The AI flight recorder for insurable autonomy.
blkbx turns consequential AI actions into tamper-evident evidence receipts — so insurers, brokers, reinsurers, auditors, and enterprises can reconstruct what happened, why, who authorized it, and whether the system stayed inside its declared controls.
Record the action. Reveal the risk. Reconstruct the claim.
- model.id
- claims-agent-v3
- version
- 3.4.1 / cfg#a91f
- input.hash
- 0x8f3a…c210
- output.hash
- 0x1d77…9e04
- retrieval
- 6 docs · policy-kb
- policy.ref
- COV-4.2 / ADV-1
- controls
- 3/3 fired ✓
- human.review
- approved · ops-224
- tool.calls
- ledger.write ✓
- action
- claim.denied
- prev.hash
- 0x77aa…01bd
We make AI actions reconstructable enough to insure.
AI now approves, denies, refunds, escalates, binds, mutates records, drafts regulated communications, reviews claims, scores borrowers, and triggers workflows with financial or legal consequences.
But when something goes wrong, the evidence is fragmented. Logs do not line up, model versions are unclear, retrieval context is missing, human review is hard to prove, and compliance is reconstructed manually.
Insurers cannot price what they cannot reconstruct.
“Black box AI” usually means opacity. We are reversing the term. The black box should not be the mystery — it should be the recorder.
In aviation, the flight recorder made complex machine failure investigable. blkbx does the same for autonomous decision systems.
Every consequential AI action gets a signed receipt.
blkbx sits inside enterprise AI workflows and records the evidence behind high-consequence actions. Each receipt preserves the local evidence around the action and is tested for reconstructability.
A log says something happened.
A Black Box receipt shows whether the event can be reconstructed into an insurer-reviewable story.
Model / agent identity
Shows which system acted
Version and configuration
Supports reproducibility
Input and output hashes
Preserves decision context
Retrieval context
Shows what knowledge was used
Policy reference
Maps action to rules and obligations
Control checks
Shows whether safeguards fired
Human review status
Proves oversight and authority
Tool calls
Captures downstream execution
Action taken
Records the consequence
Timestamp and signature
Makes the event tamper-evident
Prior receipt hash
Preserves chronology
Verification status
Allows independent review
Fast path
Low-latency recorder
Routine consequential actions get an immediate signed receipt without slowing production workflows.
action → capture → signature → hash-chain → verification
Deep path
Reconstruction engine
High-risk, disputed, regulated, or claim-relevant events build a structured evidence cover and are tested for coherence.
cover → restriction maps → diagnostics → obstruction → packet
Five surfaces, one evidence layer.
Receipt API
A developer-facing API for issuing signed receipts from AI workflows.
AI platforms · agent frameworks · model gateways · underwriting engines
Evidence Ledger
A tamper-evident chronology of AI actions.
Audit trails · claims reconstruction · regulatory inquiry
Reconstructability Engine
A sheaf-native diagnostic layer measuring whether evidence fragments form a coherent global story.
High-risk decisions · incident review · disputed actions
Insurance Packet Builder
A structured export for brokers, carriers, reinsurers, and enterprise risk teams.
AI liability submissions · renewal evidence · claims files
Obstruction Dashboard
A risk console that shows where AI evidence fails to cohere.
Model risk · operational risk · vendor risk · claims QA
We do not sell insurance. We make AI systems more insurable.
Enterprise
Prove AI controls worked
Signed receipts and control evidence
Broker
Package AI risk clearly
Submission-ready evidence packets
Carrier
Underwrite with confidence
Reconstructable decision records
Claims team
Investigate AI-related loss
Event chronology and obstruction map
Reinsurer
Understand portfolio exposure
Structured AI activity data
Regulator
Review consequential decisions
Policy-mapped evidence trail
Join the evidence layer for insurable AI.
We are stealth. The thesis is not. If an AI action can create a loss, it should create a receipt.
Read the receipt schema · See an example claim reconstruction packet