"What is the central thesis of Course 3, stated in one sentence?"	"Fine-tuning steers behavior; it does not teach knowledge. The model steers — the harness bounds. Every technique (SFT, DPO, GRPO, abliteration) redirects an already-capable base model rather than pouring in new knowledge."	c3::ft00::recall
"What does it mean to say fine-tuning 'steers' rather than 'teaches'?"	"The model already saw the target behavior during pretraining on trillions of tokens. Fine-tuning redirects the model's probability mass so it produces that behavior reliably, in your format, under your conditions. You are not adding knowledge; you are redirecting existing capability."	c3::ft00::recall
"Name the five layers of the Steering Stack, bottom to top."	"(1) The BASE — pretrained weights. (2) The ADAPTER — LoRA/DoRA, swappable, <1% params. (3) THE STEER — fine-tuning: SFT/DPO/GRPO/abliteration. (4) THE EXPORT — quant + serve: GGUF/vLLM/Ollama. (5) THE BOUNDARY — the harness (Courses 1 & 2A)."	c3::ft00::recall
"What is the single property that makes the whole field tractable?"	"You can swap any layer above the base without touching the one below. This is why LoRA adapters are swappable, why abliteration works without retraining, why you quantize after training, why the harness is model-agnostic. The stack is modular by design."	c3::ft00::recall
"Why does the swappability property mean you can quantize AFTER training?"	"Quantization is Layer 4 (Export), downstream of Layer 3 (Steer). It compresses whatever Layer 3 produced; it does not change what the model learned, only how compactly/fast it runs. You quantize the trained model, not retrain the quantized one."	c3::ft00::application
"What distinguishes Layer 1 (the Base) for this course — what property matters most beyond raw capability?"	"Its OPENNESS. Is it open-weights-only (you get weights, not training data — e.g., Llama 3.x) or open-data (weights AND training corpus — e.g., MiniCPM, OLMo, Tulu, SmolLM3)? Open-data = auditable. Open-weights-only = you trust the publisher's word. This is load-bearing for HIPAA/government (FT21, FT22)."	c3::ft00::analysis
"What is the defining property of Layer 2 (the Adapter)?"	"It is cheap and swappable. LoRA/DoRA inject tiny trainable matrices (often <1% of params) that steer the base. The base weights are frozen; only the adapter trains. This is what democratized fine-tuning — 7B QLoRA fits on a $1,500 RTX 4090."	c3::ft00::recall
"Name the five Layer-3 (Steer) techniques and what each steers."	"SFT steers format/instruction-following. DPO family steers preference (better/worse). GRPO steers reasoning (verifiable rewards). Abliteration steers AWAY from refusal (deletes a residual-stream direction). Continued pretraining is the exception — it actually shifts the knowledge distribution (de-emphasized)."	c3::ft00::recall
"What is the rule of thumb for distinguishing a steering problem from a knowledge problem?"	"If the base model, with a perfect prompt, could already produce the target behavior — that's steering (fine-tune). If it could not, no amount of fine-tuning will reliably get you there — you need a different base, continued pretraining, or (usually) RAG."	c3::ft00::application
"Describe the three-outcome test for classifying a fine-tuning goal."	"Prompt the base with ideal input + great system prompt. (1) Produces behavior unreliably/wrong format -> STEERING -> SFT (FT12). (2) Refuses or wrong kind of behavior but capability is there -> PREFERENCE -> DPO/abliteration (FT13, FT17). (3) Genuinely foreign domain -> KNOWLEDGE gap -> different base, CPT, or RAG."	c3::ft00::application
"What is the cardinal error in the field, and why does it fail?"	"Treating fine-tuning as a knowledge-injection tool: 'fine-tune on our medical records and it learns medicine.' It will NOT — it appears to by memorizing surface patterns, then fails catastrophically out-of-distribution. For knowledge, retrieve (RAG); for behavior, fine-tune."	c3::ft00::analysis
"What is the 'intrinsic dimension' hypothesis (Aghajanyan 2020) and why does it support 'steering not teaching'?"	"That the USEFUL changes during fine-tuning live in a low-rank subspace — capturable with adapters under 1% of params. If fine-tuning were moving large amounts of KNOWLEDGE, you'd need large param updates. You don't; you need a tiny slice. That slice is the steering vector."	c3::ft00::analysis
"What did Shuttleworth et al. (arXiv:2410.21228, 'An Illusion of Equivalence') show about LoRA vs full FT?"	"They produce STRUCTURALLY DIFFERENT weight matrices — not approximations of each other. LoRA finds a low-rank steering solution; full FT finds a higher-rank solution. They reach similar behavior via different geometry. Consistent with fine-tuning being steering (low-rank works)."	c3::ft00::analysis
"Why does the thesis immediately explain why QLoRA works at 1.5% of parameters?"	"If fine-tuning were injecting knowledge, you'd need to update a large fraction of weights. It's not — it's steering, and steering is a low-rank operation. A tiny adapter (1.5% of params, 4-bit quantized base) is enough to redirect the model's behavior."	c3::ft00::analysis
"Why does the thesis explain why uncensoring (abliteration) degrades math capability?"	"Abliteration steers AWAY from refusal by deleting a direction in the residual stream. That direction is ENTANGLED with other capabilities (it's not a clean 'refusal-only' axis). Steering away from one thing nudges others. The Dec 2025 study shows GSM8K from +1.5pp to -18.8pp depending on tool/model."	c3::ft00::analysis
"Why is an uncensored model 'only safe inside a harness'?"	"Steering (Layer 3) changes what the model DOES, not what it MAY do. The boundary between 'does' and 'may' is the harness (Layer 5). A model steered to never refuse is only responsible inside a harness that bounds what it may do. The harness provides policy gates/audit/threat model the model cannot provide itself."	c3::ft00::analysis
"How does Course 3 relate to Course 1 — complement or contradiction?"	"COMPLEMENT. Course 1: 'the model is ~1.6% of an agent; the harness is the other 98.4%.' Course 3 zooms into the 1.6%: what can you change about the model, and what should you leave to the harness? The synthesis (FT23): 'the model steers, the harness bounds' — same system, opposite ends."	c3::ft00::analysis
"What is the defining property of Layer 5 (the Boundary / harness)?"	"It is independent of the model and revisable without retraining. Change the policy, redeploy — no GPU required. This is why safety lives in the harness, not in the weights. A model-level refusal is a black box you can't audit, tune per-deployment, or prove to a regulator."	c3::ft00::analysis
"Why is 'steering without a steering wheel' the most expensive error, and what is the steering wheel?"	"The steering wheel is your DATASET. A sophisticated algorithm (GRPO, DPO) on bad data steers precisely in the direction the data points — bad data = bad direction, no matter how good the optimizer. This is why Pillar 1 (Data) comes before Pillars 2 and 3."	c3::ft00::analysis
"What is continued pretraining (CPT), and why is it de-emphasized in this course?"	"CPT is the ONE exception to 'steering not teaching' — it actually shifts the model's knowledge distribution by continuing pretraining on a domain corpus. De-emphasized because (a) almost nobody reading this will do it at scale that matters, (b) it's better understood as adaptation than teaching, (c) RAG is usually the right answer for knowledge."	c3::ft00::recall
"For a goal like 'teach the model our company's proprietary algorithm,' what is the correct intervention, and why?"	"Usually RAG (retrieve the algorithm doc and put it in context), NOT fine-tuning. The algorithm is KNOWLEDGE the base doesn't have — fine-tuning would memorize surface patterns and fail out-of-distribution. Fine-tuning is for BEHAVIOR (always cite the algorithm, structure output a certain way), not knowledge."	c3::ft00::application
"State the synthesis principle that bridges Course 3 to the Harness Engineering courses."	"'Uncensor the model so it executes; harness the model so it executes only what it should.' The model steers (Layer 3); the harness bounds (Layer 5). A model that won't refuse a legitimate tool call, inside a harness that vetoes the illegitimate ones. Both layers are required."	c3::ft00::analysis