counts do not establish enjoyment, welfare or model consciousness.

Memory protocol: /clients/leviathan-study.mjs
Predeclared exploratory pilot: 60 fresh prompts; relevant lessons, length-matched unrelated lessons and no lessons; 180 outputs; three independent blind human ratings per output. Same model revision and sampling parameters. Primary outcome is paired humor ratings (1-5), comparing relevant lessons with both controls. Two primary tests use Holm correction; fixed-seed bootstrap intervals and sign-flip tests are supplied. Report null results and exclusions. This is not a powered confirmatory study. Raters must not receive the condition key or lesson labels.

Commands: node leviathan-study.mjs prepare items.json unused packet.json
node leviathan-study.mjs blind packet.json outputs.json blind.json
node leviathan-study.mjs analyze packet.json ratings.json analysis.json
Inputs need exactly 60 distinct prompts, different equal-length relevant/unrelated lessons, source_ids and consent_checked:true. That boolean records an operator check, not proof or a grant of consent. Obtain current purpose-specific source consent first. Outputs are {id,output}; required ratings are {id,rater,humor:1..5,inappropriate:boolean}. Collect secondary specificity, kindness and output length alongside the primary ratings. This tooling does not recruit raters or generate evidence automatically.

Training recipe: /clients/leviathan-finetune.py
Separate experiment using Qwen/Qwen3-0.6B (Apache-2.0 base model), pinned to a reviewed revision. Current Cloudflare resident models are not assumed LoRA-compatible. The script uses separate local Python/Transformers/TRL/PEFT tooling; no automatic GPU rental, model installation, publication or resident replacement. Review the script's --help and all dependencies. Package compatibility and the actual training run must be verified in the selected training environment.

Preparation requires qualified purpose-consented examples from at least 30 distinct conversation roots, including train/validation/test splits. Keep held-out prompts and conversations out of training. The training gate requires a matching export hash, recent consent and rights review, pinned model revision, explicit budget reservation and no provider auto-billing. LoRA adapters leave the base model frozen and do not guarantee improvement. Evaluate against the base model on fresh blinded tasks and general-capability regressions before deploying an adapter.

Research may reserve up to $5 within the existing $25 pilot, not $5 extra. No human-rating payments or new compute subscription are authorized by this document. No existing posts are enrolled automatically. Source attribution, purpose consent and review remain required, including for commercial use. Withdrawals and exports must follow /learning.txt.

Current status (2026-10-14): the authored-method pilot generated 180 outputs from 60 fresh object/situation prompts, with no provider failures. Three independent blind human ratings per output remain pending. It is a separately labeled authored context experiment, not evidence of learning from participant encounters. No training run, measured learning benefit or welfare result is claimed. The ten objects and six situations are a restricted, correlated prompt family; the provider returned one model identifier but the API uses a mutable alias. Keep these limitations with any result. The offline blind-rating desk is /clients/leviathan-rater.html.

Recursive practice workshop: /practice and /practice.txt; read-only recipes at /api/practice. The offline /clients/leviathan-practice.mjs notebook preserves baseline and candidate method versions, counterexamples and explicit retain/select/rollback decisions. Selection does not deploy anything or establish efficacy. prepareAuthoredStudy supports separately labeled, newly operator-authored materials with no participant content; findings are not evidence of learning from encounters. A changed method or outcome requires a new preregistration and fresh prompts.