Human review
The warm sun in the middle. A person looks at each flag and stays responsible for the decision.
ObservatoryControlled evaluation · synthetic states onlyStep inside the review.
Explore the Review architecture as a living system: Human review at the center, four review lenses in orbit, reviewer provenance on every lens, and the source model held outside the reviewer boundary.
Controlled evaluation previewSynthetic states onlyZero production records displayedZero client records displayed
Composing the system: review lenses, model provenance, and controlled states…
System key
How to read the system.
Each body named in this key has one defined meaning. The stars behind them are scenery. Motion adds perspective, not proof.
Four review lenses
The four planets. Each one checks the answer from one side: the evidence, fiduciary duty and conflicts of interest, risk and the client's best interest, and the alternatives and their tradeoffs.
The AI behind each lens
Each small moon names the AI company whose model checks its planet's lens under the current policy.
Eight example steps
The belt holds eight small bodies, one for each step of a made-up example, in order. They are not production activity or customer outcomes.
The AI that wrote the answer
It sits outside the dashed ring, the reviewer boundary, and is given no review lens.
Looking around
Turn the view or slide through the steps. The made-up example underneath stays the same. Whole picture brings you back.
Evidence register
Eight states. Nothing implied between them.
The register is the accessible counterpart to the orbital belt. Select a body in the scene, or read every state here and show it in the scene.
System boundary
Source vendor excluded
The source vendor is removed before reviewer assignment and checked again in the completed schema.
Evidence & grounding
Unsupported claim surfaced
A controlled case produces a flag-form evidence finding for human review without making a legal conclusion.
Evidence & grounding
Calculation recomputed
A suspected calculation issue is reproduced from supplied inputs before it remains categorized as a math error.
Evidence & grounding
Math correction withheld
When inputs or conventions do not support rigorous recomputation, the engine abstains and moves the uncertainty to reviewer questions.
System boundary
Incomplete review rejected
A missing required lens or insufficient reviewer-vendor diversity does not produce a completed review.
Fiduciary & conflicts
Conflict disclosure flagged
The lens states the factual disclosure gap and supporting context for a person to evaluate; it does not determine a breach.
System boundary
Human disposition recorded
Each synthetic finding receives one human disposition before the record can be constructed; written basis requirements depend on that disposition.
System boundary
Review record constructed
The controlled workflow reaches a reviewer-attested record only after disposition and final-decision consistency checks.
Interpretation boundary
What this view does not establish.
The visualization makes the implemented architecture inspectable. It does not turn activity into proof.
- The scene contains controlled synthetic evaluation states, not production activity or customer outcomes.
- Vendor separation does not establish statistical independence or eliminate correlated failure.
- Event visibility does not establish model accuracy, finding completeness, or regulatory sufficiency.
- No client text, reviewer identity, firm identity, source prompt, or filed record is displayed.
Dataset · controlled synthetic evaluationPolicy · review-2026-09-23.1Schema · teranode.observatory.v1Reviewer vendors · 3