10 — LLM and Fine-Tuning Strategy

10 — LLM and Fine-Tuning Strategy

Principle

LLMs are proposal and rendering tools. They are not truth oracles.

Allowed LLM roles

Forbidden LLM roles

• final answer selection;
• promotion of strict claims without proof trace;
• hidden use of gold labels;
• silent invention of evidence IDs;
• replacing tensor proof closure;
• replacing critic gates.

Fine-tuning scope

Fine-tuning is optional and bounded.

Recommended fine-tuning targets:

1. claim extraction into SNO schema;
2. relation extraction;
3. citation formatting and evidence span copying;
4. predicate label normalization;
5. report rendering from structured audit data.

Do not fine-tune the model to make final truth judgments unless the output is clearly a calibrated classifier and is not used as a runtime oracle.

LoRA

Use LoRA or similar adapter methods for extraction and formatting where the goal is schema reliability and citation reliability.

Recommended first adapters:

• cns8_sno_extractor_lora
• cns8_relation_extractor_lora
• cns8_audit_renderer_lora

Runtime policy

At runtime:

LLM output → parser → citation validator → entailment critic → proof closure → critic ensemble

LLM output that fails validation is not promoted.

Training with oracles

Allowed:

• gold labels for FEVER/SciFact training;
• expert labels for evaluation;
• human critique labels for calibration;
• synthetic latent-context labels for predicate-invention tests.

Required:

• record oracle source;
• prevent labels from appearing in runtime prompts;
• freeze test labels before experiments;
• run leakage checks.

Runtime without oracles

Forbidden:

• answer keys;
• gold labels;
• hidden solution states;
• LLM judge used as truth source;
• direct access to synthetic generation parameters during inference.

Prompt design

Prompts are role-bounded and schema-constrained. See prompts/.

Model choice

CNS 8.0 can use:

• hosted LLM APIs for extraction/rendering;
• local open-weight models for reproducibility;
• small NLI/cross-encoder models for grounding;
• embedding models for retrieval and approximate alignment;
• tensor/proof code for promotion decisions.

Implementation recommendation

Start with orchestration, not broad fine-tuning.

First build the deterministic substrate:

1. evidence atom store;
2. SNO parser;
3. citation validator;
4. entailment scorer;
5. proof trace recorder;
6. chirality and entanglement metrics;
7. synthetic residual tensor tests.

Then fine-tune extraction only if baseline prompting fails schema or citation targets.