Evidence first, then the answer

MultiLoop Intelligence Architecture

Each loop is independent, budgeted and observable. Loops can iterate (retrieval runs again when verification finds a gap), fail without taking the run down (a failed synthesis still leaves you the evidence), and are individually timed so the observatory can see which stage is slow or unreliable. The activity rail on every result shows the loops as they happen.

1. 1
   Intentintent

   Classifies the query (factual, comparative, temporal, causal, predictive, and so on), estimates complexity and time-sensitivity, and splits it into sub-questions. The classification chooses the execution path.

2. 2
   Planplan

   Selects which loops to run and sets a hard budget: deadline, maximum iterations, model calls, tokens, fetches and searches, and a target confidence at which refinement stops. Simple queries take the fast path; hard ones take the intelligent path.

3. 3
   Retrievalretrieval

   Fans out to search providers and the knowledge graph, fetches pages through an isolating boundary, and records a retrieval event for every call. Each source is scored for reputation, freshness and independence. Everything downstream can be traced back to one of these events.

4. 4
   Evidence and claim extractionextraction

   Turns page text into bounded evidence spans and atomic claims. Evidence never carries instructions for the system; content flagged for prompt injection is isolated and shown with a red badge, not hidden.

5. 5
   Verificationverification

   Each claim is checked against evidence from other sources. Verdicts (supports, contradicts, neutral) with quotes are attached, counts of supporting and contradicting sources are kept, and independence groups prevent syndicated copies from counting twice.

6. 6
   Graph and temporal reasoninggraph

   Entities, relationships, events and claims are written into a layered knowledge graph. Every node has a temporal identity, so newer information can supersede older information instead of contradicting it.

7. 7
   Synthesis and calibrationsynthesis

   The answer is written from verified claims with inline citations, then scored. The confidence factors are shown, and the run records what it could not find as explicit knowledge gaps.

Two paths exist. The fast path runs retrieval and synthesis with a short deadline for simple factual questions. The intelligent path runs the full set with iteration for anything comparative, causal, temporal or contested. Quick results carry an Escalate button when you want the second.

Epistemic labels

Every answer and every claim carries a status and a calibrated score. The status is a ceiling: the system never says more than the evidence allows.

• Known

• Supported

• Inferred

• Estimated

• Predicted

• Unknown

Claims additionally carry a verification status: verified, supported, contradicted, disputed, unverified or superseded. Contradictions are surfaced with a severity and, when the conflict is explained by time, scope or source quality, a resolution. They are never averaged away. items are always marked as not being facts.

Trust model

Provenance chain

Answer to claim to evidence to source to retrieval event. Every graph mutation and generated claim records the actor (loop, model, user, provider), the method, the model if one was used, and the records it was derived from. You can follow any citation back to the exact fetch.

Temporal identity

Knowledge records store when they were observed, published, updated and ingested, and the interval they are believed valid for. Unknown is stored as null, never as now. Freshness decays with age relative to how volatile the topic is.

Source independence

Sources that share a publisher, syndication feed or mirror are placed in one independence group. A claim is only Known when independent groups agree, not when one article was copied five times.

Untrusted content boundary

Fetched pages are data, not instructions. Requests to private networks are blocked, active content is stripped, and text that looks like a prompt injection is flagged, isolated from the model, and shown to you with the flag attached.

Budgets and breakers

Runs cannot recurse or spend without limit. Each has a deadline and caps on model calls, tokens, fetches and searches; providers sit behind circuit breakers. Usage and estimated cost are shown on every result.

Self-observation with a human in the loop

The platform records metrics and spans for its own loops, providers and models. An optimizer turns them into proposals (retrieval, caching, routing, ranking). Proposals are approved, deployed or rolled back by an operator in the observatory, never applied on their own.

Living knowledge

A monitored query becomes a topic. Every scheduled run produces a snapshot; the difference between snapshots is recorded as a changelog of new, changed and contradicted claims, sources that appeared or vanished, and confidence shifts. Topics show their current state, their confidence history and their recent changes, and can notify you by email when something moves.