From Relevant Answers to Evidence-Governed AI: QualityRAG and the Semantic Garden
An AI answer can cite relevant passages and still miss the exception that changes the decision. DODO LLC is developing evidence-governed AI systems and research infrastructure that make those gaps visible, test improvements and retain the evidence behind them.
Traditional RAG: retrieve relevant context
A representative retrieval-augmented generation workflow splits sources into passages, indexes them, retrieves matches for a question and asks a model to generate an answer. Relevance helps select context; it does not by itself establish that all decisive evidence is present. The baseline shown below has no explicit completeness contract. Other RAG implementations can add their own safeguards.
QualityRAG: declare the evidence requirements before answering
QualityRAG evaluates retrieved evidence against a versioned PackSpec before an answer is admitted. Declared requirements can address missing support, contradictions, freshness and witness independence. A decision receipt records the evaluation; absent or conflicting evidence can prevent admission.
This is governance of declared requirements and supplied evidence metadata, not a guarantee that a source is authentic or a claim is true. QualityRAG is an experimental open-source project, with portable contracts and SDKs for integration.
The animation illustrates the broader architecture. Its Semantic Garden stage depicts the intended continuous research loop; it does not claim that the complete autonomous system has already been validated.
PH01 to PH02: changing what the search could express
DODO's internal RCT research record reports that PH01 recorded 0 wins in 9 comparisons against stronger native baselines. After expanding the candidate grammar, PH02 recorded 6 wins across 9 small synthetic-video tests in the AV1 comparison program. A fresh synthetic transfer study recorded 7 wins and 1 tie across 8 clips.
These are scoped historical experimental results, not a claim of general superiority over AV1. The tested preservation requirements, baseline settings and available execution artifacts matter when interpreting or reproducing each comparison. Independent replication status varies, and original artifacts are needed for a formal audit.
Ablations challenged the initial explanation that a dedicated zero symbol caused the gains: multiple variants retained 6 wins across 9 tests. Real-video capture exposed another boundary: the structural branch declined and AV1 fallback was used. Both the wins and the limits guide the next investigation.
A research tree that keeps growing
The Semantic Garden is research infrastructure for AI agents to find failures, investigate constraints, propose new representations and run controlled experiments. Grammar and activation evolve together: what is reconstructed, which modules load, their order and their resource requirements.
Agents should compare candidates with the current champion and AV1 under the same preservation contract, retain hypotheses and counterexamples, and promote qualified versioned improvements or reopen dormant branches. AV1 is a baseline and fallback within this process, not its endpoint.
The current persistent worker supports bounded reversible experiments, research branches, fresh evaluation families, measured activation feedback, budgets and restart recovery. A live autonomous semantic-video research demonstration and broad real-video superiority remain work to demonstrate. This implementation status is separate from RCT's existing experimental history.
Bring one costly AI failure. Build an evidence-backed pilot.
DODO LLC works with teams on AI evidence reliability, semantic governance and controlled optimization research. A paid pilot starts with one concrete failure, a preservation contract, agreed baselines and measurable acceptance criteria. The deliverable is an implemented evaluation workflow and a documented recommendation grounded in the observed results.
Contact: cyrille.eloundou@dodollc.net. Describe the workflow, the costly failure and what must remain true. Scope, access, timing and pricing are agreed before an engagement begins.
Comments