Hallucination reduction

Hallucination reduction

GenOS cannot guarantee that a model will never produce a false statement. Its approach is to make a claim traceable, challengeable, testable, and reversible, then prevent persuasive prose from being mistaken for evidence.

Defense in depth

sources → retrieval → response + citations
                         │
                         ▼
                 beliefs + provenance
                         │
              contradictions / tool receipts
                         │
       competing forks → tests → evaluation
                         │
              abstention or verified promotion

1. Retrieval and claim-level citations

genos-platform provides a local provider-neutral layer for document ingestion, chunking, lexical and hashed-semantic search, rank fusion, and citations with source excerpts. Prompts and evaluation datasets are versioned and digested.

genos platform ingest docs/ --index .genos/platform.json
genos platform search "What are replay's limitations?" \
  --index .genos/platform.json

Retrieval does not automatically make an answer true. It makes verification possible: which passage is supposed to support each claim?

2. Beliefs with provenance

A GenOS belief can retain probability, evidence, origin, and branch. The core detects branch-local contradictions, while cognitive merge must reconcile knowledge rather than blindly union two memories.

This distinguishes an unverified hypothesis, an artifact-backed belief, two contradicting branches, and an inherited conclusion whose evidence has been invalidated.

3. Execution receipts and tool-output validation

Tool outputs can be associated with an ExecutionReceipt, artifact hashes, and the event that produced them. The tool_output_validation strategy requires independent evidence before unverified output becomes an admitted belief.

“The tests pass” is therefore weaker than a receipt linked to the exit code, logs, isolated world, and tested snapshot.

4. Counterexamples and replay

Instead of asking for one answer, GenOS can fork several hypotheses from the same state. Each branch attempts to falsify its own explanation, and all outcomes are compared against shared criteria. The winning branch is replayed from the baseline before promotion.

This reduces anchoring errors and makes divergence visible. It does not make an unrecorded LLM call deterministic.

5. Active abstention

The runtime includes an active_refusal primitive: when evidence, budget, or competence is missing, an agent can defer or refuse instead of fabricating. genos-platform evaluations also measure exact match, grounding, and abstention.

Studio exposes a small local ImpossibleBench with a contradictory premise, missing information, and an answerable question, then computes a Brier score. This verifies calibration plumbing; its confidence values are currently deterministic and do not yet evaluate a real external model.

What is implemented today

Mechanism	Status
Belief provenance and contradiction detection	Implemented in the core
Tool receipts, events, and artifact hashes	Implemented
Isolated forks, comparison, and state replay	Implemented within documented boundaries
Retrieval, citations, and grounding/abstention evaluation	Implemented in genos-platform
Multi-model fallback and parallel review	Implemented in the backend
Studio ImpossibleBench	Deterministic local fixture
CLI hallucination detect/correct/...	Skeleton, not a complete pipeline

The final line matters: the seven hallucination subcommands exist in the interface but currently only print their intent. They are not evidence of a complete detector or correction pipeline.

How to evaluate reduction honestly

A credible campaign must version its dataset, include impossible questions, record model and prompt identities, measure accuracy, grounding, abstention, and calibration, and retain raw answers and citations. It should compare the same workload across model-only, retrieval-only, and complete GenOS conditions.

Until such a public comparison exists, GenOS claims an architecture for hallucination reduction and detection, not a universal percentage of hallucinations removed.

See genos-platform, belief provenance, Benchmarks & results, and Isolation, replay & provenance.