Independent verification for aerospace AI.
The SPAIDER Trust Layer attaches evidence provenance, consistency checks, confidence estimates, and review state to agent inputs and outputs. It is a technical layer shared by AI Foundations, SAGAN, SPOCK, and KEPLER.
Requirement SYS-021 conflicts with the stored safe-mode power constraint. Both source passages are attached to the review record.
Signals about an answer and the evidence behind it.
Agents generate outputs. The Trust Layer independently evaluates the information used, the consistency of the result, its uncertainty, and the required review state.
Independent checks
Verification mechanisms are separate from the model or agent producing the answer.
Machine-readable output
Trust results are structured signals that interfaces, policies, reviewers, and downstream systems can use.
Continuous operation
Fast local checks are designed to run throughout a workflow rather than only during a final review.
Explicit human authority
Verification supports engineering and operational decisions; it does not remove accountable human approval.
Six connected verification mechanisms.
The modules have different maturity levels. The status labels below distinguish implemented foundations, active research, and product-development work.
Contradiction screening
Compares new facts, memory writes, and retrieved passages with existing knowledge. It returns agree, unrelated, or contradict signals with confidence.
Confidence estimation
Estimates uncertainty from semantic variation in model outputs. Low-confidence results can be labelled, routed to stronger verification, or escalated for review.
Evidence and provenance
Records the libraries, files, paths, passages, and retrieval operations used by an answer in a machine-readable response graph.
Memory consistency
Screens proposed memory updates, preserves operation history, and attaches conflict information so long-term context remains inspectable.
Epistemic work measurement
Measures how much a workflow stage reduces uncertainty, providing an alternative to treating answer length or token count as progress.
Review and escalation
Presents flagged evidence, conflicts, uncertainty, and approval state to the responsible engineer or operator and records the resolution.
Methods selected for continuous verification.
Sparse-signal contradiction detection
Read selected internal representation signals and classify agreement or contradiction without generating a second answer.
Semantic-entropy confidence
Sample outputs, cluster them by meaning, and calculate residual uncertainty without requiring hosted-model log probabilities.
Per-response provenance graphs
Record the exact evidence set behind each response as structured data rather than only displaying formatted citations.
Screening-funnel economics
Apply fast local checks broadly and reserve more expensive model or human review for flagged results.
The confidence research includes semantic-entropy methods described by Farquhar et al., Nature 2024. SPAIDER's contradiction and provenance architectures are developed as internal technical programs.
Review the Trust Layer for your workflow.
Discuss evidence requirements, verification signals, review gates, and deployment constraints with the SPAIDER team.