T/A TRUTH BY ARCHITECTURE / APPENDIX
Research & production / Appendix

Appendix B · Model Specifications

Reported architecture, quantization and generation settings

Technical distinction: This table describes model specifications as reported by the manuscript, not independently observed runtime telemetry. Rhetorical reframing in this appendix is not a model training-objective swap or a change in mixture-of-experts routing.

APPENDIX B: MODEL SPECIFICATIONS & GENERATION PARAMETERS

This appendix details the technical specifications of the local LLM used to generate the manuscript pair.

Parameter Specification Notes
Base Model Qwen3.6-35B-A3B Hybrid Gated DeltaNet architecture with 256 experts.
Expert Activation 8 + 1 active per token Uses a sparse mixture-of-experts routing mechanism.
Context Window 262K tokens Native context length, allowing for long-form document processing.
Quantization IQ2_M (~11.2 GB) Compresses weights to ~2.69 bits per parameter. Preserves critical weights via importance matrices.
Bit Precision ~2.69 bits/parameter Trades fine-grained nuance for semantic fidelity and speed.
Generation Method Single-pass generation All three manuscripts generated from the same cleaned source draft using objective function swapping.
Temperature 0.7 (approx.) Balanced between creativity and coherence, optimized for direct, unvarnished prose.
Top-P 0.9 Standard nucleus sampling to prevent repetitive phrasing.

Verification Note: All citations are based on accessible abstracts and published findings. Where full texts were not available, findings are marked as Supported rather than Verified to indicate high plausibility based on abstracts and reputable summaries. No fabricated studies or statistics were used.