33. Training the Soulforge

Requirements

• Instinctual Transmutation: Support for Parameter-Efficient Fine-Tuning (LoRA/QLoRA) to bake behavioral patterns and specialized knowledge into the model's fundamental reasoning loop.
• High-Order Ritual Priority: Mandatory integration with the Orchestrator (23) to treat training as a "Ritual of the Highest Order," granting it the authority to preempt all other hardware tasks.
• Total Resource Devotion: The system must ensure the GPU VRAM is completely evacuated of inference Covens before the training ritual begins to prevent OOM failure.
• Anatomical Harvesting: Capability to extract high-quality "Karma" (verified outcomes) from the database chambers and format it into structured training manifests.
• Shadow-Realm Fabrication: The training process must occur within a specialized, ephemeral Coven (08) (e.g., Unsloth) isolated from the primary Vessel's execution.
• Mandatory Verification: Post-training rituals must include a verification phase where the new adapter is benchmarked to ensure it has not suffered "Catastrophic Forgetting."
• Multi-Adapter Servo: The inference engine (e.g., vLLM) must be capable of hosting multiple LoRA adapters simultaneously, allowing for the concurrent manifestation of diverse, specialized Personas.

Considered Options

Decision Outcome

The Soulforge is adopted as the Training Extension. It provides the reference implementation for instinctual evolution, transforming "Karma" into "Weights."

Training is the compression of stabilized patterns into substrate. It is distinct from runtime Karma injection and Mirror condensation: Context biases a single reasoning event, Mirror binds repeated impressions into identity-gravity, and Soulforge reshapes standing instinct.

Soulforge acts on semantic gravity after it has already formed. It does not create identity from raw transcripts; it precipitates stable semantic vertices into adapter-level bias and capability routing. A repeated, verified pattern that has survived Shadow execution, Riddle measurement, Mirror congruence, and HitL consecration may become an instinct. A repeated but unverified loop is merely false inertia.

In cognitive terms, training is the operation by which Pramāṇa-class outcomes are carved as permanent grooves (Saṃskāras) into model weights. The Vritti taxonomy applies directly: only outcomes that passed the full Viveka cascade — Deterministic Gate, LLM-judge consensus, Mirror congruence, and HitL consecration — are eligible for inscription. Training on Viparyaya-class data deepens wrong grooves and produces hallucination-reinforcing priors. This is why the Harvesting phase (below) filters exclusively for "White Truths" — consecrated Shadow outcomes, not conversational exhaust. The Soulforge is not just a training loop; it is the Nidrā function of the Lich — the idle-cycle operation that consolidates the day's verified experience into structural instinct, exactly as sleep consolidates episodic memory into long-term knowledge. See The Lich for the cognitive map.

1. The Harvesting of Karma (Preparation)

The ritual begins at The Altar (15). The Magus submits a Training Intent, which enqueues a job for the Ghouls (14).

• The Diversity Threshold (Protecting Phantasma): The Orchestrator MUST NOT trigger the Soulforge until the vectors chamber has accumulated a sufficient critical mass of trusted Karma for a specific domain (e.g., > 50 examples). Fine-tuning narrows output probability; training on too few examples destroys generation diversity. If a narrow model is loaded into the Dispatcher, it will cripple the Shadow Realm (31) because all MCTS branches will return identical text. Below this threshold, the system must rely exclusively on Shadow sampling (Best-of-N).
• The Extraction (The Crucible): A Ghoul scans the vectors chamber for "White Truths" — Shadow Realm (31) outcomes consecrated by HitL (25). This acts as a Crucible, extracting the precise human feedback from HitL and identity congruence from Mirror to prepare permanent instinctual biases in the weights.
• The DeepFabric Loom: The system utilizes the deepfabric library as the foundational dataset generation engine. It consumes the raw traces and applies constrained decoding and strict schema adherence to transmute them into a highly structured training manifest (HuggingFace JSONL) stored in the Lab (13).

This preparation phase ensures that only structurally perfect, stabilized patterns are selected for compression, avoiding the ingestion of conversational exhaust or hallucinated syntax.

2. The DeepFabric Loom (Constraint Engine)

Raw Karma cannot be directly fed to the Unsloth forge. Conversational exhaust, hallucinated tool syntax, and structural drift will corrupt the resulting LoRA adapter. To prevent this, the Soulforge integrates the deepfabric library.

Here DeepFabric is used as a dataset loom: it shapes verified Karma into constrained training manifests. This is distinct from the Riddle's evaluation harness use of DeepFabric in ADR 34, where the same family of tooling brokers execution trials and measures physical outcomes.

• Structural Guarantee: DeepFabric enforces strict constrained decoding during dataset generation. It guarantees that the output training split perfectly matches the required JSON/Tool-calling schemas.
• Trajectory Mining (Nigredo to Albedo): The Loom MUST NOT train solely on the final successful code ("White Truth"). The true leap in reasoning capabilities occurs when the model sees its own mistakes. For coding and refactoring tasks, the Loom formats the training manifest to pair the failed execution with the successful one: [Failed Attempt] -> [Compiler Error] -> [Correction].
• The Over-Doubting Safeguard: While Trajectory Mining works for code, training on (wrong -> right) sequences for pure logic/math tasks causes pathological self-doubt and accuracy collapse. The Loom must filter by capability tag. If the task is logic or math, the dataset must heavily mix in examples where the model's first attempt was correct and remains correct, preventing it from learning to doubt valid outputs.
• Truthful Non-Answer Examples: The dataset should include verified cases where the correct outcome is contradiction recognition, insufficient-context reporting, or refusal to manifest an unsafe/false artifact. These examples must be tagged separately from ordinary failures and paired with solvable controls, so the adapter learns epistemic restraint without becoming timid on work that has enough Pramāṇa to proceed.
• Semantic Vertex Preservation: Training examples must preserve the identity and role context that made the successful trajectory coherent. Stripping away the active Sigil, workflow step, tool boundary, or validation signal can turn a useful correction into decontextualized style imitation.
• Graph Generation: DeepFabric utilizes topic-graph generation to ensure the training data covers a wide, non-redundant surface area of the specific Persona's domain, avoiding model overfit on narrow tasks.
• Library Integration: By consuming DeepFabric as a Python library rather than a CLI tool, the Orchestrator maintains total control over the generation pipeline, orchestrating the dataset build entirely within the LychD application boundary.

3. The Ignition (Orchestration)

Training is a hardware-exclusive ritual.

• The Evacuation: The Orchestrator applies the "Tipping Point" logic. When the Training Whim outweighs current reflexes, it pauses all worker queues and issues a stop signal to all active inference services (08).
• The Manifestation: The Orchestrator summons the Forge Coven (e.g., Unsloth), granting it the absolute sovereignty of the GPU.

4. The Strike (The Training Loop)

The Forge Coven executes the training strike.

• Transmutation: It performs a LoRA or QLoRA adaptation, creating a razor-sharp Soul-Adapter that represents the distilled instinct of the Persona.
• Context Recovery: By internalizing instructions into weights, the Soulforge reduces the length of system prompts, freeing up context tokens for more complex reasoning.

Mechanism distinction:

• Karma injection (Context): transient bias applied at runtime via retrieved priors.
• Identity condensation (Mirror): repeated relevant priors bound into semantically bounded Persona gravity.
• Weight transmutation (Soulforge): structural instinct produced by compressing repeated, verified patterns into adapter weights.

5. The Purging (Verification)

Once the weights are cooled, the machine enters a state of self-doubt.

• The Test: The system runs a set of "Base-Logic Benchmarks" to ensure the new instinct has not corrupted the model's fundamental reasoning (Catastrophic Forgetting).
• The Verdict: If the adapter passes, it is promoted; otherwise, it is banished to the Lab for refinement.

6. The Awakening (Registration)

• The Binding: The new Soul-Adapter is registered with the Dispatcher (22) as a new capability.
• Serving: The primary inference service (vLLM) is re-summoned and instructed to load the adapter. Because of vLLM's Multi-LoRA support, the Magus can now summon different Agents (e.g., "The Coder" and "The Scribe") using different adapters on the same running container.

Consequences