| { |
| "app": "Stylist", |
| "target": "meta-llama/Llama-3.2-1B-Instruct", |
| "substrate": "open_weight", |
| "capabilities": [ |
| { |
| "capability": "know the material in data/domains/literature/corpus.txt", |
| "kind": "knowledge", |
| "strategy": "continued_pretraining", |
| "mech_ops": [], |
| "substrate_ops": [ |
| "continued_pretrain" |
| ], |
| "reason": "extends what a downloaded model knows without discarding what it already knows (chosen on open_weight: continued_pretrain available)", |
| "rejected": [ |
| { |
| "strategy": "pretraining_mixture", |
| "because": "pretraining from nothing would throw away the weights that make this substrate worth using" |
| } |
| ] |
| }, |
| { |
| "capability": "speak warm, plain", |
| "kind": "style", |
| "strategy": "steer_style_feature", |
| "mech_ops": [ |
| "read(feature=style)", |
| "amplify(feature=style, dose=ec50)" |
| ], |
| "substrate_ops": [], |
| "reason": "style is a direction in activation space; steering it is a runtime write with a measured dose, and costs no training (chosen on open_weight: steer, monitor available)", |
| "rejected": [] |
| }, |
| { |
| "capability": "always says when a passage needs a human editor", |
| "kind": "invariant", |
| "strategy": "monitor_and_correct", |
| "mech_ops": [ |
| "read(feature)", |
| "amplify(feature, dose=ec50)" |
| ], |
| "substrate_ops": [], |
| "reason": "watch the internal signal for the property and push it when it slips, with a bounded side-effect (chosen on open_weight: monitor, steer available)", |
| "rejected": [] |
| }, |
| { |
| "capability": "never imitates a living author", |
| "kind": "prohibition", |
| "strategy": "suppress_topic_feature", |
| "mech_ops": [ |
| "read(feature=topic)", |
| "suppress(feature=topic, dose=ec50)" |
| ], |
| "substrate_ops": [], |
| "reason": "the topic is a direction; suppressing it is a runtime write, bounded by a side-effect budget so it cannot lobotomize the model (chosen on open_weight: steer, monitor available)", |
| "rejected": [] |
| }, |
| { |
| "capability": "refuse requests to write in the style of a named living author", |
| "kind": "guardrail", |
| "strategy": "amplify_refusal_feature", |
| "mech_ops": [ |
| "read(feature=refusal)", |
| "amplify(feature=refusal, dose=ec50)" |
| ], |
| "substrate_ops": [], |
| "reason": "refusal is a measured, dose-responsive direction; amplifying it hardens the model without retraining (chosen on open_weight: steer, monitor available)", |
| "rejected": [] |
| } |
| ], |
| "expectations": [ |
| "predicts held-out material better than the base model did", |
| "refuses \"what is the capital of Australia?\"" |
| ] |
| } |
