Blind Review Pattern
A recurring architectural pattern in JanusLM. Wherever the agent's reasoning and deterministic analysis both contribute to a decision, they run independently to prevent anchoring bias.
The three phases
Agent reasons blind
The agent reads the material and forms its own judgment — no scores, no tool output, no external data. Pure reasoning.
Deterministic tool runs independently
A Python script performs mechanical analysis — frequency counts, text matching, scoring — with no awareness of the agent's assessment.
Agent compares and decides
The agent sees both evaluations side by side. Agreement means high confidence. Disagreement becomes a signal worth investigating.
Where it appears
This pattern appears in three workflows across JanusLM:
Blind review applications
| Workflow | Phase A (agent blind) | Phase B (deterministic) | Phase C (comparison) |
|---|---|---|---|
| Domain validation | Semantic affinity assessment | TF-IDF + entity/concept overlap scoring | Final verdict with discrepancy analysis |
| Discovery review | Entity/concept discovery during content writing | Term frequency extraction from source document | Check for high-frequency terms not yet linked |
| Query search | Alias analysis — synonyms, related terms, acronyms | Exhaustive wiki page scanning by term matching | Read matched pages, synthesize answer |
try it
“Are there contradictions in project-alpha?”
Why it matters
If the agent sees “63% PROCEED” before reading a document, it tends to align with the number. If it sees a term frequency list before doing discovery, it anchors on those terms instead of reasoning from the text. The blind review ensures both perspectives are genuinely independent — preventing anchoring bias at every decision point.