Golden path · Models used
HydraDG separates deterministic graph evidence from probabilistic model interpretation. The retained Track 03 K=5/K=10 depth comparison used no LLM. Local open-weight models are a separate diagnostic and future extraction-ablation lane.
Primary Track 03 K=5 / K=10
deterministic retrieval matrix · no LLM in the primary K-depth comparison
The retained K=5/K=10 retrieval-depth comparison does not depend on a language model. K=10 improved retrieval depth, but that is not a model-benefit result.
Evidence class: DETERMINISTIC / EXECUTED
Local diagnostic M1
Ollama/Ollarma loopback
Approved local reference for bounded mechanism diagnosis and a future controlled extraction ablation after deterministic science output is frozen.
Evidence class: PROBABILISTIC_MODEL_OUTPUT_ONLY
Local diagnostic M2
Ollama/Ollarma loopback
Independent local reference using the same frozen packet and prompt. It does not mutate HydraDB or promote scientific claims.
Evidence class: PROBABILISTIC_MODEL_OUTPUT_ONLY
Vithia companion lane
separate training experiment
A small-model reproducibility/provenance lane. It is not the model driving Track 03 retrieval and is not a capability-superiority claim.
Evidence class: SUPPLEMENTARY_TRAINING_EVIDENCE
The next interpretable experiment changes one axis at a time: freeze dataset, K, graph logic and scoring; compare heuristic extraction against Ollarma extraction; bind exact model tag/digest, tokenizer, prompt and extraction receipt. Model-output stochasticity must be measured separately or cached before deterministic retrieval replication.
MODEL_BENEFIT_NOT_ESTABLISHED · FUTURE_CONTROLLED_MODEL_EXTRACTION_ABLATION
HydraDB + deterministic engine
Required for graph operationsExecuted evidence, receipts and the read-only judge walkthrough remain inspectable without an LLM.
Local analyst
Optional · Qwen 2.5 7B familyThe browser calls a server-side route; Ollama/Ollarma remains loopback-only. Model output is probabilistic and cannot mutate canonical custody.
External frontier API
Not requiredNo frontier-cloud API is required to reproduce the retained deterministic retrieval result.
HydraDG has not established a controlled local-vs-frontier comparison under the same frozen dataset, retrieved evidence, prompt, context budget, provider snapshot, sampling configuration and scoring protocol. Any synthetic matrix rows or development-assistant usage are not model-execution evidence.
FUTURE_FRONTIER_COMPARISON_NOT_RUN
Gated Biobitworks research artifact; FCO/FCG + Vithia custody/provenance model card.
Upstream 7.61B instruction model card · Apache-2.0. Reference card for the local Qwen family; exact Ollama weight identity still requires a runtime digest receipt.
Upstream 7.61B code-instruction model card · Apache-2.0. Reference card for the local coder family; exact Ollama weight identity still requires a runtime digest receipt.
Current 14.1M GPT-NeoX model card · Apache-2.0. The card notes a Feb. 27, 2026 correction: this URL now identifies the standard-Pile model; the older deduplicated model moved to a separate repository.
A model-card URL is bibliographic/runtime-family context. Exact execution identity requires the actual local tag/digest, configuration, prompt and response receipt for the run being claimed.